Canalblog
Editer l'article Suivre ce blog Administration + Créer mon blog
Publicité
NEWS OFF AMERICAN INTERNET COUNCIL COMPANY - DIACONESCO.TV Phone : 06.32.17.36.33/ 06.50.34.10.26 diaconesco@gmail.com
NEWS OFF AMERICAN INTERNET COUNCIL COMPANY - DIACONESCO.TV Phone : 06.32.17.36.33/ 06.50.34.10.26 diaconesco@gmail.com
Publicité
Visiteurs
Hier 11
Ce mois ci 198
Depuis la création 1 237 469
Newsletter
122 abonnés
NEWS OFF AMERICAN INTERNET COUNCIL COMPANY - DIACONESCO.TV Phone : 06.32.17.36.33/ 06.50.34.10.26 diaconesco@gmail.com
Derniers commentaires
Archives
27 mars 2018

SCIENCES TO DAY AND THE FUTURE IN 2030 WITH THE ARTIFICIAL INTELLIGENCE !

 

LBYgxc43

ARTIFICIAL INTELLIGENCE AND LIFE IN 2030 ONE HUNDRED YEAR STUDY ON ARTIFICIAL INTELLIGENCE | REPORT OF THE 2015 STUDY PANEL | SEPTEMBER 2016 PREFACE

C6F2KknWQAAW8kA 

The One Hundred Year Study on Artificial Intelligence, launched in the fall of 2014, is a long-term investigation of the field of Artificial Intelligence (AI) and its influences on people, their communities, and society. It considers the science, engineering, and deployment of AI-enabled computing systems. As its core activity, the Standing Committee that oversees the One Hundred Year Study forms a Study Panel every five years to assess the current state of AI. The Study Panel reviews AI’s progress in the years following the immediately prior report, envisions the potential advances that lie ahead, and describes the technical and societal challenges and opportunities these advances raise, including in such arenas as ethics, economics, and the design of systems compatible with human cognition.

The overarching purpose of the One Hundred Year Study’s periodic expert review is to provide a collected and connected set of reflections about AI and its influences as the field advances. The studies are expected to develop syntheses and assessments that provide expert-informed guidance for directions in AI research, development, and systems design, as well as programs and policies to help ensure that these systems broadly benefit individuals and society.

1 The One Hundred Year Study is modeled on an earlier effort informally known as the “AAAI Asilomar Study.” During 2008-2009, the then president of the Association for the Advancement of Artificial Intelligence (AAAI), Eric Horvitz, assembled a group of AI experts from multiple institutions and areas of the field, along with scholars of cognitive science, philosophy, and law. Working in distributed subgroups, the participants addressed near-term AI developments, long-term possibilities, and legal and ethical concerns, and then came together in a three-day meeting at Asilomar to share and discuss their findings. A short written report on the intensive meeting discussions, amplified by the participants’ subsequent discussions with other colleagues, generated widespread interest and debate in the field and beyond.

 The impact of the Asilomar meeting, and important advances in AI that included AI algorithms and technologies starting to enter daily life around the globe, spurred the idea of a long-term recurring study of AI and its influence on people and society. The One Hundred Year Study was subsequently endowed at a university to enable 1 “One Hundred Year Study on Artificial Intelligence (AI100),” Stanford University, accessed August 1, 2016,

https://ai100.stanford.edu

The overarching purpose of the One Hundred Year Study’s periodic expert review is to provide a collected and connected set of reflections about AI and its influences as the field advances.

2-3 As one consequence of the decision to focus on life in North American cities, military applications were deemed to be outside the scope of this initial report. This is not to minimize the importance of careful monitoring and deliberation about the implications of AI advances for defense and warfare, including potentially destabilizing developments and deployments. The report is designed to address four intended audiences. For the general public, it aims to provide an accessible, scientifically and technologically accurate portrayal of the current state of AI and its potential. For industry, the report describes relevant technologies and legal and ethical challenges, and may help guide resource allocation. The report is also directed to local, national, and international governments to help them better plan for AI in governance.

Finally, the report can help AI researchers, as well as their institutions and funders, to set priorities and consider the ethical and legal issues raised by AI research and its applications. Given the unique nature of the One Hundred Year Study on AI, we expect that future generations of Standing Committees and Study Panels, as well as research scientists, policy experts, leaders in the private and public sectors, and the general public, will reflect on this assessment as they make new assessments of AI’s future. We hope that this first effort in the series stretching out before us will be useful for both its failures and successes in accurately predicting the trajectory and influences of AI.

55653

The Standing Committee is grateful to the members of the Study Panel for investing their expertise, perspectives, and significant time to the creation of this inaugural report. We especially thank Professor Peter Stone for agreeing to serve as chair of the study and for his wise, skillful, and dedicated leadership of the panel, its discussions, and creation of the report. Standing Committee of the One Hundred Year Study of Artificial Intelligence Barbara J. Grosz, Chair Russ Altman Eric Horvitz Alan Mackworth Tom Mitchell Deidre Mulligan Yoav Shoham extended deep thought and cross-disciplinary scholarly investigations that could inspire innovation and provide intelligent advice to government agencies and industry.

This report is the first in the planned series of studies that will continue for at least a hundred years. The Standing Committee defined a Study Panel charge for the inaugural Study Panel in the summer of 2015 and recruited Professor Peter Stone, at the University of Texas at Austin, to chair the panel. The seventeen-member Study Panel, comprised of experts in AI from academia, corporate laboratories and industry, and AI-savvy scholars in law, political science, policy, and economics, was launched in mid-fall 2015. The participants represent diverse specialties and geographic regions, genders, and career stages. The Standing Committee extensively discussed ways to frame the Study Panel charge to consider both recent advances in AI and potential societal impacts on jobs, the environment, transportation, public safety, healthcare, community engagement, and government.

The committee considered various ways to focus the study, including surveying subfields and their status, examining a particular technology such as machine learning or natural language processing, and studying particular application areas such as healthcare or transportation. The committee ultimately chose a thematic focus on “AI and Life in 2030” to recognize that AI’s various uses and impacts will not occur independently of one another, or of a multitude of other societal and technological developments. Acknowledging the central role cities have played throughout most of human experience, the focus was narrowed to the large urban areas where most people live.

The Standing Committee further narrowed the focus to a typical North American city in recognition of the great variability of urban settings and cultures around the world, and limits on the first Study Panel’s efforts. The Standing Committee expects that the projections, assessments, and proactive guidance stemming from the study will have broader global relevance and is making plans for future studies to expand the scope of the project internationally. STUDY PANEL Peter Stone, University of Texas at Austin, Chair Rodney Brooks, Rethink Robotics Erik Brynjolfsson, Massachussets Institute of Technology Ryan Calo, University of Washington Oren Etzioni, Allen Institute for AI Greg Hager, Johns Hopkins University Julia Hirschberg, Columbia University Shivaram Kalyanakrishnan, Indian Institute of Technology Bombay Ece Kamar, Microsoft Research Sarit Kraus, Bar Ilan University Kevin Leyton-Brown, University of British Columbia David Parkes, Harvard University William Press, University of Texas at Austin AnnaLee (Anno) Saxenian, University of California, Berkeley Julie Shah, Massachussets Institute of Technology Milind Tambe, University of Southern California Astro Teller, X Acknowledgments: The members of the Study Panel gratefully acknowledge the support of and valuable input from the Standing Committee, especially the chair, Barbara Grosz, who handled with supreme grace the unenviable role of mediating between two large, very passionate committees.

We also thank Kerry Tremain for his tireless and insightful input on the written product during the extensive editing and polishing process, which unquestionably strengthened the report considerably.

C6F1FW3XMAAs83x

TABLE OF CONTENTS PREFACE EXECUTIVE SUMMARY 4 OVERVIEW 6 SECTION I: WHAT IS ARTIFICIAL INTELLIGENCE?

12 Defining AI 12 AI Research Trends 14 SECTION II: AI BY DOMAIN 18 Transportation 18 Home/Service Robots 24 Healthcare 25 Education 31 Low-resource Communities 35 Public Safety and Security 36 Employment and Workplace 38 Entertainment 40 SECTION III: PROSPECTS AND RECOMMENDATIONS FOR AI PUBLIC POLICY 42 AI Policy, Now and in the Future 42 APPENDIX I: A SHORT HISTORY OF AI 50 4 5 labor is augmented or replaced by AI, creating new challenges for the economy and society more broadly. Application design and policy decisions made in the near term are likely to have long-lasting influences on the nature and directions of such developments, making it important for AI researchers, developers, social scientists, and policymakers to balance the imperative to innovate with mechanisms to ensure that AI’s economic and social benefits are broadly shared across society. If society approaches these technologies primarily with fear and suspicion, missteps that slow AI’s development or drive it underground will result, impeding important work on ensuring the safety and reliability of AI technologies.

On the other hand, if society approaches AI with a more open mind, the technologies emerging from the field could profoundly transform society for the better in the coming decades. Study Panel: Peter Stone, Chair, University of Texas at Austin, Rodney Brooks, Rethink Robotics, Erik Brynjolfsson, Massachussets Institute of Technology, Ryan Calo, University of Washington, Oren Etzioni, Allen Institute for AI, Greg Hager, Johns Hopkins University, Julia Hirschberg, Columbia University, Shivaram Kalyanakrishnan, Indian Institute of Technology Bombay, Ece Kamar, Microsoft Research, Sarit Kraus, Bar Ilan University. Kevin Leyton-Brown, University of British Columbia, David Parkes, Harvard University, William Press, University of Texas at Austin, AnnaLee (Anno) Saxenian, University of California, Berkeley, Julie Shah, Massachussets Institute of Technology, Milind Tambe, University of Southern California, Astro Teller, X Standing Committee of the One Hundred Year Study of Artificial Intelligence: Barbara J. Grosz, Chair, Russ Altman, Eric Horvitz, Alan Mackworth, Tom Mitchell, Deidre Mulligan, Yoav Shoham EXECUTIVE SUMMARY Artificial Intelligence (AI) is a science and a set of computational technologies that are inspired by—but typically operate quite differently from—the ways people use their nervous systems and bodies to sense, learn, reason, and take action.

C8EYcRtXUAAfZli

While the rate of progress in AI has been patchy and unpredictable, there have been significant advances since the field’s inception sixty years ago. Once a mostly academic area of study, twenty-first century AI enables a constellation of mainstream technologies that are having a substantial impact on everyday lives. Computer vision and AI planning, for example, drive the video games that are now a bigger entertainment industry than Hollywood. Deep learning, a form of machine learning based on layered representations of variables referred to as neural networks, has made speech-understanding practical on our phones and in our kitchens, and its algorithms can be applied widely to an array of applications that rely on pattern recognition.

Natural Language Processing (NLP) and knowledge representation and reasoning have enabled a machine to beat the Jeopardy champion and are bringing new power to Web searches. While impressive, these technologies are highly tailored to particular tasks. Each application typically requires years of specialized research and careful, unique construction. In similarly targeted applications, substantial increases in the future uses of AI technologies, including more self-driving cars, healthcare diagnostics and targeted treatments, and physical assistance for elder care can be expected.

AI and robotics will also be applied across the globe in industries struggling to attract younger workers, such as agriculture, food processing, fulfillment centers, and factories. They will facilitate delivery of online purchases through flying drones, self-driving trucks, or robots that can get up the stairs to the front door. This report is the first in a series to be issued at regular intervals as a part of the One Hundred Year Study on Artificial Intelligence (AI100). Starting from a charge given by the AI100 Standing Committee to consider the likely influences of AI in a typical North American city by the year 2030, the 2015 Study Panel, comprising experts in AI and other relevant areas focused their attention on eight domains they considered most salient: transportation; service robots; healthcare; education; low-resource communities; public safety and security; employment and workplace; and entertainment. In each of these domains, the report both reflects on progress in the past fifteen years and anticipates developments in the coming fifteen years.

Though drawing from a common source of research, each domain reflects different AI influences and challenges, such as the difficulty of creating safe and reliable hardware (transportation and service robots), the difficulty of smoothly interacting with human experts (healthcare and education), the challenge of gaining public trust (low-resource communities and public safety and security), the challenge of overcoming fears of marginalizing humans (employment and workplace), and the social and societal risk of diminishing interpersonal interactions (entertainment). The report begins with a reflection on what constitutes Artificial Intelligence, and concludes with recommendations concerning AI-related policy.

Automatisation-robots

These recommendations include accruing technical expertise about AI in government and devoting more resources—and removing impediments—to research on the fairness, security, privacy, and societal impacts of AI systems. Contrary to the more fantastic predictions for AI in the popular press, the Study Panel found no cause for concern that AI is an imminent threat to humankind. No machines with self-sustaining long-term goals and intent have been developed, nor are they likely to be developed in the near future. Instead, increasingly useful applications of AI, with potentially profound positive impacts on our society and economy are likely to emerge between now and 2030, the period this report considers.

At the same time, many of these developments will spur disruptions in how human Substantial increases in the future uses of AI applications, including more self-driving cars, healthcare diagnostics and targeted treatment, and physical assistance for elder care can be expected. While drawing on common research and technologies, AI systems are specialized to accomplish particular tasks. Each application requires years of focused research and a careful, unique construction. 6 7 An accurate and sophisticated picture of AI—one that competes with its popular portrayal—is hampered at the start by the difficulty of pinning down a precise definition of artificial intelligence. In the approaches the Study Panel considered, none suggest there is currently a “general purpose” AI.

While drawing on common research and technologies, AI systems are specialized to accomplish particular tasks, and each application requires years of focused research and a careful, unique construction. As a result, progress is uneven within and among the eight domains. A prime example is Transportation, where a few key technologies have catalyzed the widespread adoption of AI with astonishing speed. Autonomous transportation will soon be commonplace and, as most people’s first experience with physically embodied AI systems, will strongly influence the public’s perception of AI.

As cars become better drivers than people, city-dwellers will own fewer cars, live further from work, and spend time differently, leading to an entirely new urban organization. In the typical North American city in 2030, physically embodied AI applications will not be limited to cars, but are likely to include trucks, flying vehicles, and personal robots. Improvements in safe and reliable hardware will spur innovation over the next fifteen years, as they will with Home/Service Robots, which have already entered people’s houses, primarily in the form of vacuum cleaners. 

onu-sophia-robot-8fd8a6-1@1x

Better chips, low-cost 3D sensors, cloud-based machine learning, and advances in speech understanding will enhance future robots’ services and their interactions with people. Special purpose robots will deliver packages, clean offices, and enhance security. But technical constraints and the high costs of reliable mechanical devices will continue to limit commercial opportunities to narrowly defined applications for the foreseeable future. In Healthcare, there has been an immense forward leap in collecting useful data from personal monitoring devices and mobile apps, from electronic health records (EHR) in clinical settings and, to a lesser extent, from surgical robots designed to assist with medical procedures and service robots supporting hospital operations. AI-based applications could improve health outcomes and the quality of life for millions of people in the coming years.

Though clinical applications have been slow to move from the computer science lab to the real-world, there are hopeful signs that the pace of innovation will improve. Advances in healthcare can be promoted via the development of incentives and mechanisms for sharing data and for removing overbearing policy, regulatory, and commercial obstacles. For many applications, AI systems will have to work closely with care providers and patients to gain their trust. Advances in how intelligent machines interact naturally with caregivers, patients, and patients’ families are crucial. Enabling more fluid interactions between people and promising AI technologies also remains a critical challenge in Education, which has seen considerable progress in the same period.

Though quality education will always require active engagement by human teachers, AI promises to enhance education at all levels, especially by providing personalization at scale. Interactive machine tutors are now being matched to students for teaching science, math, language, and other disciplines. Natural Language Processing, machine learning, and crowdsourcing have boosted online learning and enabled teachers in higher education to multiply the size of their classrooms while addressing individual students’ learning needs and styles. Over the next fifteen years in a typical North American city, the use of these technologies in the classroom and in the home is likely to expand significantly, provided they can be meaningfully integrated with face-to-face learning.

C5b6VIoXEAEMAfJ

Beyond education, many opportunities exist for AI methods to assist Low-resource Communities by providing mitigations and solutions to a variety of social problems. Traditionally, funders have underinvested in AI research lacking commercial application. With targeted incentives and funding priorities, OVERVIEW The frightening, futurist portrayals of Artificial Intelligence that dominate films and novels, and shape the popular imagination, are fictional. In reality, AI is already changing our daily lives, almost entirely in ways that improve human health, safety, and productivity. Unlike in the movies, there is no race of superhuman robots on the horizon or probably even possible. And while the potential to abuse AI technologies must be acknowledged and addressed, their greater potential is, among other things, to make driving safer, help children learn, and extend and enhance people’s lives.

In fact, beneficial AI applications in schools, homes, and hospitals are already growing at an accelerated pace. Major research universities devote departments to AI studies, and technology companies such as Apple, Facebook, Google, IBM, and Microsoft spend heavily to explore AI applications they regard as critical to their futures. Even Hollywood uses AI technologies to bring its dystopian AI fantasies to the screen. Innovations relying on computer-based vision, speech recognition, and Natural Language Processing have driven these changes, as have concurrent scientific and technological advances in related fields. AI is also changing how people interact with technology. Many people have already grown accustomed to touching and talking to their smart phones. People’s future relationships with machines will become ever more nuanced, fluid, and personalized as AI systems learn to adapt to individual personalities and goals. These AI applications will help monitor people’s well-being, alert them to risks ahead, and deliver services when needed or wanted. For example, in a mere fifteen years in a typical North American city—the time frame and scope of this report— AI applications are likely to transform transportation toward self-driving vehicles with on-time pickup and delivery of people and packages. This alone will reconfigure the urban landscape, as traffic jams and parking challenges become obsolete. This study’s focus on a typical North American city is deliberate and meant to highlight specific changes affecting the everyday lives of the millions of people who inhabit them.

The Study Panel further narrowed its inquiry to eight domains where AI is already having or is projected to have the greatest impact: transportation, healthcare, education, low-resource communities, public safety and security, employment and workplace, home/service robots, and entertainment. Though drawing from a common source of research, AI technologies have influenced and will continue to influence these domains differently. Each domain faces varied AI-related challenges, including the difficulty of creating safe and reliable hardware for sensing and effecting (transportation and service robots), the difficulty of smoothly interacting with human experts (healthcare and education), the challenge of gaining public trust (low-resource communities and public safety and security), the challenge of overcoming fears of marginalizing humans (employment and workplace) and the risk of diminishing interpersonal interaction (entertainment).

C4xz-uVWAAEGfPs

Some domains are primarily business sectors, such as transportation and healthcare, while others are more oriented to consumers, such as entertainment and home service robots. Some cut across sectors, such as employment/workplace and low-resource communities. In each domain, even as AI continues to deliver important benefits, it also raises important ethical and social issues, including privacy concerns. Robots and other AI technologies have already begun to displace jobs in some sectors. As a society, we are now at a crucial juncture in determining how to deploy AI-based technologies in ways that promote, not hinder, democratic values such as freedom, equality, and transparency. For individuals, the quality of the lives we lead and how our contributions are valued are likely to shift gradually, but markedly. Over the next several years, AI research, systems development, and social and regulatory frameworks will shape how the benefits of AI are weighed against its costs and risks, and how broadly these benefits are spread. Many have already grown accustomed to touching and talking to their smart phones.

People’s future relationships with machines will become ever more nuanced, fluid, and personalized. Society is now at a crucial juncture in determining how to deploy AI-based technologies in ways that promote rather than hinder democratic values such as freedom, equality, and transparency. 8 9 markets for data-driven products, and the economic incentives to find new products and markets, have also stimulated research advances. Now, as it becomes a central force in society, the field of AI is shifting toward building intelligent systems that can collaborate effectively with people, and that are more generally human-aware, including creative ways to develop interactive and scalable ways for people to teach robots. These trends drive the currently “hot” areas of AI research into both fundamental methods and application areas: Large-scale machine learning concerns the design of learning algorithms, as well as scaling existing algorithms, to work with extremely large data sets. Deep learning, a class of learning procedures, has facilitated object recognition in images, video labeling, and activity recognition, and is making significant inroads into other areas of perception, such as audio, speech, and natural language processing.

Reinforcement learning is a framework that shifts the focus of machine learning from pattern recognition to experience-driven sequential decision-making. It promises to carry AI applications forward toward taking actions in the real world. While largely confined to academia over the past several decades, it is now seeing some practical, real-world successes. Robotics is currently concerned with how to train a robot to interact with the world around it in generalizable and predictable ways, how to facilitate manipulation of objects in interactive environments, and how to interact with people. Advances in robotics will rely on commensurate advances to improve the reliability and generality of computer vision and other forms of machine perception. Computer vision is currently the most prominent form of machine perception.

universal-robots-2-jpg

It has been the sub-area of AI most transformed by the rise of deep learning. For the first time, computers are able to perform some vision tasks better than people. Much current research is focused on automatic image and video captioning. Natural Language Processing, often coupled with automatic speech recognition, is quickly becoming a commodity for widely spoken languages with large data sets. Research is now shifting to develop refined and capable systems that are able to interact with people through dialog, not just react to stylized requests. Great strides have also been made in machine translation among different languages, with more real-time person-to-person exchanges on the near horizon. Collaborative systems research investigates models and algorithms to help develop autonomous systems that can work collaboratively with other systems and with humans. Crowdsourcing and human computation research investigates methods to augment computer systems by making automated calls to human expertise to solve problems that computers alone cannot solve well.

Algorithmic game theory and computational social choice draw attention to the economic and social computing dimensions of AI, such as how systems can handle potentially misaligned incentives, including self-interested human participants or firms and the automated AI-based agents representing them. Internet of Things (IoT) research is devoted to the idea that a wide array of devices, including appliances, vehicles, buildings, and cameras, can be interconnected to collect and share their abundant sensory information to use for intelligent purposes. Neuromorphic computing is a set of technologies that seek to mimic biological neural networks to improve the hardware efficiency and robustness of computing systems, often replacing an older emphasis on separate modules for input/ output, instruction-processing, and memory. AI technologies could help address the needs of low-resource communities, and budding efforts are promising.

Using data mining and machine learning, for example, AI has been used to create predictive models to help government agencies address issues such as prevention of lead poisoning in at-risk children and distribution of food efficiently. These budding efforts suggest more could be done, particularly if agencies and organizations can engage and build trust with these communities. Gaining public trust is also a challenge for AI use by Public Safety and Security professionals. North American cities and federal agencies have already begun to deploy AI technologies in border administration and law enforcement. By 2030, they will rely heavily upon them, including improved cameras and drones for surveillance, algorithms to detect financial fraud, and predictive policing. The latter raises the specter of innocent people being unjustifiably monitored, and care must be taken to avoid systematizing human bias and to protect civil liberties.

harfang-329182-jpg_208067

Well-deployed AI prediction tools have the potential to provide new kinds of transparency about data and inferences, and may be applied to detect, remove, or reduce human bias, rather than reinforcing it. Social and political decisions are likewise at play in AI’s influences on Employment and Workplace trends, such as the safety nets needed to protect people from structural changes in the economy. AI is poised to replace people in certain kinds of jobs, such as in the driving of taxis and trucks. However, in many realms, AI will likely replace tasks rather than jobs in the near term, and will also create new kinds of jobs. But the new jobs that will emerge are harder to imagine in advance than the existing jobs that will likely be lost. AI will also lower the cost of many goods and services, effectively making everyone better off. Longer term, AI may be thought of as a radically different mechanism for wealth creation in which everyone should be entitled to a portion of the world’s AI-produced treasures. It is not too soon for social debate on how the economic fruits of AI technologies should be shared.

Entertainment has been transformed by social networks and other platforms for sharing and browsing blogs, videos, and photos, which rely on techniques actively developed in NLP, information retrieval, image processing, crowdsourcing, and machine learning. Some traditional sources of entertainment have also embraced AI to compose music, create stage performances, and even to generate 3D scenes from natural language text. The enthusiasm with which people have already responded to AI-driven entertainment has been surprising. As with many aspects of AI, there is ongoing debate about the extent to which the technology replaces or enhances sociability. AI will increasingly enable entertainment that is more interactive, personalized, and engaging. Research should be directed toward understanding how to leverage these attributes for individuals’ and society’s benefit. What’s next for AI research? The research that fuels the AI revolution has also seen rapid changes. Foremost among them is the maturation of machine learning, stimulated in part by the rise of the digital economy, which both provides and leverages large amounts of data.

Other factors include the rise of cloud computing resources and consumer demand for widespread access to services such as speech recognition and navigation support. Machine learning has been propelled dramatically forward by impressive empirical successes of artificial neural networks, which can now be trained with huge data sets and large-scale computing. This approach has been come to be known as “deep learning.” The leap in the performance of information processing algorithms has been accompanied by significant progress in hardware technology for basic operations such as sensing, perception, and object recognition. New platforms and Longer term, AI may be thought of as a radically different mechanism for wealth creation in which everyone should be entitled to a portion of the world’s AI-produced treasures.

The field of AI is shifting toward building intelligent systems that can collaborate effectively with people, including creative ways to develop interactive and scalable ways for people to teach robots. 10 11 the abilities and efficiency of people who have access to them. Policies should be evaluated as to whether they foster democratic values and equitable sharing of AI’s benefits, or concentrate power and benefits in the hands of a fortunate few. As this report documents, significant AI-related advances have already had an impact on North American cities over the past fifteen years, and even more substantial developments will occur over the next fifteen. Recent advances are largely due to the growth and analysis of large data sets enabled by the internet, advances in sensory technologies and, more recently, applications of “deep learning.”

C5XmbWeWMAA_wfd

In the coming years, as the public encounters new AI applications in domains such as transportation and healthcare, they must be introduced in ways that build trust and understanding, and respect human and civil rights. While encouraging innovation, policies and processes should address ethical, privacy, and security implications, and should work to ensure that the benefits of AI technologies will be spread broadly and fairly. Doing so will be critical if Artificial Intelligence research and its applications are to exert a positive influence on North American urban life in 2030 and beyond. AI policy, now and in the future The measure of success for AI applications is the value they create for human lives. In that light, they should be designed to enable people to understand AI systems successfully, participate in their use, and build their trust.

Public policies should help ease society’s adaptation to AI applications, extend their benefits, and mitigate their inevitable errors and failures. Debate about how AI is deployed, including concerns about how privacy is protected and AI’s benefits fairly shared, should be encouraged. Given the speed with which AI technologies are being realized, and concomitant concerns about their implications, the Study Panel recommends that all layers of government acquire technical expertise in AI. Further, research on the fairness, security, privacy, and societal implications of AI systems should be encouraged by removing impediments and increasing private and public spending to support it. Currently in the United States, at least sixteen separate agencies govern sectors of the economy related to AI technologies. Rapid advances in AI research and, especially, its applications require experts in these sectors to develop new concepts and metaphors for law and policy.

Who is responsible when a self-driven car crashes or an intelligent medical device fails?

How can AI applications be prevented from promulgating racial discrimination or financial cheating?

Who should reap the gains of efficiencies enabled by AI technologies and what protections should be afforded to people whose skills are rendered obsolete?

Wiki-background

As people integrate AI more broadly and deeply into industrial processes and consumer products, best practices need to be spread, and regulatory regimes adapted. While the Study Panel does not consider it likely that near-term AI systems will autonomously choose to inflict harm on people, it will be possible for people to use AI-based systems for harmful as well as helpful purposes. And though AI algorithms may be capable of making less biased decisions than a typical person, it remains a deep technical challenge to ensure that the data that inform AI-based decisions can be kept free from biases that could lead to discrimination based on race, sexual orientation, or other factors. Faced with the profound changes that AI technologies can produce, pressure for “more” and “tougher” regulation is probably inevitable. Misunderstandings about what AI is and is not could fuel opposition to technologies with the potential to benefit everyone. Inappropriate regulatory activity would be a tragic mistake.

Poorly informed regulation that stifles innovation, or relocates it to other jurisdictions, would be counterproductive.

2 Fortunately, principles that guide successful regulation of current digital technologies provide a starting point. In privacy regulation, broad legal mandates coupled with tough transparency requirements and meaningful enforcement—rather than strict controls—encourage companies to develop processes and professional staff to enforce privacy controls, engage with outside stakeholders, and adapt their practices to technological advances. This in turn supports the development of professional trade associations and standards committees that spread best practices. In AI, too, regulators can strengthen a virtuous cycle of activity involving internal and external accountability, transparency, and professionalization, rather than narrow compliance. A vigorous and informed debate about how to best steer AI in ways that enrich our lives and our society, while encouraging creativity in the field, is an urgent and vital need.

AI technologies could widen existing inequalities of opportunity if access to them—along with the high-powered computation and large-scale data that fuel many of them—is unfairly distributed across society. These technologies will improve 2 Kate Crawford, “Artificial Intelligence’s White Guy Problem,” The New York Times, June 25, 2016, accessed August 1, 2016,

Misunderstandings about what AI is and is not could fuel opposition to technologies with the potential to benefit everyone. Poorly informed regulation that stifles innovation would be a tragic mistake. 12 13 The human measure Notably, the characterization of intelligence as a spectrum grants no special status to the human brain. But to date human intelligence has no match in the biological and artificial worlds for sheer versatility, with the abilities “to reason, achieve goals, understand and generate language, perceive and respond to sensory inputs, prove mathematical theorems, play challenging games, synthesize and summarize information, create art and music, and even write histories.”

C5b7kb4XEAIqgw1

6 - This makes human intelligence a natural choice for benchmarking the progress of AI. It may even be proposed, as a rule of thumb, that any activity computers are able to perform and people once performed should be counted as an instance of intelligence. But matching any human ability is only a sufficient condition, not a necessary one. There are already many systems that exceed human intelligence, at least in speed, such as scheduling the daily arrivals and departures of thousands of flights in an airport. AI’s long quest—and eventual success—to beat human players at the game of chess offered a high-profile instance for comparing human to machine intelligence. Chess has fascinated people for centuries. When the possibility of building computers became imminent, Alan Turing, who many consider the father of computer science, “mentioned the idea of computers showing intelligence with chess as a paradigm.”

7 - Without access to powerful computers, “Turing played a game in which he simulated the computer, taking about half an hour per move.” But it was only after a long line of improvements in the sixties and seventies— contributed by groups at Carnegie Mellon, Stanford, MIT, The Institute for Theoretical and Experimental Physics at Moscow, and Northwestern University— that chess-playing programs started gaining proficiency. The final push came through a long-running project at IBM, which culminated with the Deep Blue program beating Garry Kasparov, then the world chess champion, by a score of 3.5-2.5 in 1997. Curiously, no sooner had AI caught up with its elusive target than Deep Blue was portrayed as a collection of “brute force methods” that wasn’t “real intelligence.”

8 - In fact, IBM’s subsequent publication about Deep Blue, which gives extensive details about its search and evaluation procedures, doesn’t mention the word “intelligent” even once!

9 - Was Deep Blue intelligent or not? Once again, the frontier had moved. An operational definition AI can also be defined by what AI researchers do. This report views AI primarily as a branch of computer science that studies the properties of intelligence by synthesizing intelligence.

10 - Though the advent of AI has depended on the rapid progress of hardware computing resources, the focus here on software reflects a trend in the AI community. More recently, though, progress in building hardware tailored for neural-network-based computing11 has created a 6 Nilsson, The Question for Artificial Intelligence. 7 Nilsson, The Question for Artificial Intelligence, 89. 8 McCorduck, Machines Who Think, 433. 9 Murray Campbell, A. Joseph Hoane Jr., and Feng-hsiung Hsu, “Deep Blue,” Artificial Intelligence 134, nos. 1 and 2 (2002): 57–83. 10 Herbert A. Simon, “Artificial Intelligence: An Empirical Science,” Artificial Intelligence 77, no. 2 (1995):95–127. 11 Paul Merolla John V. Arthur, Rodrigo Alvarez-Icaza, Andrew S. Cassidy, Jun Sawada, Filipp Akopyan, Bryan L. Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, Bernard Brezzo, Ivan Vo, Steven K. Esser, Rathinakumar Appuswamy, Brian Taba, Arnon Amir, Myron D. Flickner, William P. Risk, Rajit Manohar, and Dharmendra S. Modha, “A Million Spiking-Neuron Integrated Circuit with a Scalable Communication Network and Interface,” accessed August 1, 2016.

C5g4ZlKWYAEw0NZ

SECTION I: WHAT IS ARTIFICIAL INTELLIGENCE? This section describes how researchers and practitioners define “Artificial Intelligence,” and the areas of AI research and application that are currently thriving. It proffers definitions of what AI is and is not, and describes some of the currently “hot” areas of AI Research. This section lays the groundwork for Section II, which elaborates on AI’s impacts and future in eight domains and Section III, which describes issues related to AI design and public policy and makes recommendations for encouraging AI innovation while protecting democratic values. DEFINING AI Curiously, the lack of a precise, universally accepted definition of AI probably has helped the field to grow, blossom, and advance at an ever-accelerating pace. Practitioners, researchers, and developers of AI are instead guided by a rough sense of direction and an imperative to “get on with it.” Still, a definition remains important and Nils J. Nilsson has provided a useful one: “Artificial intelligence is that activity devoted to making machines intelligent, and intelligence is that quality that enables an entity to function appropriately and with foresight in its environment.”

3 - From this perspective, characterizing AI depends on the credit one is willing to give synthesized software and hardware for functioning “appropriately” and with “foresight.” A simple electronic calculator performs calculations much faster than the human brain, and almost never makes a mistake.

4 - Is a calculator intelligent? Like Nilsson, the Study Panel takes a broad view that intelligence lies on a multidimensional spectrum. According to this view, the difference between an arithmetic calculator and a human brain is not one of kind, but of scale, speed, degree of autonomy, and generality. The same factors can be used to evaluate every other instance of intelligence—speech recognition software, animal brains, cruise-control systems in cars, Go-playing programs, thermostats—and to place them at some appropriate location in the spectrum. Although our broad interpretation places the calculator within the intelligence spectrum, such simple devices bear little resemblance to today’s AI. The frontier of AI has moved far ahead and functions of the calculator are only one among the millions that today’s smartphones can perform. AI developers now work on improving, generalizing, and scaling up the intelligence currently found on smartphones. In fact, the field of AI is a continual endeavor to push forward the frontier of machine intelligence. Ironically, AI suffers the perennial fate of losing claim to its acquisitions, which eventually and inevitably get pulled inside the frontier, a repeating pattern known as the “AI effect” or the “odd paradox”—AI brings a new technology into the common fold, people become accustomed to this technology, it stops being considered AI, and newer technology emerges. 5The same pattern will continue in the future. AI does not “deliver” a life-changing product as a bolt from the blue. Rather, AI technologies continue to get better in a continual, incremental way. 3 Nils J. Nilsson, The Quest for Artificial Intelligence: A History of Ideas and Achievements (Cambridge, UK: Cambridge University Press, 2010). 4 Wikimedia Images, accessed August 1, 2016.

C8EXzVJXwAAhxN7

5 - Pamela McCorduck, Machines Who Think: A Personal Inquiry into the History and Prospects of Artificial Intelligence, 2nd ed. (Natick, MA: A. K. Peters, Ltd., 2004; San Francisco: W. H. Freeman, 1979), Citations are to the Peters edition. Intelligence lies on a multi-dimensional spectrum. According to this view, the difference between an arithmetic calculator and a human brain is not one of kind, but of scale, speed, degree of autonomy, and generality. An accurate and sophisticated picture of AI—one that competes with its popular portrayal—is hampered by the difficulty of pinning down a precise definition of artificial intelligence. 14 15 labeling, activity recognition, and several variants thereof. Deep learning is also making significant inroads into other areas of perception, such as audio, speech, and natural language processing. Reinforcement learning Whereas traditional machine learning has mostly focused on pattern mining, reinforcement learning shifts the focus to decision making, and is a technology that will help AI to advance more deeply into the realm of learning about and executing actions in the real world. It has existed for several decades as a framework for experience-driven sequential decision-making, but the methods have not found great success in practice, mainly owing to issues of representation and scaling.

However, the advent of deep learning has provided reinforcement learning with a “shot in the arm.” The recent success of AlphaGo, a computer program developed by Google Deepmind that beat the human Go champion in a five-game match, was due in large part to reinforcement learning. AlphaGo was trained by initializing an automated agent with a human expert database, but was subsequently refined by playing a large number of games against itself and applying reinforcement learning. Robotics Robotic navigation, at least in static environments, is largely solved. Current efforts consider how to train a robot to interact with the world around it in generalizable and predictable ways. A natural requirement that arises in interactive environments is manipulation, another topic of current interest. The deep learning revolution is only beginning to influence robotics, in large part because it is far more difficult to acquire the large labeled data sets that have driven other learning-based areas of AI. Reinforcement learning (see above), which obviates the requirement of labeled data, may help bridge this gap but requires systems to be able to safely explore a policy space without committing errors that harm the system itself or others. Advances in reliable machine perception, including computer vision, force, and tactile perception, much of which will be driven by machine learning, will continue to be key enablers to advancing the capabilities of robotics. Computer vision Computer vision is currently the most prominent form of machine perception. It has been the sub-area of AI most transformed by the rise of deep learning. Until just a few years ago, support vector machines were the method of choice for most visual classification tasks. But the confluence of large-scale computing, especially on GPUs, the availability of large datasets, especially via the internet, and refinements of neural network algorithms has led to dramatic improvements in performance on benchmark tasks (e.g., classification on ImageNet17).

For the first time, computers are able to perform some (narrowly defined) visual classification tasks better than people. Much current research is focused on automatic image and video captioning. Natural Language Processing Often coupled with automatic speech recognition, Natural Language Processing is another very active area of machine perception. It is quickly becoming a commodity for mainstream languages with large data sets. Google announced that 20% of current mobile queries are done by voice,18 and recent demonstrations have proven the possibility of real-time translation. Research is now shifting towards developing refined and capable systems that are able to interact with people through dialog, not just react to stylized requests. 17 ImageNet, Stanford Vision Lab, Stanford University, Princeton University, 2016, accessed August 1, 2016.

http://www.image-net.org/

C4kTJRrWQAAMQu_

18 - Greg Sterling, “Google says 20% of mobile queries are voice searches,” Search Engine Land, May 18, 2016, accessed August 1, 2016,

Tighter coupling between hardware and software in advancing AI. “Intelligence” remains a complex phenomenon whose varied aspects have attracted the attention of several different fields of study, including psychology, economics, neuroscience, biology, engineering, statistics, and linguistics. Naturally, the field of AI has benefited from the progress made by all of these allied fields. For example, the artificial neural network, which has been at the heart of several AI-based solutions12 13 was originally inspired by thoughts about the flow of information in biological neurons.14 AI RESEARCH TRENDS Until the turn of the millennium, AI’s appeal lay largely in its promise to deliver, but in the last fifteen years, much of that promise has been redeemed.15 AI technologies already pervade our lives. As they becomes a central force in society, the field is shifting from simply building systems that are intelligent to building intelligent systems that are human-aware and trustworthy. Several factors have fueled the AI revolution. Foremost among them is the maturing of machine learning, supported in part by cloud computing resources and wide-spread, web-based data gathering. Machine learning has been propelled dramatically forward by “deep learning,” a form of adaptive artificial neural networks trained using a method called backpropagation.16 This leap in the performance of information processing algorithms has been accompanied by significant progress in hardware technology for basic operations such as sensing, perception, and object recognition. New platforms and markets for data-driven products, and the economic incentives to find new products and markets, have also contributed to the advent of AI-driven technology. All these trends drive the “hot” areas of research described below.

This compilation is meant simply to reflect the areas that, by one metric or another, currently receive greater attention than others. They are not necessarily more important or valuable than other ones. Indeed, some of the currently “hot” areas were less popular in past years, and it is likely that other areas will similarly re-emerge in the future. Large-scale machine learning Many of the basic problems in machine learning (such as supervised and unsupervised learning) are well-understood. A major focus of current efforts is to scale existing algorithms to work with extremely large data sets. For example, whereas traditional methods could afford to make several passes over the data set, modern ones are designed to make only a single pass; in some cases, only sublinear methods (those that only look at a fraction of the data) can be admitted. Deep learning The ability to successfully train convolutional neural networks has most benefited the field of computer vision, with applications such as object recognition, video 12 Gerald Tesauro, “Practical Issues in Temporal Difference Learning,” Machine Learning, no. 8 (1992): 257–77. 13 David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis, “Mastering the game of Go with deep neural networks and tree search,” Nature 529 (2016): 484–489. 14 W. McCulloch and W. Pitts, W., “A logical calculus of the ideas immanent in nervous activity,” Bulletin of Mathematical Biophysics, 5 (1943): 115–133. 15 Appendix I offers a short history of AI, including a description of some of the traditionally core areas of research, which have shifted over the past six decades.

16 - Backpropogation is an abbreviation for “backward propagation of errors,” a common method of training artificial neural networks used in conjunction with an optimization method such as gradient descent. The method calculates the gradient of a loss function with respect to all the weights in the network. Human intelligence has no match in the biological and artificial worlds for sheer versatility, with the abilities “to reason, achieve goals, understand and generate language... create art and music, and even write histories.” AI technologies already pervade our lives. As they become a central force in society, the field is shifting from simply building systems that are intelligent to building intelligent systems that are human-aware and trustworthy. 16 17 actively pursuing alternative models of computing—especially those that are inspired by what is known about biological neural networks—with the aim of improving the hardware efficiency and robustness of computing systems. At the moment, such “neuromorphic” computers have not yet clearly demonstrated big wins, and are just beginning to become commercially viable. But it is possible that they will become commonplace (even if only as additions to their von Neumann cousins) in the near future.

C5dBLk0XEAE74yL

Deep neural networks have already created a splash in the application landscape. A larger wave may hit when these networks can be trained and executed on dedicated neuromorphic hardware, as opposed to simulated on standard von Neumann architectures, as they are today. Overall trends and the future of AI research The resounding success of the data-driven paradigm has displaced the traditional paradigms of AI. Procedures such as theorem proving and logic-based knowledge representation and reasoning are receiving reduced attention, in part because of the ongoing challenge of connecting with real-world groundings. Planning, which was a mainstay of AI research in the seventies and eighties, has also received less attention of late due in part to its strong reliance on modeling assumptions that are hard to satisfy in realistic applications.

Model-based approaches—such as physics-based approaches to vision and traditional control and mapping in robotics—have by and large given way to data-driven approaches that close the loop with sensing the results of actions in the task at hand. Bayesian reasoning and graphical models, which were very popular even quite recently, also appear to be going out of favor, having been drowned by the deluge of data and the remarkable success of deep learning. Over the next fifteen years, the Study Panel expects an increasing focus on developing systems that are human-aware, meaning that they specifically model, and are specifically designed for, the characteristics of the people with whom they are meant to interact. There is a lot of interest in trying to find new, creative ways to develop interactive and scalable ways to teach robots. Also, IoT-type systems— devices and the cloud—are becoming increasingly popular, as is thinking about social and economic dimensions of AI.

In the coming years, new perception/object recognition capabilities and robotic platforms that are human-safe will grow, as will data-driven products and their markets. The Study Panel also expects a reemergence of some of the traditional forms of AI as practitioners come to realize the inevitable limitations of purely end-to-end deep learning approaches. We encourage young researchers not to reinvent the wheel, but rather to maintain an awareness of the significant progress in many areas of AI during the first fifty years of the field, and in related fields such as control theory, cognitive science, and psychology. Collaborative systems Research on collaborative systems investigates models and algorithms to help develop autonomous systems that can work collaboratively with other systems and with humans. This research relies on developing formal models of collaboration, and studies the capabilities needed for systems to become effective partners. There is growing interest in applications that can utilize the complementary strengths of humans and machines—for humans to help AI systems to overcome their limitations, and for agents to augment human abilities and activities. Crowdsourcing and human computation Since human abilities are superior to automated methods for accomplishing many tasks, research on crowdsourcing and human computation investigates methods to augment computer systems by utilizing human intelligence to solve problems that computers alone cannot solve well.

Introduced only about fifteen years ago, this research now has an established presence in AI. The best-known example of crowdsourcing is Wikipedia, a knowledge repository that is maintained and updated by netizens and that far exceeds traditionally-compiled information sources, such as encyclopedias and dictionaries, in scale and depth. Crowdsourcing focuses on devising innovative ways to harness human intelligence. Citizen science platforms energize volunteers to solve scientific problems, while paid crowdsourcing platforms such as Amazon Mechanical Turk provide automated access to human intelligence on demand. Work in this area has facilitated advances in other subfields of AI, including computer vision and NLP, by enabling large amounts of labeled training data and/or human interaction data to be collected in a short amount of time. Current research efforts explore ideal divisions of tasks between humans and machines based on their differing capabilities and costs. Algorithmic game theory and computational social choice New attention is being drawn to the economic and social computing dimensions of AI, including incentive structures. Distributed AI and multi-agent systems have been studied since the early 1980s, gained prominence starting in the late 1990s, and were accelerated by the internet. A natural requirement is that systems handle potentially misaligned incentives, including self-interested human participants or firms, as well as automated AI-based agents representing them.

C5f2CXwWUAAkrvM

Topics receiving attention include computational mechanism design (an economic theory of incentive design, seeking incentive-compatible systems where inputs are truthfully reported), computational social choice (a theory for how to aggregate rank orders on alternatives), incentive aligned information elicitation (prediction markets, scoring rules, peer prediction) and algorithmic game theory (the equilibria of markets, network games, and parlor games such as Poker—a game where significant advances have been made in recent years through abstraction techniques and no-regret learning). Internet of Things (IoT) A growing body of research is devoted to the idea that a wide array of devices can be interconnected to collect and share their sensory information. Such devices can include appliances, vehicles, buildings, cameras, and other things. While it’s a matter of technology and wireless networking to connect the devices, AI can process and use the resulting huge amounts of data for intelligent and useful purposes.

Currently, these devices use a bewildering array of incompatible communication protocols. AI could help tame this Tower of Babel. Neuromorphic Computing Traditional computers implement the von Neumann model of computing, which separates the modules for input/output, instruction-processing, and memory. With the success of deep neural networks on a wide array of tasks, manufacturers are A growing body of research is devoted to the idea that a wide array of devices can be interconnected to collect and share their sensory information. Such devices can include appliances, vehicles, buildings, cameras, and other things. Natural Language Processing is a very active area of machine perception. Research is now shifting towards developing systems that are able to interact with people through dialog, not just react to stylized requests. 18 19 They already had a number of functionalities that combined real-time sensing with perception and decision-making such as Anti-lock Braking Systems (ABS), airbag control, Traction Control Systems (TCS), and Electronic Stability Control (ESC).22 Automated capabilities have been introduced into commercial cars gradually since 2003 as summarized in the following table. Context Automated Functionality Release Date Parking Intelligent Parking Assist System Since 200323 Parking Summon Since 201624 Arterial & Highway Lane departure system Since 2004 in North America25 Arterial & Highway Adaptive cruise control Since 2005 in North America26 Highway Blind spot monitoring 200727 Highway Lane changing 201528 These functionalities assist drivers or completely take over well-defined activities for increased safety and comfort.

Current cars can park themselves, perform adaptive cruise control on highways, steer themselves during stop-and-go traffic, and alert drivers about objects in blind spots during lane changes. Vision and radar technology were leveraged to develop pre-collision systems that let cars autonomously brake when risk of a collision is detected. Deep learning also has been applied to improve automobiles’ capacity to detect objects in the environment and recognize sound. 29 Self-driving vehicles Since the 1930s, science fiction writers dreamed of a future with self-driving cars, and building them has been a challenge for the AI community since the 1960s. By the 2000s, the dream of autonomous vehicles became a reality in the sea and sky, and even on Mars, but self-driving cars existed only as research prototypes in labs.

C5dTFN-XQAAHAqg

Driving in a city was considered to be a problem too complex for automation due to factors like pedestrians, heavy traffic, and the many unexpected events that can happen outside of the car’s control. Although the technological components required to 22 Carl Liersch, “Vehicle Technology Timeline: From Automated to Driverless,” Robert Bosch (Australia) Pty. Ltd., 2014, accessed August 1, 2016,

23 - “Intelligent Parking Assist System,” Wikipedia, last modified July 26, 2016, accessed August 1, 2016,

https://en.wikipedia.org/wiki/Intelligent_Parking_Assist_System

24 - The Tesla Motors Team, “Summon Your Tesla from Your Phone,” Tesla, January 10, 2016, accessed August 1, 2016,

https://www.teslamotors.com/blog/summon-your-tesla-your-phone

25 - Lane departure warning system,” Wikipedia, last modified July 24, 2016, accessed August 1, 2016,

https://en.wikipedia.org/wiki/Lane_departure_warning_system

26 - “Autonomous cruise control system,” Wikipedia, last modified July 30, 2016, accessed August 1, 2016,

https://en.wikipedia.org/wiki/Autonomous_cruise_control_system

27 - “Blind spot monitor,” Wikipedia, last modified April 20, 2016, accessed August 1, 2016,

https://en.wikipedia.org/wiki/Blind_spot_monitor

28 - Dana Hull, “Tesla Starts Rolling Out Autopilot Features,” Boomberg Technology, October 14, 2015, accessed August 1, 2016,

29 - Aaron Tilley, “New Qualcomm Chip Brings Deep Learning To Cars,” Forbes, January 5, 2016, accessed August 1, 2016,

SECTION II: AI BY DOMAIN Though different instances of AI research and practice share common technologies, such as machine learning, they also vary considerably in different sectors of the economy and society. We call these sectors “domains,” and in this section describe the different states of AI research and implementation, as well as impacts and distinct challenges, in eight of them: transportation; home/service robotics; healthcare; education; low-resource communities; public safety and security; employment and workplace; and entertainment. Based on these analyses, we also predict trends in a typical North American city over the next fifteen years. Contrary to AI’s typical depiction in popular culture, we seek to offer a balanced overview of the ways in which AI is already beginning to transform everyday life, and how those transformations are likely to grow by the year 2030. TRANSPORTATION Transportation is likely to be one of the first domains in which the general public will be asked to trust the reliability and safety of an AI system for a critical task. Autonomous transportation will soon be commonplace and, as most people’s first experience with physically embodied AI systems, will strongly influence the public’s perception of AI.

Once the physical hardware is made sufficiently safe and robust, its introduction to daily life may happen so suddenly as to surprise the public, which will require time to adjust. As cars will become better drivers than people, city-dwellers will own fewer cars, live further from work, and spend time differently, leading to an entirely new urban organization. Further, in the typical North American city in 2030, changes won’t be limited to cars and trucks, but are likely to include flying vehicles and personal robots, and will raise social, ethical and policy issues. A few key technologies have already catalyzed the widespread adoption of AI in transportation. Compared to 2000, the scale and diversity of data about personal and population-level transportation available today—enabled by the adoption of smartphones and decreased costs and improved accuracies for variety of sensors—is astounding. Without the availability of this data and connectivity, applications such as real-time sensing and prediction of traffic, route calculations, peer-to-peer ridesharing and self-driving cars would not be possible.

Smarter cars GPS was introduced to personal vehicles in 2001 with in-car navigation devices and has since become a fundamental part of the transportation infrastructure.19 GPS assists drivers while providing large-scale information to technology companies and cities about transportation patterns. Widespread adoption of smartphones with GPS technology further increased connectivity and the amount of location data shared by individuals. Current vehicles are also equipped with a wide range of sensing capabilities. An average automobile in the US is predicted to have seventy sensors including gyroscopes, accelerometers, ambient light sensors, and moisture sensors.20 Sensors are not new to vehicles. Automobiles built before 2000 had sensors for the internal state of the vehicle such as its speed, acceleration, and wheel position.21 19 Mark Sullivan, “A brief history of GPS,” PCWorld, August 9, 2012, accessed August 1, 2016,

http://www.pcworld.com/article/2000276/a-brief-history-of-gps.html

20 - William J. Fleming, “New Automotive Sensors - A Review,” IEEE Sensors Journal 8, no 11, (2008): 1900-1921. 21 Jean Jacques Meneu, ed., “Automotive Sensors: Now and in the Future,” Arrow, September 24, 2015, accessed August 1, 2016,

https://www.arrow.com/en/research-and-events/articles/automotivesensors-now-and-in-the-future

As cars will become better drivers than people, citydwellers will own fewer cars, live further from work, and spend time differently, leading to an entirely new urban organization. Autonomous transportation will soon be commonplace and, as most people’s first experience with physically embodied AI systems, will strongly influence the public’s perception of AI. 20 21 commuter in US spends twenty-five minutes driving each way.37 With self-driving car technology, people will have more time to work or entertain themselves during their commutes. And the increased comfort and decreased cognitive load with self-driving cars and shared transportation may affect where people choose to live. The reduced need for parking may affect the way cities and public spaces are designed. Self-driving cars may also serve to increase the freedom and mobility of different subgroups of the population, including youth, elderly and disabled. Self-driving cars and peer-to-peer transportation services may eliminate the need to own a vehicle.

C5cbGr_UYAAE6yG

The effect on total car use is hard to predict. Trips of empty vehicles and people’s increased willingness to travel may lead to more total miles driven. Alternatively, shared autonomous vehicles—people using cars as a service rather than owning their own—may reduce total miles, especially if combined with well-constructed incentives, such as tolls or discounts, to spread out travel demand, share trips, and reduce congestion. The availability of shared transportation may displace the need for public transportation—or public transportation may change form towards personal rapid transit, already available in four cities,38which uses small capacity vehicles to transport people on demand and point-to-point between many stations.39 As autonomous vehicles become more widespread, questions will arise over their security, including how to ensure that technologies are safe and properly tested under different road conditions prior to their release. Autonomous vehicles and the connected transportation infrastructure will create a new venue for hackers to exploit vulnerabilities to attack. Ethical questions are also involved in programming cars to act in situations in which human injury or death is inevitable, especially when there are split-second choices to be made about whom to put at risk.

The legal systems in most states in the US do not have rules covering self-driving cars. As of 2016, four states in the US (Nevada, Florida, California, and Michigan), Ontario in Canada, the United Kingdom, France, and Switzerland have passed rules for the testing of self-driving cars on public roads. Even these laws do not address issues about responsibility and assignment of blame for an accident for self-driving and semi-selfdriving cars.40 Transportation planning By 2005, cities had started investing in the transportation infrastructure to develop sensing capabilities for vehicle and pedestrian traffic.41The sensors currently used include inductive loops, video cameras, remote traffic microwave sensors, radars, and GPS.42 For example, in 2013 New York started using a combination of microwave sensors, a network of cameras, and pass readers to detect vehicle traffic in the city.43 37 Brian McKenzie and Melanie Rapino, “Commuting in the United States: 2009,” American Community Survey Reports, United States Census Bureau, September 2011, accessed August 1, 2016,

https://www.census.gov/prod/2011pubs/acs-15.pdf

38 - Morgantown, West Virginia; Masdar City, UAE; London, England; and Suncheon, South Korea. 39 “Personal rapid transit,” Wikipedia, Last modified July 18, 2016, accessed August 1, 2016,

https://en.wikipedia.org/wiki/Personal_rapid_transit

40 - Patrick Lin, “The Ethics of Autonomous Cars,” The Atlantic, October 8, 2013, accessed August 1, 2016,

http://www.theatlantic.com/technology/archive/2013/10/the-ethics-ofautonomous-cars/280360/

41 - Steve Lohr, “Bringing Efficiency to the Infrastructure,” The New York Times, April 29, 2009, accessed August 1, 2016,

42 - “Intelligent transportation system,” Wikipedia, last modified July 28, 2016, accessed August 1, 2016,

https://en.wikipedia.org/wiki/Intelligent_transportation_system

43 - Access Science Editors, “Active traffic management: adaptive traffic signal control,” Access Science, 2014, accessed August 1, 2016,

http://www.accessscience.com/content/active-traffic-managementadaptive-traffic-signal-control/BR0106141

make such autonomous driving possible were available in 2000—and indeed some autonomous car prototypes existed30 31 32—few predicted that mainstream companies would be developing and deploying autonomous cars by 2015. During the first Defense Advanced Research Projects Agency (DARPA) “grand challenge” on autonomous driving in 2004, research teams failed to complete the challenge in a limited desert setting. But in eight short years, from 2004-2012, speedy and surprising progress occurred in both academia and industry. Advances in sensing technology and machine learning for perception tasks has sped progress and, as a result, Google’s autonomous vehicles and Tesla’s semi-autonomous cars are driving on city streets today. Google’s selfdriving cars, which have logged more than 1,500,000 miles (300,000 miles without an accident),33 are completely autonomous—no human input needed. Tesla has widely released self-driving capability to existing cars with a software update.34 Their cars are semi-autonomous, with human drivers expected to stay engaged and take over if they detect a potential problem.

It is not yet clear whether this semi-autonomous approach is sustainable, since as people become more confident in the cars’ capabilities, they are likely to pay less attention to the road, and become less reliable when they are most needed. The first traffic fatality involving an autonomous car, which occurred in June of 2016, brought this question into sharper focus.35 In the near future, sensing algorithms will achieve super-human performance for capabilities required for driving. Automated perception, including vision, is already near or at human-level performance for well-defined tasks such as recognition and tracking. Advances in perception will be followed by algorithmic improvements in higher level reasoning capabilities such as planning. A recent report predicts self-driving cars to be widely adopted by 2020.36 And the adoption of self-driving capabilities won’t be limited to personal transportation. We will see self-driving and remotely controlled delivery vehicles, flying vehicles, and trucks. Peer-to-peer transportation services such as ridesharing are also likely to utilize self-driving vehicles.

Beyond self-driving cars, advances in robotics will facilitate the creation and adoption of other types of autonomous vehicles, including robots and drones. It is not yet clear how much better self-driving cars need to become to encourage broad acceptance. The collaboration required in semi-self-driving cars and its implications for the cognitive load of human drivers is not well understood. But if future self-driving cars are adopted with the predicted speed, and they exceed human-level performance in driving, other significant societal changes will follow. Self-driving cars will eliminate one of the biggest causes of accidental death and injury in United States, and lengthen people’s life expectancy. On average, a 30 “Navlab,” Wikipedia, last updated June 4, 2016, accessed August 1, 2016,

31- “Navlab: The Carnegie Mellon University Navigation Laboratory,” Carnegie Mellon University, accessed August 1, 2016,

http://www.cs.cmu.edu/afs/cs/project/alv/www/

32 - “Eureka Prometheus Project,” Wikipedia, last modified February 12, 2016, accessed August 1, 2016, https://en.wikipedia.org/wiki/Eureka_Prometheus_Project. 33 “Google Self-Driving Car Project,” Google, accessed August 1, 2016,

33 - Molly McHugh, “Tesla’s Cars Now Drive Themselves, Kinda,” Wired, October 14, 2015, accessed August 1, 2016,

http://www.wired.com/2015/10/tesla-self-driving-overair-update-live/

34 - Molly McHugh, “Tesla’s Cars Now Drive Themselves, Kinda,” Wired, October 14, 2015, accessed August 1, 2016,

http://www.wired.com/2015/10/tesla-self-driving-over-air-update-live/

35 - Anjali Singhvi and Karl Russell, “Inside the Self-Driving Tesla Fatal Accident,” The New York Times, Last updated July 12, 2016, accessed August 1, 2016,

36 - John Greenough, “10 million self-driving cars will be on the road by 2020,” Business Insider, June 15, 2016, accessed August 1, 2016,

http://www.businessinsider.com/report-10-million-selfdriving-cars-will-be-on-the-road-by-2020-2015-5-6

Shared transportation may displace the need for public transportation— or public transportation may change form towards personal rapid transit that uses small capacity vehicles to transport people on demand. We will see self-driving and remotely controlled delivery vehicles, flying vehicles, and trucks. Peer-to-peer transportation services such as ridesharing are also likely to utilize selfdriving vehicles. 22 23 AI is likely to have an increasing impact on city infrastructure. Accurate predictive models of individuals’ movements, their preferences, and their goals are likely to emerge with the greater availability of data. The ethical issues regarding such an emergence are discussed in Section III of this report. The United States Department of Transportation released a call for proposals in 2016 asking medium-size cities to imagine smart city infrastructure for transportation.55 This initiative plans to award forty million dollars to a city to demonstrate how technology and data can be used to reimagine the movement of people as well as goods.

C5YDXqYVcAAeYlz

One vision is a network of connected vehicles that can reach a high level of safety in driving with car-to-car communication.56 If this vision becomes reality, we expect advances in multi-agent coordination, collaboration, and planning will have a significant impact on future cars and play a role in making the transportation system more reliable and efficient. Robots are also likely to take part in transportation by carrying individuals and packages (c.f., Segway robot). For transportation of goods, interest in drones has increased, and Amazon is currently testing a delivery system using them,57 although questions remain about the appropriate safety rules and regulations. The increased sensing capabilities, adoption of drones, and the connected transportation infrastructure will also raise concerns about the privacy of individuals and the safety of private data. In coming years, these and related transportation issues will need to be addressed either by preemptive action on the part of industry or within the legal framework. As noted in the Section III policy discussion, how well this is done will affect the pace and scope of AI-related advances in the transportation sector.

On-demand transportation On-demand transportation services such as Uber and Lyft have emerged as another pivotal application of sensing, connectivity, and AI,58 with algorithms for matching drivers to passengers by location and suitability (reputation modeling).59 60 Through dynamic pricing, these services ration access by willingness-to-pay, with dynamic pricing also encouraging an increase in the supply of drivers, and have become a popular method for transportation in cities. With their rapid advance have come multiple policy and legal issues, such as competition with existing taxi services and concerns about lack of regulation and safety. On-demand transportation services seem likely to be a major force towards self-driving cars. Carpooling and ridesharing have long been seen as a promising approach to decrease traffic congestion and better utilize personal transportation resources. Services such as Zimride and Nuride bring together people sharing similar routes for a joint trip. But this approach to carpooling has failed to gain traction on a large scale. 55 “U.S. Department of Transportation Launches Smart City Challenge to Create a City of the Future,” Transportation.gov, U.S. Department of Transportation, December 7, 2015, accessed August 1, 2016,

https://www.transportation.gov/briefing-room/us-department-transportationlaunches-smart-city-challenge-create-city-future

56 - Will Knight, “Car-to-Car Communication: A simple wireless technology promises to make driving much safer.,” MIT Technology Review, accessed August 1, 2016,

57 - “Amazon Prime Air,” Amazon, accessed August 1, 2016,

http://www.amazon.com/b?node=8037720011

58 - Jared Meyer, “Uber and Lyft are changing the way Americans move about their country,” National Review, June 7, 2016, accessed August 1, 2016,

59 - Alexander Howard, “How Digital Platforms Like LinkedIn, Uber And TaskRabbit Are Changing The On-Demand Economy,” The Huffington Post, July 14, 2015, accessed August 1, 2016,

http://www.huffingtonpost.com/entry/online-talent-platforms_us_55a03545e4b0b8145f72ccf6

60 - “Announcing UberPool,” Uber Newsroom, August 5, 2014, accessed August 1, 2016,

Cities use AI methods to optimize services in several ways, such as bus and subway schedules, and tracking traffic conditions to dynamically adjust speed limits or apply smart pricing on highways, bridges, and HOV lanes.44 45 46Using sensors and cameras in the road network, they can also optimize traffic light timing for improving traffic flow and to help with automated enforcement.47 48These dynamic strategies are aimed at better utilizing the limited resources in the transportation network, and are made possible by the availability of data and the widespread connectivity of individuals.

C5ceuKaWQAAYinv

Before the 2000s, transportation planners were forced to rely on static pricing strategies tied to particular days or times of day, to manage demand. As dynamic pricing strategies are adopted, this raises new issues concerning the fair distribution of public goods, since market conditions in high-demand situations may make services unavailable to segments of the public. The availability of large-scale data has also made transportation an ideal domain for machine learning applications. Since 2006, applications such as Mapquest, Google Maps, and Bing Maps have been widely used by the public for routing trips, using public transportation, receiving real-time information and predictions about traffic conditions, 49 50 and finding services around a location.51 52Optimal search algorithms have been applied to the routing of vehicles and pedestrians to a given destination (i.e.,53 54). Despite these advances, the widespread application of sensing and optimization techniques to city infrastructure has been slower than the application of these techniques to individual vehicles or people. Although individual cities have implemented sensing and optimization applications, as yet there is no standardization of the sensing infrastructure and AI techniques used. Infrastructure costs, differing priorities among cities, and the high coordination costs among the parties involved have slowed adoption, as have public concerns over privacy related to sensing. Still, 44 Kitae Jang, Koohong Chung, and Hwasoo Yeo, “A Dynamic Pricing Strategy for High Occupancy Toll Lanes,” Transportation Research Part A: Policy and Practice 67 (2014): 69–80. 45 “Seattle Variable Tolling Study,” City of Seattle Department of Transportation, May 2009, accessed August 1, 2016,

46 - James F. Peltz, “Dynamic Pricing Is Catching On in the Public and Private Sectors,” Government Technology, March 21, 2016, accessed August 1, 2016,

http://www.govtech.com/budgetfinance/Dynamic-Pricing-Is-Catching-On-in-the-Public-and-Private-Sectors.html

47 - Arthur G Sims and Kenneth W. Dobinson. “The Sydney Coordinated Adaptive Traffic (SCAT) System Philosophy and Benefits.” IEEE Transactions on Vehicular Technology 29, no. 2 (1980): 130–137.

48 - “New York City Launches Nation’s Most Sophisticated Active Traffic Management System Powered by TransCore’s TransSuite Traffic Management Software and RFID Technology,” Business Wire, September 27, 2009, accessed August 1, 2016,

49 - Eric Horvitz, Johnson Apacible, Raman Sarin, and Lin Liao, “Prediction, Expectation, and Surprise: Methods, Designs, and Study of a Deployed Traffic Forecasting Service,” Proceedings of the Twenty-First Conference on Uncertainty and Artificial Intelligence (2005) (Arlington, Virginia: AUAI Press, July 2005), 275–283.

50 - Timothy Hunter, Ryan Herring, Pieter Abbeel, and Alexandre Bayen, “Path and Travel Time Inference from GPS Probe Vehicle Data,” NIPS Analyzing Networks and Learning with Graphs 12, no. 1 (2009).

51 - John Krumm and Eric Horvitz, “Predestination: Inferring Destinations from Partial Trajectories,” UbiComp 2006: Ubiquitous Computing, Proceedings of the 8th International Conference, September 2006, (Springer Berlin, Heidelberg, 2006), 243–260.

52 - Jill Duffy, “Get Organized: Using Location-Based Reminders,” PC Magazine, June 30, 2014, accessed August 1, 2016,

http://www.pcmag.com/article2/0,2817,2460207,00.asp

53 - Robert J. Szczerba, Peggy Galkowski, I. S. Glicktein, and Noah Ternullo. “Robust Algorithm for Real-time Route Planning,” IEEE Transactions on Aerospace and Electronic Systems 36, no. 3 (2000): 869–878.

54 - Matt Duckham and Lars Kulik, “Simplest” Paths: Automated Route Selection for Navigation,” Spatial Information Theory. Foundations of Geographic Information Science, Proceedings of the International Conference, COSIT 2003, September 2003 (Springer-Verlag Berlin Heidelberg, 2003),

169-185. Our Study Panel doesn’t expect drones that can fly, swim, and drive, or flying quadcoptors to become a common means of transportation by 2030 (although prototypes exist today). Ethical questions arise when programming cars to act in situations in which human injury or death is inevitable, especially when there are split-second choices to be made about whom to put at risk. 24 25 for functions such as image labeling, and 3D object recognition, while common at AI conferences, are still only a few years into development as products. Home robots 2030 Despite the slow growth to date of robots in the home, there are signs that this will change in the next fifteen years. Corporations such as Amazon Robotics and Uber are developing large economies of scale using various aggregation technologies. Also: System in Module (SiM), with a lot of System on Chip (SoC) subsystems, are now being pushed out the door by phone-chip makers (Qualcomm’s SnapDragon, Samsung’s Artik, etc.). These are better than supercomputers of less than ten years ago with eight or more sixty-four-bit cores, and specialized silicon for cryptography, camera drivers, additional DSPs, and hard silicon for certain perceptual algorithms. This means that low cost devices will be able to support much more onboard AI than we have been able to consider over the last fifteen years. Cloud (“someone else’s computer”) is going to enable more rapid release of new software on home robots, and more sharing of data sets gathered in many different homes, which will in turn feed cloud-based machine learning, and then power improvements to already deployed robots.

The great advances in speech understanding and image labeling enabled by deep learning will enhance robots’ interactions with people in their homes. Low cost 3D sensors, driven by gaming platforms, have fueled work on 3D perception algorithms by thousands of researchers worldwide, which will speed the development and adoption of home and service robots. In the past three years, low cost and safe robot arms have been introduced to hundreds of research labs around the world, sparking a new class of research on manipulation that will eventually be applicable in the home, perhaps around 2025. More than half a dozen startups around the world are developing AI-based robots for the home, for now concentrating mainly on social interaction. New ethics and privacy issues may surface as a result. HEALTHCARE For AI technologies, healthcare has long been viewed as a promising domain. AI-based applications could improve health outcomes and quality of life for millions of people in the coming years—but only if they gain the trust of doctors, nurses, and patients, and if policy, regulatory, and commercial obstacles are removed. Prime applications include clinical decision support, patient monitoring and coaching, automated devices to assist in surgery or patient care, and management of healthcare systems. Recent successes, such as mining social media to infer possible health risks, machine learning to predict patients at risk, and robotics to support surgery, have expanded a sense of possibility for AI in healthcare. Improvements in methods for interacting with medical professionals and patients will be a critical challenge. As in other domains, data is a key enabler.

C5btemXWMAI7zL0

There has been an immense forward leap in collecting useful data from personal monitoring devices and mobile apps, from electronic health records (EHR) in clinical settings and, to a lesser extent, from robots designed to assist with medical procedures and hospital operations. But using this data to enable more finely-grained diagnostics and treatments for both individual patients and patient populations has proved difficult. Research and deployment have been slowed by outdated regulations and incentive structures. Poor human-computer interaction methods and the inherent difficulties and risks of implementing technologies in such a large and complex system have slowed realization of AI’s Interacting with people For decades, people have imagined wildly different, futuristic-looking transportation vehicles. Although future cars will be smarter and drones will be available widely, it is unlikely that by 2030 we will have widely adopted transportation vehicles that look and function differently than the ones we have today. Our Study Panel doesn’t expect drones that can fly, swim, and drive, or flying quadcoptors to become a common means of transportation in this time horizon (although prototypes exist today). We do expect humans to become partners to self-driving cars and drones in their training, execution, and evaluation. This partnering will happen both when humans are co-located with machines and also virtually. We predict advances in algorithms to facilitate machine learning from human input. We also expect models and algorithms for modeling of human attention, and to support communication and coordination between humans and machine.

This is an integral part of the development of future vehicles. HOME/SERVICE ROBOTS Robots have entered people’s homes in the past fifteen years. Disappointingly slow growth in the diversity of applications has occurred simultaneously with increasingly sophisticated AI deployed on existing applications. AI advances are often inspired by mechanical innovations, which in turn prompt new AI techniques to be introduced. Over the next fifteen years, coincident advances in mechanical and AI technologies promise to increase the safe and reliable use and utility of home robots in a typical North American city. Special purpose robots will deliver packages, clean offices, and enhance security, but technical constraints and the high costs of reliable mechanical devices will continue to limit commercial opportunities to narrowly defined applications for the foreseeable future. As with self-driving cars and other new transportation machines, the difficulty of creating reliable, market-ready hardware is not to be underestimated. Vacuum cleaners In 2001, after many years of development, the Electrolux Trilobite, a vacuum cleaning robot, became the first commercial home robot. It had a simple control system to do obstacle avoidance, and some navigation. A year later, iRobot introduced Roomba, which was a tenth the price of the Trilobite and, with only 512 bytes of RAM, ran a behavior based controller. The most intelligent thing it did was to avoid falling down stairs. Since then, sixteen million Roombas have been deployed all over the world and several other competing brands now exist. As the processing power and RAM capacity of low cost embedded processors improved from its dismal state in the year 2000, the AI capabilities of these robots also improved dramatically. Simple navigation, self-charging, and actions for dealing with full dust bins were added, followed by ability to deal with electrical cords and rug tassels, enabled by a combination of mechanical improvements and sensor based perception.

C8aq8W3XsAAlswL

More recently, the addition of full VSLAM (Visual Simultaneous Location and Mapping)— an AI technology that had been around for twenty years— has enabled the robots to build a complete 3D world model of a house as they clean, and become more efficient in their cleaning coverage. Early expectations that many new applications would be found for home robots have not materialized. Robot vacuum cleaners are restricted to localized flat areas, while real homes have lots of single steps, and often staircases; there has been very little research on robot mobility inside real homes. Hardware platforms remain challenging to build, and there are few applications that people want enough to buy. Perceptual algorithms Special purpose robots will deliver packages, clean offices, and enhance security, but technical constraints and high costs will continue to limit commercial opportunities for the foreseeable future. Over the next fifteen years, coincident advances in mechanical and AI technologies promise to increase the safe and reliable use and utility of home robots in a typical North American city.

26 27 - delivery as medical procedures and lifetime clinical records for hundreds of millions of individuals become available. Similarly, the automated capture of personal environmental data from wearable devices will expand personalized medicine.

These activities are becoming more commercially viable as vendors discover ways to engage large populations (e.g. ShareCare)63 and then to create population-scale data that can be mined to produce individualized analytics and recommendations. Unfortunately, the FDA has been slow to approve innovative diagnostic software, and there are many remaining barriers to rapid innovation. HIPAA (Health Insurance Portability and Accountability Act) requirements for protecting patient privacy create legal barriers to the flow of patient data to applications that could utilize AI technologies. Unanticipated negative effects of approved drugs could show up routinely, sooner, and more rigorously than they do today, but mobile apps that analyze drug interactions may be blocked from pulling the necessary information from patient records. More generally, AI research and innovation in healthcare are hampered by the lack of widely accepted methods and standards for privacy protection. The FDA has been slow to approve innovative software, in part due to an unclear understanding of the cost/benefit tradeoffs of these systems. If regulators (principally the FDA) recognize that effective post-marketing reporting is a dependable hedge against some safety risks, faster initial approval of new treatments and interventions may become possible. Automated image interpretation has also been a promising subject of study for decades. Progress on interpreting large archives of weakly-labeled images, such as large photo archives scraped from the web, has been explosive. At first blush, it is surprising that there has not been a similar revolution in interpretation of medical images. Most medical imaging modalities (CT, MR, ultrasound) are inherently digital, the images are all archived, and there are large, established companies with internal R&D (e.g. Siemens, Philips, GE) devoted to imaging. But several barriers have limited progress to date. Most hospital image archives have only gone digital over the past decade. More importantly, the problem in medicine is not to recognize what is in the image (is this a liver or a kidney?), but rather to make a fine-grained judgement about it (does the slightly darker smudge in the liver suggest a potentially cancerous tumor?).

Strict regulations govern these high-stakes judgements. Even with state-of-the-art technologies, a radiologist will still likely have to look at the images, so the value proposition is not yet compelling. Also, healthcare regulations preclude easy federation of data across institutions. Thus, only very large organizations of integrated care, such as Kaiser Permanente, are able to attack these problems. Still, automated/augmented image interpretation has started to gain momentum. The next fifteen years will probably not bring fully automated radiology, but initial forays into image “triage” or second level checking will likely improve the speed and costeffectiveness of medical imaging. When coupled with electronic patient record systems, large-scale machine learning techniques could be applied to medical image data. For example, multiple major healthcare systems have archives of millions of patient scans, each of which has an associated radiological report, and most have an associated patient record. Already, papers are appearing in the literature showing that deep neural networks can be trained to produce basic radiological findings, with high reliability, by training from this data.64 63 Sharecare, accessed August 1, 2016,

64 - Hoo-Chang Shin, Holger R. Roth, Mingchen Gao, Le Lu, Ziyue Xu, Isabella Nogues, Jianhua Yao, Daniel Mollura, and Ronald M. Summers, “Deep Convolutional Neural Networks for Computer-aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning,” IEEE Transactions on Medical Imaging 35, no. 5 (2016):

1285–1298 - promise in healthcare.61The reduction or removal of these obstacles, combined with innovations still on the horizon, have the potential to significantly improve health outcomes and quality of life for millions of people in the coming years.

C4llxPzWMAMWip8

The clinical setting For decades, the vision of an AI-powered clinician’s assistant has been a near cliché. Although there have been successful pilots of AI-related technology in healthcare,

62 - the current healthcare delivery system unfortunately remains structurally ill-suited to absorb and deploy rapid advances. Incentives provided by the Affordable Care Act have accelerated the penetration of electronic health records (EHRs) into clinical practice, but implementation has been poor, eroding clinicians’ confidence in their usefulness. A small group of companies control the EHR market, and user interfaces are widely considered substandard, including annoying pop-ups that physicians routinely dismiss. The promise of new analytics using data from EHRs, including AI, remains largely unrealized due to these and other regulatory and structural barriers. Looking ahead to the next fifteen years, AI advances, if coupled with sufficient data and well-targeted systems, promise to change the cognitive tasks assigned to human clinicians. Physicians now routinely solicit verbal descriptions of symptoms from presenting patients and, in their heads, correlate patterns against the clinical presentation of known diseases. With automated assistance, the physician could instead supervise this process, applying her or his experience and intuition to guide the input process and to evaluate the output of the machine intelligence.

The literal “hands-on” experience of the physician will remain critical. A significant challenge is to optimally integrate the human dimensions of care with automated reasoning processes. To achieve future advances, clinicians must be involved and engaged at the outset to ensure that systems are well-engineered and trusted. Already, a new generation of more tech savvy physicians routinely utilize specialized apps on mobile devices. At the same time, workloads on primary care clinicians have increased to the point that they are grateful for help from any quarter. Thus, the opportunity to exploit new learning methods, to create structured patterns of inference by mining the scientific literature automatically, and to create true cognitive assistants by supporting free-form dialogue, has never been greater. Provided these advances are not stymied by regulatory, legal, and social barriers, immense improvements to the value of healthcare are within our grasp. Healthcare analytics At the population level, AI’s ability to mine outcomes from millions of patient clinical records promises to enable finer-grained, more personalized diagnosis and treatment. Automated discovery of genotype-phenotype connections will also become possible as full, once-in-a-lifetime genome sequencing becomes routine for each patient. A related (and perhaps earlier) capability will be to find “patients like mine” as a way to inform treatment decisions based on analysis of a similar cohort.

Traditional and non-traditional healthcare data, augmented by social platforms, may lead to the emergence of self-defined subpopulations, each managed by a surrounding ecosystem of healthcare providers augmented with automated recommendation and monitoring systems. These developments have the potential to radically transform healthcare

61 - LeighAnne Olsen, Dara Aisner, and J. Michael McGinnis, eds., “Institute of Medicine (US) Roundtable on Evidence-Based Medicine,” The Learning Healthcare System: Workshop Summary. (Washington (DC): National Academies Press (US); 2007), accessed August 1, 2016,

62 - Katherine E. Henry, David N. Hager, Peter J. Pronovost, and Suchi Saria, “A Targeted Realtime Early Warning Score (TREWScore) for Septic Shock,” Science Translational Medicine 7, (299), 299ra122. A small group of companies control the EHR market, and user interfaces are widely considered substandard, including annoying pop-ups that physicians routinely dismiss. AI-based applications could improve health outcomes and quality of life for millions of people in the coming years—but only if they gain the trust of doctors, nurses, and patients.

28 - 29 be relatively simple once a patient is standing in a walker (though will certainly not be trivial for patients recovering from surgery and/or elderly patients, especially in corridors crowded with equipment and other people). Driving a needle to place a suture is relatively straightforward once the needle is correctly placed.

72 - This implies that many future systems will involve intimate interaction between people and machines and require technologies that facilitate collaboration between them. The growth of automation will enable new insights into healthcare processes. Historically, robotics has not been a strongly data-driven or data-oriented science. This is changing as (semi)automation infiltrates healthcare. As the new surgical, delivery, and patient care platforms come online, the beginnings of quantification and predictive analytics are being built on top of data coming from these platforms.

73 - This data will be used to assess quality of performance, identify deficiencies, errors, or potential optimizations, and will be used as feedback to improve performance. In short, these platforms will facilitate making the connection between what is done, and the outcome achieved, making true “closed-loop” medicine a real possibility.

Mobile health To date, evidence-driven analytics on healthcare have relied on traditional healthcare data—mainly the electronic medical records discussed above. In the clinical setting, there are hopeful trends towards bringing new data to bear. For example, TeleLanguage enables a human clinician to conduct language therapy sessions with multiple patients simultaneously with the aid of an AI agent trained by the clinician. And Lifegraph, which extracts behavioral patterns and creates alerts from data passively collected from a patient’s smartphone, has been adopted by psychiatrists in Israel to detect early signs of distressful behavior in patients. Looking ahead, driven by the mobile computing revolution, the astonishing growth of “biometrics in the wild”—and the explosion of platforms and applications that use them—is a hopeful and unanticipated trend. Thousands of mobile apps now offer information, introduce behavior modification, or identify groups of “people like me.” This, combined with the emerging trend of more specialized motion tracking devices, such as Fitbit, and the emerging (inter)connectedness between the home environment and health-monitoring devices, has created a vibrant new sector of innovation. By combining social and healthcare data, some healthcare apps can perform data mining, learning, and prediction from captured data, though their predictions are relatively rudimentary.

The convergence of data and functionality across applications will likely spur new and even obvious products, such as an exercise app that not only proposes a schedule for exercise but also suggests the best time to do it, and provides coaching to stick to that schedule.

72 - Azad Shademan, Ryan S. Decker, Justin D. Opfermann, Simon Leonard, Axel Krieger, and Peter CW Kim, “Supervised Autonomous Robotic Soft Tissue Surgery,” Science Translational Medicine 8, no. 337 (2016): 337ra64–337ra64. 73 Carolyn Chen, Lee White, Timothy Kowalewski, Rajesh Aggarwal, Chris Lintott, Bryan Comstock, Katie Kuksenok, Cecilia Aragon, Daniel Holst, and Thomas Lendvay, “Crowd-Sourced Assessment of Technical Skills: a novel method to evaluate surgical performance.” Journal of Surgical Research 187, no. 1 (2014):

65–71. - Healthcare robotics Fifteen years ago, healthcare robotics was largely science fiction. One company called Robodoc,

65 - a spin-out from IBM, developed robotic systems for orthopedic surgeries, such as hip and knee replacements. The technology worked, but the company struggled commercially, and was ultimately shut down and acquired for its technology.

66 - More recently, though, the research and practical use of surgical robotics has exploded. In 2000 Intuitive Surgical67 introduced the da Vinci system, a novel technology initially marketed to support minimally invasive heart bypass surgery, and then gained substantial market traction for treatment of prostate cancer and merged with its only major competition, Computer Motion, in 2003.

Sophia-Robot

The Da Vinci, now in its fourth generation, provides 3D visualization (as opposed to 2D monocular laparoscopy) and wristed instruments in an ergonomic platform. It is considered the standard of care in multiple laparoscopic procedures, and used in nearly three quarters of a million procedures a year,68 providing not only a physical platform, but also a new data platform for studying the process of surgery. The da Vinci anticipates a day when much greater insight into how medical professionals carry out the process of providing interventional medical care will be possible. The presence of the da Vinci in day-to-day operation has also opened the doors to new types of innovation—from new instrumentation to image fusion to novel biomarkers—creating its own innovation ecosystem. The success of the platform has inspired potential competitors in robotic surgery, most notably the Alphabet spin-off Verb, in collaboration with J&J/Ethicon.

69 - There are likely to be many more, each exploring a unique niche or space and building out an ecosystem of sensing, data analytics, augmentation, and automation. Intelligent automation in hospital operations has been less successful. The story is not unlike surgical robotics. Twenty years ago, one company, HelpMate, created a robot for hospital deliveries,

70 - such as meals and medical records, but ultimately went bankrupt. More recently, Aethon

71 - introduced TUG Robots for basic deliveries, but few hospitals have invested in this technology to date. However, robotics in other service industries such as hotels and warehouses, including Amazon Robotics (formerly Kiva), are demonstrating that these technologies are practical and cost effective in at least some large-scale settings, and may ultimately spur additional innovation in health care. Looking ahead, many tasks that appear in healthcare will be amenable to augmentation, but will not be fully automated.

For example, robots may be able to deliver goods to the right room in a hospital, but then require a person to pick them up and place them in their final location. Walking a patient down the corridor may 65 ROBODOC, accessed August 1, 2016,

http://www.robodoc.com/professionals.html

66 - THINK Surgical, accessed August 1, 2016,

http://thinksurgical.com/history

67 - Intuitive Surgical, accessed August 1, 2016,

http://www.intuitivesurgical.com

68 - Trefis Team, “Intuitive Surgical Maintains Its Growth Momentum With Strong Growth In Procedure Volumes,” Forbes, January 22, 2016, accessed August 1, 2016,

69 - Evan Ackerman, “Google and Johnson & Johnson Conjugate to Create Verb Surgical, Promise Fancy Medical Robots,” IEEE Spectrum, December 17, 2015, accessed August 1, 2016,

http://spectrum.ieee.org/automaton/robotics/medical-robots/google-verily-johnson-johnsonverb-surgical-medical-robots

70 - John M. Evans and Bala Krishnamurthy, “HelpMate®, the trackless robotic courier: A perspective on the development of a commercial autonomous mobile robot,” Lecture Notes in Control and Information Sciences 236, June 18, 2005 (Springer-Verlag London Limited, 1998), 182–210, accessed August 1, 2016,

http://link.springer.com/chapter/10.1007%2FBFb0030806

71 - Aethon, accessed August 1, 2016,

http://www.aethon.com

Specialized motion tracking devices... and the emerging (inter) connectedness between the home environment and health-monitoring devices have created a vibrant new sector of innovation. The problem in medicine is not to recognize what is in the image—is this a liver or a kidney?—but rather to make a fine-grained judgement about it. Strict regulations govern these high-stakes judgements.

30 31 - EDUCATION The past fifteen years have seen considerable AI advances in education. Applications are in wide use by educators and learners today, with some variation between K-12 and university settings. Though quality education will always require active engagement by human teachers, AI promises to enhance education at all levels, especially by providing personalization at scale. Similar to healthcare, resolving how to best integrate human interaction and face-to-face learning with promising AI technologies remains a key challenge. Robots have long been popular educational devices, starting with the early Lego Mindstorms kits developed with the MIT Media Lab in the 1980s. Intelligent Tutoring Systems (ITS) for science, math, language, and other disciplines match students with interactive machine tutors.

Natural Language Processing, especially when combined with machine learning and crowdsourcing, has boosted online learning and enabled teachers to multiply the size of their classrooms while simultaneously addressing individual students’ learning needs and styles. The data sets from large online learning systems have fueled rapid growth in learning analytics. Still, schools and universities have been slow in adopting AI technologies primarily due to lack of funds and lack of solid evidence that they help students achieve learning objectives. Over the next fifteen years in a typical North American city, the use of intelligent tutors and other AI technologies to assist teachers in the classroom and in the home is likely to expand significantly, as will learning based on virtual reality applications. But computer-based learning systems are not likely to fully replace human teaching in schools. Teaching robots Today, more sophisticated and versatile kits for use in K-12 schools are available from a number of companies that create robots with new sensing technologies programmable in a variety of languages.

Ozobot is a robot that teaches children to code and reason deductively while configuring it to dance or play based on color-coded patterns.

75 - Cubelets help teach children logical thinking through assembling robot blocks to think, act, or sense, depending upon the function of the different blocks.

76 - Wonder Workshop’s Dash and Dot span a range of programming capabilities. Children eight years old and older can create simple actions using a visual programming language, Blockly, or build iOS and Android applications using C or Java.

77 - PLEO rb is a robot pet that helps children learn biology by teaching the robot to react to different aspects of the environment.

78 - However, while fun and engaging for some, in order for such kits to become widespread, there will need to be compelling evidence that they improve students’ academic performance. Intelligent Tutoring Systems (ITS) and online learning ITS have been developed from research laboratory projects such as Why-2 Atlas, which supported human-machine dialogue to solve physics problems early in the era.

79 - The rapid migration of ITS from laboratory experimental stages to real use is 75 Ozobot, accessed August 1, 2016,

http://ozobot.com/

.76 - “Cubelets,” Modular Robotics, accessed August 1, 2016, http://www.modrobotics.com/ cubelets. 77 “Meet Dash,” Wonder Workshop, accessed August 1, 2016,

78 - “Pleo rb,” Innvo Labs, accessed August 1, 2016,

79 - Kurt VanLehn, Pamela W. Jordan, Carolyn P. Rosé, Dumisizwe Bhembe, Michael Böttner, Andy Gaydos, Maxim Makatchev, Umarani Pappuswamy, Michael Ringenberg, Antonio Roque, Stephanie Siler, and Ramesh Srivastava, “The Architecture of Why2-Atlas: A Coach for Qualitative Elder care Over the next fifteen years the number of elderly in the United States will grow by over 50%.

74 - The National Bureau of Labor Statistics projects that home health aides will grow 38% over the next ten years. Despite the broad opportunities in this domain—basic social support, interaction and communication devices, home health monitoring, a variety of simple in-home physical aids such as walkers, and light meal preparation—little has happened over the past fifteen years. But the coming generational shift will accompany a change in technology acceptance among the elderly. Currently, someone who is seventy was born in 1946 and may have first experienced some form of personalized IT in middle age or later, while a fifty-year-old today is far more technology-friendly and savvy. As a result, there will be a growing interest and market for already available and maturing technologies to support physical, emotional, social, and mental health.

Here are a few likely examples by category: Life quality and independence • Automated transportation will support continued independence and expanded social horizons. • Sharing of information will help families remain engaged with one another at a distance, and predictive analytics may be used to “nudge” family groups toward positive behaviors, such as reminders to “call home.” • Smart devices in the home will help with daily living activities when needed, such as cooking and, if robot manipulation capabilities improve sufficiently, dressing and toileting. Health and wellness • Mobile applications that monitor movement and activities, coupled with social platforms, will be able to make recommendations to maintain mental and physical health. • In-home health monitoring and health information access will be able to detect changes in mood or behavior and alert caregivers. • Personalized health management will help mitigate the complexities associated with multiple co-morbid conditions and/or treatment interactions. Treatments and devices • Better hearing aids and visual assistive devices will mitigate the effects of hearing and vision loss, improving safety and social connection. • Personalized rehabilitation and in-home therapy will reduce the need for hospital or care facility stays. • Physical assistive devices (intelligent walkers, wheel chairs, and exoskeletons) will extend the range of activities of an infirm individual. The Study Panel expects an explosion of low-cost sensing technologies that can provide substantial capabilities to the elderly in their homes. In principle, social agents with a physical presence and simple physical capabilities (e.g. a mobile robot with basic communication capabilities) could provide a platform for new innovations. However, doing so will require integration across multiple areas of AI—Natural Language Processing, reasoning, learning, perception, and robotics—to create a system that is useful and usable by the elderly.

violin

These innovations will also introduce questions regarding privacy within various circles, including friends, family, and care-givers, and create new challenges to accommodate an evermore active and engaged population far past retirement. 74 Jennifer M. Ortman, Victoria A. Velkoff, and Howard Hogan, “An Aging Nation: The Older Population in the United States: Population Estimates and Projections,” Current Population Reports, U.S Census Bureau (May 2014), accessed August 1, 2016,

Though quality education will always require active engagement by human teachers, AI promises to enhance education at all levels, especially by providing personalization at scale. Better hearing aids and visual assistive devices will mitigate the effects of hearing and vision loss, improving safety and social connection. Personalized rehabilitation and in-home therapy will reduce the need for hospital stays.

32 33 - also possible, such as those designed to assess vocabulary,86 wh (who/what/when/ where/why) questions,

87 - and multiple choice questions,88 using electronic resources such as WordNet, Wikipedia, and online ontologies. With the explosion of online courses, these techniques are sure to be eagerly adopted for use in online education. Although the long term impact of these systems will have on the educational system remains unclear, the AI community has learned a great deal in a very short time. Learning analytics Data sets being collected from massive scale online learning systems, ranging from MOOCs to Khan Academy, as well as smaller scale online programs, have fueled the rapid growth of the field of learning analytics. Online courses are not only good for widespread delivery, but are natural vehicles for data collection and experimental instrumentation that will contribute to scientific findings and improving the quality of learning at scale. Organizations such as the Society for Learning Analytics Research (SOLAR), and the rise of conferences including the Learning Analytics and Knowledge Conference

89 - and the Learning at Scale Conference (L@S)

90 - reflect this trend. This community applies deep learning, natural language processing, and other AI techniques to analysis of student engagement, behavior, and outcomes.

Current projects seek to model common student misconceptions, predict which students are at risk of failure, and provide real-time student feedback that is tightly integrated with learning outcomes. Recent work has also been devoted to understanding the cognitive processes involved in comprehension, writing, knowledge acquisition, and memory, and to applying that understanding to educational practice by developing and testing educational technologies. Challenges and opportunities One might have expected more and more sophisticated use of AI technologies in schools, colleges, and universities by now. Much of its absence can be explained by the lack of financial resources of these institutions as well as the lack of data establishing the technologies’ effectiveness. These problems are being addressed, albeit slowly, by private foundations and by numerous programs to train primarily secondary school teachers in summer programs. As in other areas of AI, excessive hype and promises about the capabilities of MOOCs have meant that expectations frequently exceed the reality.

The experiences of certain institutions, such as San Jose State University’s experiment with Udacity,91 have led to more sober assessment of the potential of the new educational technologies. 86 Jonathan C. Brown, Gwen A. Frishkoff , and Maxine Eskenazi, “Automatic Question Generation for Vocabulary Assessment,” Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), Vancouver, October 2005, (Association for Computational Linguistics, 2005),

819–826. 87 - Michael Heilman, “Automatic Factual Question Generation from Text,” PhD thesis CMULTI-11-004, (Carnegie Mellon University, 2011), accessed August 1, 2016,

88 - Tahani Alsubait, Bijan Parsia, and Uli Sattler, “Generating Multiple Choice Questions from Ontologies: How Far Can We Go?,” in eds. P. Lambrix, E. Hyvönen. E. Blomqvist, V. Presutti, G. Qi, U. Sattler, Y. Ding, and C. Ghidini, Knowledge Engineering and Knowledge Management: EKAW 2014 Satellite Events, VISUAL, EKM1, and ARCOE-Logic Linköping, Sweden, November 24–28, 2014 Revised Selected Papers, (Switzerland: Springer International Publishing, 2015),

66–79. 89 - The 6th International Learning Analytics & Knowledge Conference, accessed August 1, 2016,

http://lak16.solaresearch.org/

90 - Third Annual ACM Conference on Learning at Scale,

91 - Ry Rivard, “Udacity Project on ‘Pause’,” Inside Higher Ed, July 18, 2013, accessed August 1, 2016,

https://www.insidehighered.com/news/2013/07/18/citing-disappointing-student-outcomessan-jose-state-pauses-work-udacity

Surprising and welcome downloadable software and online systems such as Carnegie Speech or Duolingo provide foreign language training using Automatic Speech Recognition (ASR) and NLP techniques to recognize language errors and help users correct them.80 Tutoring systems such as the Carnegie Cognitive Tutor81 have been used in US high schools to help students learn mathematics. Other ITS have been developed for training in geography, circuits, medical diagnosis, computer literacy and programming, genetics, and chemistry. Cognitive tutors use software to mimic the role of a good human tutor by, for example, providing hints when a student gets stuck on a math problem. Based on the hint requested and the answer provided, the tutor offers context specific feedback. Applications are growing in higher education. An ITS called SHERLOCK

82 - is beginning to be used to teach Air Force technicians to diagnose electrical systems problems in aircraft. And the University of Southern California’s Information Sciences Institute has developed more advanced avatar-based training modules to train military personnel being sent to international posts in appropriate behavior when dealing with people from different cultural backgrounds. New algorithms for personalized tutoring, such as Bayesian Knowledge Tracing, enable individualized mastery learning and problem sequencing.83 Most surprising has been the explosion of the Massive Open Online Courses (MOOCs) and other models of online education at all levels—including the use of tools like Wikipedia and Khan Academy as well as sophisticated learning management systems that build in synchronous as well as asynchronous education and adaptive learning tools.

Since the late 1990s, companies such as the Educational Testing Service and Pearson have been developing automatic NLP assessment tools to co-grade essays in standardized testing.

84 - Many of the MOOCs which have become so popular, including those created by EdX, Coursera, and Udacity, are making use of NLP, machine learning, and crowdsourcing techniques for grading short-answer and essay questions as well as programming assignments.

85 - Online education systems that support graduate-level professional education and lifelong learning are also expanding rapidly. These systems have great promise because the need for face-to-face interaction is less important for working professionals and career changers. While not the leaders in AI-supported systems and applications, they will become early adopters as the technologies are tested and validated. It can be argued that AI is the secret sauce that has enabled instructors, particularly in higher education, to multiply the size of their classrooms by a few orders of magnitude—class sizes of a few tens of thousands are not uncommon. In order to continually test large classes of students, automated generation of the questions is Physics Essay Writing,” Intelligent Tutoring Systems: Proceedings of the 6th International Conference, (Springer Berlin Heidelberg, 2002),

158–167. 80 - VanLehn et al, “The Architecture of Why2-Atlas.” 81 “Resources and Support,” Carnegie Learning, accessed August 1, 2016,

82 - Alan Lesgold, Suzanne Lajoie, Marilyn Bunzo, and Gary Eggan, “SHERLOCK: A Coached Practice Environment for an Electronics Troubleshooting Job,” in J. H. Larkin and R. W. Chabay, eds., Computer-Assisted Instruction and Intelligent Tutoring Systems: Shared Goals and Complementary Approaches (Hillsdale, New Jersey: Lawrence Erlbaum Associates, 1988).

83 - Michael V. Yudelson, Kenneth R. Koedinger, and Geoffrey J. Gordon, (2013). “ Individualized Bayesian Knowledge Tracing Models,” Artificial Intelligence in Education, (Springer Berlin Heidelberg, 2013), 171–180.

84 - Jill Burstein, Karen Kukich, Susanne Wolff, Chi Lu, Martin Chodorow, Lisa Braden-Harder, and Mary Dee Harris, “Automated Scoring Using a Hybrid Feature Identification Technique” in Proceedings of the Annual Meeting of the Association of Computational Linguistics, Montreal, Canada, August 1998, accessed August 1, 2016,

https://www.ets.org/Media/Research/pdf/erater_acl98.pdf

85 - EdX,

https://www.edx.org/

Coursera,

https://www.coursera.org/

Udacity,

all accessed August 1, 2016. It can be argued that AI is the secret sauce that has enabled instructors, particularly in higher education, to multiply the size of their classrooms by a few orders of magnitude—class sizes of a few tens of thousands are not uncommon. The current absence of sophisticated use of AI technologies in schools, colleges, and universities may be explained by the lack of financial resources as well as the lack of data establishing the technologies’ effectiveness.

34 35 - education by providing tools and relatively simple amounts of training in their use. For example, large numbers of educational apps, many of them free, are being developed for the iPad. On the negative side, there is already a major trend among students to restrict their social contacts to electronic ones and to spend large amounts of time without social contact, interacting with online programs. If education also occurs more and more online, what effect will the lack of regular, face-to-face contact with peers have on students’ social development? Certain technologies have even been shown to create neurological side effects.

95 - On the other hand, autistic children have benefited from interactions with AI systems already.

96 - LOW-RESOURCE COMMUNITIES Many opportunities exist for AI to improve conditions for people in low-resource communities in a typical North American city—and, indeed, in some cases it already has. Understanding these direct AI contributions may also inform potential contributions in the poorest parts of the developing world. There has not been a significant focus on these populations in AI gatherings, and, traditionally, AI funders have underinvested in research lacking commercial application. With targeted incentives and funding priorities, AI technologies could help address the needs of low-resource communities. Budding efforts are promising. Counteracting fears that AI may contribute to joblessness and other societal problems, AI may provide mitigations and solutions, particularly if implemented in ways that build trust in them by the affected communities.

Machine learning, data mining approaches Under the banner of “data science for social good,” AI has been used to create predictive models to help government agencies more effectively use their limited budgets to address problems such as lead poisoning,

97 - a major public health concern that has been in the news due to ongoing events in Flint, Michigan. Children may be tested for elevated lead levels, but that unfortunately means the problem is only detected after they have already been poisoned. Many efforts are underway to use predictive models to assist government agencies in prioritizing children at risk, including those who may not yet have been exposed.

98 - Similarly, the Illinois Department of Human Services (IDHS) uses predictive models to identify pregnant women at risk for adverse birth outcomes in order to maximize the impact of prenatal care. The City of Cincinnati uses them to proactively identify and deploy inspectors to properties at risk of code violations. Scheduling, planning Task assignment scheduling and planning techniques have been applied by many different groups to distribute food before it spoils from those who may have excess, such as restaurants, to food banks, community centers and individuals.

99 - 95 - Scientist have studied, for example, the way reliance on GPS may lead to changes in the hypocampus. Kim Tingley, “The Secrets of the Wave Pilots,” The New York Times, March 17, 2016, accessed August 1, 2016,

96 - Judith Newman, “To Siri, With Love: How One Boy With Autism Became BFF With Apple’s Siri,” The New York Times, October 17, 2014, accessed August 1, 2016,

97 - Eric Potash, Joe Brew, Alexander Loewi, Subhabrata Majumdar, Andrew Reece, Joe Walsh, Eric Rozier, Emile Jorgensen, Raed Mansour, and Rayid Ghani, “Predictive Modeling for Public Health: Preventing Childhood Lead Poisoning,” Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (New York: Association for Computing Machinery, 2015), 2039–2047.

98 - Data Science for Social Good, University of Chicago, accessed August 1, 2016,

http://dssg.uchicago.edu/

99 - Senay Solak, Christina Scherrer, and Ahmed Ghoniem, “The Stop-and-Drop Problem in Nonprofit Food Distribution Networks,” Annals of Operations Research 221, no. 1 (October 2014): In the next fifteen years, it is likely that human teachers will be assisted by AI technologies with better human interaction, both in the classroom and in the home. The Study Panel expects that more general and more sophisticated virtual reality scenarios in which students can immerse themselves in subjects from all disciplines will be developed. Some steps in this direction are being taken now by increasing collaborations between AI researchers and researchers in the humanities and social sciences, exemplified by Stanford’s Galileo Correspondence Project

92 - and Columbia’s Making and Knowing Project.

93 - These interdisciplinary efforts create interactive experiences with historical documents and the use of Virtual Reality (VR) to explore interactive archeological sites.

94 - VR techniques are already being used in the natural sciences such as biology, anatomy, geology and astronomy to allow students to interact with environments and objects that are difficult to engage with in the real world. The recreation of past worlds and fictional worlds will become just as popular for studies of arts and other sciences. AI techniques will increasingly blur the line between formal, classroom education and self-paced, individual learning. Adaptive learning systems, for example, are going to become a core part of the teaching process in higher education because of the pressures to contain cost while serving a larger number of students and moving students through school more quickly. While formal education will not disappear, the Study Panel believes that MOOCs and other forms of online education will become part of learning at all levels, from K-12 through university, in a blended classroom experience.

This development will facilitate more customizable approaches to learning, in which students can learn at their own pace using educational techniques that work best for them. Online education systems will learn as the students learn, supporting rapid advances in our understanding of the learning process. Learning analytics, in turn, will accelerate the development of tools for personalized education. The current transition from hard copy books to digital and audio media and texts is likely to become prevalent in education as well. Digital reading devices will also become much ‘smarter’, providing students with easy access to additional information about subject matter as they study. Machine Translation (MT) technology will also make it easier to translate educational material into different languages with a fair degree of accuracy, just as it currently translates technical manuals. Textbook translation services that currently depend only upon human translators will increasingly incorporate automatic methods to improve the speed and affordability of their services for school systems. Online learning systems will also expand the opportunity for adults and working professionals to enhance their knowledge and skills (or to retool and learn a new field) in a world where these fields are evolving rapidly. This will include the expansion of fully online professional degrees as well as professional certifications based on online coursework. Broader societal consequences In countries where education is difficult for the broad population to obtain, online resources may have a positive effect if the population has the tools to access them.

The development of online educational resources should make it easier for foundations that support international educational programs to provide quality

92 - Stanford University: Galileo Correspondence Project, accessed August 1, 2016,

.93 - The Making and Knowing Project: Reconstructing the 16th Century Workshop of BNF MS. FR. 640 at Columbia University, accessed August 1, 2016,

http://www.makingandknowing.org

94 - Paul James, “3D Mapped HTC Vive Demo Brings Archaeology to Life,” Road to VR, August 31, 2015, accessed August 1, 2016,

http://www.roadtovr.com/3d-mapped-htc-vive-demo-bringsarchaeology-to-life/

While formal education will not disappear, the Study Panel believes that MOOCs and other forms of online education will become part of learning at all levels, from K-12 through university, in a blended classroom experience. With targeted incentives and funding priorities, AI technologies could help address the needs of lowresource communities. Budding efforts are promising.

36 - 37 - evidence of police malpractice. These improvements could lead to even more widespread surveillance. Some cities have already added drones for surveillance purposes, and police use of drones to maintain security of ports, airports, coastal areas, waterways, industrial facilities is likely to increase, raising concerns about privacy, safety, and other issues. The New York Police Department’s CompStat was the first tool pointing toward predictive policing,

104 - and many police departments now use it.

105 - Machine learning significantly enhances the ability to predict where and when crimes are more likely to happen and who may commit them. As dramatized in the movie Minority Report, predictive policing tools raise the specter of innocent people being unjustifiably targeted. But well-deployed AI prediction tools have the potential to actually remove or reduce human bias, rather than reinforcing it, and research and resources should be directed toward ensuring this effect. AI techniques can be used to develop intelligent simulations for training lawenforcement personnel to collaborate. While international criminal organizations and terrorists from different countries are colluding, police forces from different countries still face difficulty in joining forces to fight them. Training international groups of law enforcement personnel to work as teams is very challenging.

The European Union, through the Horizon 2020 program, currently supports such attempts in projects such as LawTrain.

106 - The next step will be to move from simulation to actual investigations by providing tools that support such collaborations. Tools do exist for scanning Twitter and other feeds to look for certain types of events and how they may impact security. For example, AI can help in social network analysis to prevent those at risk from being radicalized by ISIS or other violent groups. Law enforcement agencies are increasingly interested in trying to detect plans for disruptive events from social media, and also to monitor activity at large gatherings of people to analyze security. There is significant work on crowd simulations to determine how crowds can be controlled. At the same time, legitimate concerns have been raised about the potential for law enforcement agencies to overreach and use such tools to violate people’s privacy. The US Transportation Security Administration (TSA), Coast Guard, and the many other security agencies that currently rely on AI will likely increase their reliance to enable significant efficiency and efficacy improvements.

107 - AI techniques— vision, speech analysis, and gait analysis— can aid interviewers, interrogators, and security guards in detecting possible deception and criminal behavior. For example, the TSA currently has an ambitious project to redo airport security nationwide.

108 - Called DARMS, the system is designed to improve efficiency and efficacy of airport security by relying on personal information to tailor security based on a person’s risk categorization and the flights being taken.

The future vision for this project is a tunnel that checks people’s security while they walk through it. Once again, developers of this technology should be careful to avoid building in bias (e.g. about a person’s risk level category) through use of datasets that reflect prior bias.

109 - 104 - Walter L. Perry, Brian McInnis, Carter C. Price, Susan Smith, and John S. Hollywood, “The Role of Crime Forecasting in Law Enforcement Operations,” Rand Corporation Report 233 (2013).

105 - “CompStat,” Wikipedia, last modified July 28, 2016, accessed August 1, 2016,

106 - LAW-TRAIN, accessed August 1, 2016,

http://www.law-train.eu/

107 - Milind Tambe, Security and Game Theory: Algorithms, Deployed Systems, Lessons Learned (New York: Cambridge University Press, 2011).

108 - Peter Neffenger, “TSA’s 2017 Budget—A Commitment to Security (Part I),” Department of Homeland Security, March 1, 2016, accessed August 1, 2016,

109 - Crawford, “AI’s White Guy Problem.” Reasoning with social networks and influence maximization Social networks can be harnessed to create earlier, less-costly interventions involving large populations. For example, AI might be able to assist in spreading healthrelated information. In Los Angeles, there are more than 5,000 homeless youth (ages thirteen-twenty-four). Individual interventions are difficult and expensive, and the youths’ mistrust of authority dictates that key messages are best spread through peer leaders. AI programs might be able to leverage homeless youth social networks to strategically select peer leaders to spread health-related information, such as how to avoid spread of HIV. The dynamic, uncertain nature of these networks does pose challenges for AI research.

100 - Care must also be taken to prevent AI systems from reproducing discriminatory behavior, such as machine learning that identifies people through illegal racial indicators, or through highly-correlated surrogate factors, such as zip codes. But if deployed with great care, greater reliance on AI may well result in a reduction in discrimination overall, since AI programs are inherently more easily audited than humans.

PUBLIC SAFETY AND SECURITY Cities already have begun to deploy AI technologies for public safety and security. By 2030, the typical North American city will rely heavily upon them. These include cameras for surveillance that can detect anomalies pointing to a possible crime, drones, and predictive policing applications. As with most issues, there are benefits and risks. Gaining public trust is crucial. While there are legitimate concerns that policing that incorporates AI may become overbearing or pervasive in some contexts, the opposite is also possible. AI may enable policing to become more targeted and used only when needed. And assuming careful deployment, AI may also help remove some of the bias inherent in human decision-making. One of the more successful uses of AI analytics is in detecting white collar crime, such as credit card fraud.

101 - Cybersecurity (including spam) is a widely shared concern, and machine learning is making an impact. AI tools may also prove useful in helping police manage crime scenes or search and rescue events by helping commanders prioritize tasks and allocate resources, though these tools are not yet ready for automating such activities. Improvements in machine learning in general, and transfer learning in particular—for speeding up learning in new scenarios based on similarities with past scenarios—may facilitate such systems.

The cameras deployed almost everywhere in the world today tend to be more useful for helping solve crimes than preventing them.

102 - 103 - This is due to the low quality of event identification from videos and the lack of manpower to look at massive video streams. As AI for this domain improves, it will better assist crime prevention and prosecution through greater accuracy of event classification and efficient automatic processing of video to detect anomalies—including, potentially,

407–426. 100 - Jordan Pearson, “Artificial Intelligence Could Help Reduce HIV Among Homeless Youths,” Teamcore, University of Southern California, February 4. 2015, accessed August 1, 2016,

101 - “RSA Adaptive Authentication,” RSA, accessed August 1, 2016,

102 - Takeshi Arikuma and Yasunori Mochizuki, “Intelligent multimedia surveillance system for safer cities” APSIPA Transactions on Signal and Information Processing 5 (2016): 1–8.

103 - “Big Op-Ed: Shifting Opinions On Surveillance Cameras,”, Talk of the Nation, NPR, April 22, 2013, accessed August 1, 2016,

http://www.npr.org/2013/04/22/178436355/big-op-ed-shiftingopinions-on-surveillance-cameras

One of the more successful uses of AI analytics is in detecting white collar crime, such as credit card fraud. Cybersecurity (including spam) is a widely shared concern, and machine learning is making an impact. As dramatized in the movie Minority Report, predictive policing tools raise the specter of innocent people being unjustifiably targeted. But well-deployed AI prediction tools have the potential to actually remove or reduce human bias.

38 - 39 - markets, which often have the effect of lowering barriers to entry and increasing participation—from app stores to AirBnB to taskrabbit. A vibrant research community within AI studies further ways of creating new markets and making existing ones operate more efficiently. While work has intrinsic value, most people work to be able to purchase goods and services they value. Because AI systems perform work that previously required human labor, they have the effect of lowering the cost of many goods and services, effectively making everyone richer.

But as exemplified in current political debates, job loss is more salient to people—especially those directly affected—than diffuse economic gains, and AI unfortunately is often framed as a threat to jobs rather than a boon to living standards. There is even fear in some quarters that advances in AI will be so rapid as to replace all human jobs—including those that are largely cognitive or involve judgment—within a single generation. This sudden scenario is highly unlikely, but AI will gradually invade almost all employment sectors, requiring a shift away from human labor that computers are able to take over. The economic effects of AI on cognitive human jobs will be analogous to the effects of automation and robotics on humans in manufacturing jobs. Many middle-aged workers have lost well-paying factory jobs and the socio-economic status in family and society that traditionally went with such jobs. An even larger fraction of the total workforce may, in the long run, lose well-paying “cognitive” jobs. As labor becomes a less important factor in production as compared to owning intellectual capital, a majority of citizens may find the value of their labor insufficient to pay for a socially acceptable standard of living. These changes will require a political, rather than a purely economic, response concerning what kind of social safety nets should be in place to protect people from large, structural shifts in the economy. Absent mitigating policies, the beneficiaries of these shifts may be a small group at the upper stratum of the society.

113 - In the short run, education, re-training, and inventing new goods and services may mitigate these effects. Longer term, the current social safety net may need to evolve into better social services for everyone, such as healthcare and education, or a guaranteed basic income. Indeed, countries such as Switzerland and Finland have actively considered such measures.

AI may be thought of as a radically different mechanism of wealth creation in which everyone should be entitled to a portion of the world’s AI-produced treasure.

114 - It is not too soon for social debate on how the economic fruits of AI-technologies should be shared. As children in traditional societies support their aging parents, perhaps our artificially intelligent “children” should support us, the “parents” of their intelligence.

113 - For example, Brynjolfsson and McAfee, Second Machine Age, have two chapters of devoted to this (Erik Brynjolfsson and Andrew McAfee, The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies, (New York: W. W. Norton & Company, Inc., 2014)) and Brynjolfsson, McAfee, and Spence describe policy responses for the combination of globalization and digital technology (Erik Brynjolfsson, Andrew McAfee, and Michael Spence, Foreign Affairs, July/August 2014, accessed August 1, 2016.

114 - GDP does not do a good job of measuring the value of many digital goods. When society can’t manage what isn’t measured, bad policy decisions result. One alternative is to look at consumer surplus, not just dollar flows. As AI is embodied in more goods, this issue becomes more salient. It may look like GDP goes down but people have better well-being through access to these digital goods. See Erik Brynjolfsson and Adam Saunders, “What the GDP Gets Wrong (Why Managers Should Care),” Sloan Management Review, vol. 51, no. 1 (October 1, 2009):

95–96. - EMPLOYMENT AND WORKPLACE While AI technologies are likely to have a profound future impact on employment and workplace trends in a typical North American city, it is difficult to accurately assess current impacts, positive or negative. In the past fifteen years, employment has shifted due to a major recession and increasing globalization, particularly with China’s introduction to the world economy, as well as enormous changes in non-AI digital technology. Since the 1990s, the US has experienced continued growth in productivity and GDP, but median income has stagnated and the employment to population ratio has fallen. There are clear examples of industries in which digital technologies have had profound impacts, good and bad, and other sectors in which automation will likely make major changes in the near future. Many of these changes have been driven strongly by “routine” digital technologies, including enterprise resource planning, networking, information processing, and search. Understanding these changes should provide insights into how AI will affect future labor demand, including the shift in skill demands. To date, digital technologies have been affecting workers more in the skilled middle, such as travel agents, rather than the very lowest-skilled or highest skilled work.

110 - On the other hand, the spectrum of tasks that digital systems can do is evolving as AI systems improve, which is likely to gradually increase the scope of what is considered routine. AI is also creeping into high end of the spectrum, including professional services not historically performed by machines. To be successful, AI innovations will need to overcome understandable human fears of being marginalized. AI will likely replace tasks rather than jobs in the near term, and will also create new kinds of jobs. But the new jobs that will emerge are harder to imagine in advance than the existing jobs that will likely be lost. Changes in employment usually happen gradually, often without a sharp transition, a trend likely to continue as AI slowly moves into the workplace. A spectrum of effects will emerge, ranging from small amounts of replacement or augmentation to complete replacement. For example, although most of a lawyer’s job is not yet automated,

111 - AI applied to legal information extraction and topic modeling has automated parts of first-year lawyers’ jobs.

112 - In the not too distant future, a diverse array of job-holders, from radiologists to truck drivers to gardeners, may be affected. AI may also influence the size and location of the workforce. Many organizations and institutions are large because they perform functions that can be scaled only by adding human labor, either “horizontally” across geographical areas or “vertically” in management hierarchies. As AI takes over many functions, scalability no longer implies large organizations. Many have noted the small number of employees of some high profile internet companies, but not of others. There may be a natural scale of human enterprise, perhaps where the CEO can know everyone in the company. Through the creation of efficiently outsourced labor markets enabled by AI, enterprises may tend towards that natural size. AI will also create jobs, especially in some sectors, by making certain tasks more important, and create new categories of employment by making new modes of interaction possible. Sophisticated information systems can be used to create new

110 - Jeremy Ashkenas and Alicia Parlapiano, “How the Recession Reshaped the Economy, in 255 Charts,” The New York Times, June 6, 2014, accessed August 1, 2016,

111 - R Dana Remus and Frank S. Levy, “Can Robots Be Lawyers? Computers, Lawyers, and the Practice of Law,” Social Science Research Network, last modified February 12, 2016, accessed August 1, 2016,

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2701092

112 - John Markoff, “Armies of Expensive Lawyers, Replaced by Cheaper Software,” The New York Times, March 4, 2011, accessed August 1, 2016,

As labor becomes a less important factor in production as compared to owning intellectual capital, a majority of citizens may find the value of their labor insufficient to pay for a socially acceptable standard of living. AI will likely replace tasks rather than jobs in the near term, and will also create new kinds of jobs. But the new jobs that will emerge are harder to imagine in advance than the existing jobs that will likely be lost.

40 - 41 - Imagining the Future The success of any form of entertainment is ultimately determined by the individuals and social groups that are its subjects. The modes of entertainment that people find appealing are diverse and change over time. It is therefore hard to predict the forms entertainment will take in the next fifteen years precisely. Nevertheless, current trends suggest at least a few features that the future entertainment landscape is likely to contain. To date, the information revolution has mostly unfolded in software. However, with the growing availability of cheaper sensors and devices, greater innovation in the hardware used in entertainment systems is expected. Virtual reality and haptics could enter our living rooms—personalized companion robots are already being developed.

123 - With the accompanying improvements in Automatic Speech Recognition, the Study Panel expects that interaction with robots and other entertainment systems will become dialogue-based, perhaps constrained at the start, but progressively more human-like. Equally, the interacting systems are predicted to develop new characteristics such as emotion, empathy, and adaptation to environmental rhythms such as time of day.

124 - Today, an amateur with a video camera and readily-available software tools can make a relatively good movie. In the future, more sophisticated tools and apps will become available to make it even easier to produce high-quality content, for example, to compose music or to choreograph dance using an avatar. The creation and dissemination of entertainment will benefit from the progress of technologies such as ASR, dubbing, and Machine Translation, which will enable content to be customized to different audiences inexpensively. This democratization and proliferation of AIcreated media makes it difficult to predict how humans’ taste for entertainment, which are already fluid, will evolve. With content increasingly delivered digitally, and large amounts of data being logged about consumers’ preferences and usage characteristics, media powerhouses will be able to micro-analyze and micro-serve content to increasingly specialized segments of the population—down to the individual.

125 - Conceivably the stage is set for the emergence of media conglomerates acting as “Big Brothers” who are able to control the ideas and online experiences to which specific individuals are exposed. It remains to be seen whether broader society will develop measures to prevent their emergence. This topic, along with others pertaining to AI-related policy, is treated in more detail in the next section. 123 Emoters, accessed August 1, 2016,

http://emoterbots.com/

124 - “Siri,” Apple, Inc., accessed August 1, 2016,

http://www.apple.com/in/ios/siri/

125 - Ryan Calo, “Digital Market Manipulation,” George Washington Law Review 82, no. 4 (2014):

995–1051. - ENTERTAINMENT With the explosive growth of the internet over the past fifteen years, few can imagine their daily lives without it. Powered by AI, the internet has established user-generated content as a viable source of information and entertainment. Social networks such as Facebook are now pervasive, and they function as personalized channels of social interaction and entertainment—sometimes to the detriment of interpersonal interaction. Apps such as WhatsApp and Snapchat enable smart-phone users to remain constantly “in touch” with peers and share sources of entertainment and information. In on-line communities such as Second Life and role-playing games such as World of Warcraft, people imagine an alternative existence in a virtual world.

115 - Specialized devices, such as Amazon’s Kindle have also redefined the essentials of long-cherished pastimes. Books can now be browsed and procured with a few swipes of the finger, stored by the thousands in a pocket-sized device, and read in much the same way as a handheld paperback. Trusted platforms now exist for sharing and browsing blogs, videos, photos, and topical discussions, in addition to a variety of other user-generated information. To operate at the scale of the internet, these platforms must rely on techniques that are being actively developed in natural language processing, information retrieval, image processing, crowdsourcing, and machine learning. Algorithms such as collaborative filtering have been developed, for example, to recommend relevant movies, songs, or articles based on the user’s demographic details and browsing history

116 - Traditional sources of entertainment have also embraced AI to keep pace with the times. As exemplified in the book and movie Moneyball, professional sport is now subjected to intensive quantitative analysis.

117 - Beyond aggregate performance statistics, on-field signals can be monitored using sophisticated sensors and cameras. Software has been created for composing music

118 - and recognizing soundtracks.

119 - Techniques from computer vision and NLP have been used in creating stage performances.

120 - Even the lay user can exercise his or her creativity on platforms such as WordsEye, which automatically generates 3D scenes from natural language text.

121 - AI has also come to the aid of historical research in the arts, and is used extensively in stylometry and, more recently, in the analysis of paintings.

122 - The enthusiasm with which humans have responded to AI-driven entertainment has been surprising and led to concerns that it reduces interpersonal interaction among human beings. Few predicted that people would spend hours on end interacting with a display. Children often appear to be genuinely happier playing at home on their devices rather than outside with their friends. AI will increasingly enable entertainment that is more interactive, personalized, and engaging. Research should be directed toward understanding how to leverage these attributes for individuals’ and society’s benefit.

115 - Second Life, accessed August 1, 2016,

http://secondlife.com

“World of Warcraft,” Blizzard Entertainment, Inc, accessed August 1, 2016,

http://us.battle.net/wow/en/

116 - John S. Breese, David Heckerman, and Carl Kadie, “Empirical Analysis of Predictive Algorithms for Collaborative Filtering,” Proceedings of the 14th Conference on Uncertainty in Artificial Intelligence (July 1998), accessed August 1, 2016,

http://arxiv.org/pdf/1301.7363.pdf

43–52. 117 - Michael Lewis, Moneyball: The Art of Winning an Unfair Game (New York: W. W. Norton & Company, Inc., 2003):

http://www.imdb.com/title/tt1210166/

 118 - MuseScore, accessed August 1, 2016,

https://musescore.org/

119 - Shazam, accessed August 1, 2016,

http://www.shazam.com/

120 - Annie Dorsen, accessed August 1, 2016,

http://www.anniedorsen.com/

121 - WordsEye, accessed August 1, 2016,

https://www.wordseye.com/

122 - “Stylometry,” Wikipedia, last modified August 4, 2016, accessed August 1, 2016,

http://arxiv.org/pdf/1408.3218v1.pdf

AI will increasingly enable entertainment that is more interactive, personalized, and engaging. Research should be directed toward understanding how to leverage these attributes for individuals’ and society’s benefit. More sophisticated tools and apps will become available to make it even easier to produce high-quality content, for example, to compose music or to choreograph dance using an avatar.

42 - 43 - Likewise, AI could widen existing inequalities of opportunity if access to AI technologies—along with the high-powered computation and large-scale data that fuel many of them—is unfairly distributed across society. These technologies will improve the abilities and efficiency of people who have access to them. A person with access to accurate Machine Translation technology will be better able to use learning resources available in different languages. Similarly, if speech translation technology is only available in English, people who do not speak English will be at a disadvantage. Further, AI applications and the data they rely upon may reflect the biases of their designers and users, who specify the data sources. This threatens to deepen existing social biases, and concentrate AI’s benefits unequally among different subgroups of society. For example, some speech recognition technologies do not work well for women and people with accents. As AI is increasingly used in critical applications, these biases may surface issues of fairness to diverse groups in society. On the other hand, compared to the well-documented biases in human decision-making, AI-based decision-making tools have the potential to significantly reduce the bias in critical decisions such as who is lent money or sent to jail. Privacy concerns about AI-enabled surveillance are also widespread, particularly in cities with pervasive instrumentation. Sousveillance, the recording of an activity by a participant, usually with portable personal devices, has increased as well. Since views about bias and privacy are based on personal and societal ethical and value judgments, the debates over how to address these concerns will likely grow and resist quick resolution. Similarly, since AI is generating significant wealth, debates will grow regarding how the economic fruits of AI technologies should be shared—especially as AI expertise and the underlying data sets that fuel applications are concentrated in a small number of large corporations.

To help address these concerns about the individual and societal implications of rapidly evolving AI technologies, the Study Panel offers three general policy recommendations:

1. Define a path toward accruing technical expertise in AI at all levels of government. Effective governance requires more experts who understand and can analyze the interactions between AI technologies, programmatic objectives, and overall societal values. Absent sufficient technical expertise to assess safety or other metrics, national or local officials may refuse to permit a potentially promising application. Or insufficiently trained officials may simply take the word of industry technologists and green light a sensitive application that has not been adequately vetted. Without an understanding of how AI systems interact with human behavior and societal values, officials will be poorly positioned to evaluate the impact of AI on programmatic objectives.

2. Remove the perceived and actual impediments to research on the fairness, security, privacy, and social impacts of AI systems. Some interpretations of federal laws such as the Computer Fraud and Abuse Act and the anti-circumvention provision of the Digital Millennium Copyright Act are ambiguous regarding whether and how proprietary AI systems may be reverse engineered and evaluated by academics, journalists, and other researchers. Such research is critical if AI systems with physical and other material consequences are to be properly vetted and held accountable. 3. Increase public and private funding for interdisciplinary studies of the societal impacts of AI. As a society, we are underinvesting resources in research on the societal implications of AI technologies. Private and public dollars should be directed toward interdisciplinary

SECTION III: PROSPECTS AND RECOMMENDATIONS FOR AI PUBLIC POLICY

The goal of AI applications must be to create value for society. Our policy recommendations flow from this goal, and, while this report is focused on a typical North American city in 2030, the recommendations are broadly applicable to other places over time. Strategies that enhance our ability to interpret AI systems and participate in their use may help build trust and prevent drastic failures. Care must be taken to augment and enhance human capabilities and interaction, and to avoid discrimination against segments of society. Research to encourage this direction and inform public policy debates should be emphasized.

Given the current sector-specific regulation of US industries, new or retooled laws and policies will be needed to address the widespread impacts AI is likely to bring. Rather than “more” or “stricter” regulation, policies should be designed to encourage helpful innovation, generate and transfer expertise, and foster broad corporate and civic responsibility for addressing critical societal issues raised by these technologies. In the long term, AI will enable new wealth creation that will require social debate on how the economic fruits of AI technologies should be shared.

AI POLICY, NOW AND IN THE FUTURE Throughout history, humans have both shaped and adapted to new technologies. This report anticipates that advances in AI technologies will be developed and fielded gradually—not in sudden, unexpected jumps in the techniques themselves—and will build on what exists today, making this adaptation easier. On the other hand, small improvements to techniques, computing power, or availability of data can occasionally lead to novel, game-changing applications. The measure of success for AI applications is the value they create for human lives. Going forward, the ease with which people use and adapt to AI applications will likewise largely determine their success. Conversely, since AI applications are susceptible to errors and failures, a mark of their success will be how users perceive and tolerate their shortcomings. As AI becomes increasingly embedded in daily lives and used for more critical tasks, system mistakes may lead to backlash from users and negatively affect their trust. Though accidents in a self-driving car may be less probable than those driven by humans, for example, they will attract more attention. Design strategies that enhance the ability of humans to understand AI systems and decisions (such as explicitly explaining those decisions), and to participate in their use, may help build trust and prevent drastic failures. Likewise, developers should help manage people’s expectations, which will affect their happiness and satisfaction with AI applications. Frustration in carrying out functions promised by a system diminishes people’s trust and reduces their willingness to use the system in the future. Another important consideration is how AI systems that take over certain tasks will affect people’s affordances and capabilities. As machines deliver super-human performances on some tasks, people’s ability to perform them may wither. Already, introducing calculators to classrooms has reduced children’s ability to do basic arithmetic operations. Still, humans and AI systems have complementary abilities.

People are likely to focus on tasks that machines cannot do as well, including complex reasoning and creative expression. Already, children are increasingly exposed to AI applications, such as interacting with personal assistants on cell phones or with virtual agents in theme parks. Having early exposure will improve children’s interactions with AI applications, which will become a natural part of their daily lives. As a result, gaps will appear in how younger and older generations perceive AI’s influences on society. AI could widen existing inequalities of opportunity if access to AI technologies—along with the high-powered computation and largescale data that fuel many of them—is unfairly distributed across society. The measure of success for AI applications is the value they create for human lives. Going forward, the ease with which people use and adapt to AI applications will likewise largely determine their success.

44 45 - companies such as Google, Facebook, and Amazon have actively lobbied to avoid being designated as critical to the economy, arguing that this would open the door to regulation that would inevitably compromise their rapid product development cycles and ability to innovate.

130 - Nonetheless, as the companies creating, operating, and maintaining critical infrastructure use AI, interest will grow in regulating that software. Some existing regulatory regimes for software safety (for example, the FDA’s regulation of high consequence medical software) require specific software engineering practices at the developer level. However, modern software systems are often assembled from library components which may be supplied by multiple vendors, and are relatively application-independent. It doesn’t seem feasible or desirable to subject all such developers to the standards required for the most critical, rare applications. Nor does it seem advisable to allow unregulated use of such components in safety critical applications. Tradeoffs between promoting innovation and regulating for safety are difficult ones, both conceptually and in practice.

At a minimum, regulatory entities will require greater expertise going forward in order to understand the implications of standards and measures put in place by researchers, government, and industry.131 Policy and legal considerations While a comprehensive examination of the ways artificial intelligence (AI) interacts with the law is beyond the scope of this inaugural report, this much seems clear: as a transformative technology, AI has the potential to challenge any number of legal assumptions in the short, medium, and long term. Precisely how law and policy will adapt to advances in AI—and how AI will adapt to values reflected in law and policy—depends on a variety of social, cultural, economic, and other factors, and is likely to vary by jurisdiction. American law represents a mixture of common law, federal, state, and local statutes and ordinances, and—perhaps of greatest relevance to AI—regulations. Depending on its instantiation, AI could implicate each of these sources of law. For example, Nevada passed a law broadly permitting autonomous vehicles and instructed the Nevada Department of Motor Vehicles to craft requirements. Meanwhile, the National Highway Transportation Safety Administration has determined that a selfdriving car system, rather than the vehicle occupants, can be considered the “driver” of a vehicle. Some car designs sidestep this issue by staying in autonomous mode only when hands are on the wheel (at least every so often), so that the human driver has ultimate control and responsibility. Still, Tesla’s adoption of this strategy did not prevent the first traffic fatality involving an autonomous car, which occurred in June of 2016. Such incidents are sure to influence public attitudes towards autonomous driving. And as most people’s first experience with embodied agents, autonomous transportation will strongly influence the public’s perception of AI. Driverless cars are, of course, but one example of the many instantiations of AI in services, products, and other contexts.

The legal effect of introducing AI into the provision of tax advice, automated trading on the stock market, or generating medical diagnoses will also vary in accordance to the regulators that govern these contexts and the rules that apply within them. Many other examples of AI applications fall within current non-technology-specific policy, including predictive

130 - Eric Engleman, “Google Exception in Obama’s Cyber Order Questioned as Unwise Gap,” Bloomberg Technology, March 4, 2013, accessed August 1, 2016

131 - Ryan Calo, “The Case for a Federal Robotics Commission,” Brookings Report, September 15, 2014, accessed August 1, 2016,

http://www.brookings.edu/research/reports2/2014/09/case-forfederal-robotics-commission

Teams capable of analyzing AI from multiple angles. Research questions range from basic research into intelligence to methods to assess and affect the safety, privacy, fairness, and other impacts of AI. Questions include: Who is responsible when a self-driven car crashes or an intelligent medical device fails? How can AI applications be prevented from unlawful discrimination? Who should reap the gains of efficiencies enabled by AI technologies and what protections should be afforded to people whose skills are rendered obsolete? As AI becomes integrated more broadly and deeply into industrial and consumer products, it enters areas in which established regulatory regimes will need to be adapted to AI innovations or in some cases fundamentally reconfigured according to broadly accepted goals and principles. The approach in the United States to date has been sector-specific, with oversight by a variety of agencies. The use of AI in devices that deliver medical diagnostics and treatments is subject to aggressive regulation by the Food and Drug Administration (FDA), both in defining what the product is and specifying the methods by which it is produced, including standards of software engineering. The use of drones in regulated airspace falls under the authority of the Federal Aviation Administration (FAA).

126 - For consumer-facing AI systems, regulation by the Federal Trade Commission (FTC) comes into play. Financial markets using AI technologies, such as in high-frequency trading, come under regulation by the Security Exchange Commission (SEC). In addition to sector-specific approaches, the somewhat ambiguous and broad regulatory category of “critical infrastructure” may apply to AI applications.

127 - The Obama Administration’s Presidential Policy Directive (PPD)

21 - broadly defines critical infrastructure as composed of “the assets, systems, and networks, whether physical or virtual, so vital to the United States that their incapacitation or destruction would have a debilitating effect on security, national economic security, national public health or safety, or any combination thereof.” Today, an enterprise does not come under federal regulation solely by falling under that broad definition. Instead, the general trend of federal policy is to seek regulation in sixteen sectors of the economy.

128 - As regards AI, critical infrastructure is notably defined by the end-user application, and not the technology or sector that actually produces AI software.

129 - Software

126 - FAA controls the ways drones fly, requires drones to be semiautonomous as opposed to autonomous, requires visual connection to the drone, and enforces no-fly zones close to airports.

127 - “Presidential Policy Directive (PPD-21)—Critical Infrastructure Security and Resilience,” The White House, February 12, 2013, accessed August 1, 2016,

https://www.whitehouse.gov/the-pressoffice/2013/02/12/presidential-policy-directive-critical-infrastructure-security-and-resil

128 - PPD 21 identifies agencies responsible in each case. Chemical: Department of Homeland Security; Commercial Facilities: Department of Homeland Security; Communications: Department of Homeland Security; Critical Manufacturing: Department of Homeland Security; Dams: Department of Homeland Security; Defense Industrial Base: Department of Defense; Emergency Services: Department of Homeland Security; Energy: Department of Energy; Financial Services: Department of the Treasury; Food and Agriculture: U.S. Department of Agriculture and Department of Health and Human Services; Government Facilities: Department of Homeland Security and General Services Administration; Healthcare and Public Health: Department of Health and Human Services; Information Technology: Department of Homeland Security; Nuclear Reactors, Materials, and Waste: Department of Homeland Security; Transportation Systems: Department of Homeland Security and Department of Transportation; Water and Wastewater Systems: Environmental Protection Agency.

129 - In “ICYMI- Business Groups Urge White House to Rethink Cyber Security Order,” Internet Association, March 5, 2013, accessed August 1, 2016,

https://internetassociation.org/030513gov-3/

“Obama’s Feb. 12 order says the government can’t designate ‘commercial information technology products’ or consumer information technology services as critical U.S. infrastructure targeted for voluntary computer security standards,’ ... ‘Obama’s order isn’t meant to get down to the level of products and services and dictate how those products and services behave,’ said David LeDuc, senior director of public policy for the Software & Information Industry Association, a Washington trade group that lobbied for the exclusions.” As a society, we are underinvesting resources in research on the societal implications of AI technologies. Private and public dollars should be directed toward interdisciplinary teams capable of analyzing AI from multiple angles. Absent sufficient technical expertise to assess safety or other metrics, national or local officials may refuse to permit a potentially promising application—or green light a sensitive application that has not been adequately vetted.

46 - 47 - of product liability—and the responsibility that falls to companies manufacturing these products— will likely grow when human actors become less responsible for the actions of a machine. Liability (criminal) If tort law expects harms to be foreseeable, criminal law goes further to expect that harms be intended. US law in particular attaches great importance to the concept of mens rea—the intending mind. As AI applications engage in behavior that, were it done by a human, would constitute a crime, courts and other legal actors will have to puzzle through whom to hold accountable and on what theory. Agency The above issues raise the question of whether and under what circumstances an AI system could operate as the agent of a person or corporation. Already regulatory bodies in the United States, Canada, and elsewhere are setting the conditions under which software can enter into a binding contract.

135 - The more AI conducts legally salient activities, the greater the challenge to principles of agency under the law. Certification The very notion of “artificial intelligence” suggests a substitution for human skill and ingenuity. And in many contexts, ranging from driving to performing surgery or practicing law, a human must attain some certification or license before performing a given task. Accordingly, law and policy will have to—and already does—grapple with how to determine competency in an AI system. For example, imagine a robotics company creates a surgical platform capable of autonomously removing an appendix. Or imagine a law firm writes an application capable of rendering legal advice. Today, it is unclear from a legal perspective who in this picture would have to pass the medical boards or legal bar, let alone where they would be required to do so.

136 - Labor As AI substitutes for human roles, some jobs will be eliminated and new jobs will be created. The net effect on jobs is ambiguous, but labor markets are unlikely to benefit everyone evenly. The demand for some types of skills or abilities will likely drop significantly, negatively affecting the employment levels and wages of people with those skills.

137 - While the ultimate effects on income levels and distribution are not inevitable, they depend substantially on government policies, on the way companies choose to organize work, and on decisions by individuals to invest in learning new skills and seeking new types of work and income opportunities. People who find their employment altered or terminated as a consequence of advances of AI may seek recourse in the legislature and courts.

This may be why Littler Mendelson LLP— perhaps the largest employment law firm in the world—has an entire practice group to address robotics and artificial intelligence. Taxation Federal, state, and local revenue sources may be affected. Accomplishing a task using AI instead of a person can be faster and more accurate—and avoid employment taxes. As a result, AI applications could increasingly shift investment from payroll and income to capital expenditure. Depending on a state budget’s reliance on payroll and income tax, such a shift could be destabilizing. AI may also display different

135 - Ian R. Kerr, “Ensuring the Success of Contract Formation in Agent-Mediated Electronic Commerce,” Electronic Commerce Research 1 (2001):

183–202. 136 - Ryan Calo, “Digital Agenda’s public discussion on ‘The effects of robotics on economics, labour and society,’” Ausschuss Digitale Agenda, (Deutsche Bundestag: Ausschussdrucksache A-Drs. 18(24)102), June 22, 2016, accessed August 1, 2016,

137 - Brynjolfsson and McAfee, “Race Against the Machine: How the Digital Revolution is Accelerating Innovation, Driving Productivity, and Irreversibly Transforming Employment and the Economy (2011)”; Brynjolfsson and McAfee, Second Machine Age. policing, non-discriminatory loans, healthcare applications such as eldercare and drug delivery, systems designed to interact with children (for example, autonomous tutoring systems are required to respect laws in regard to balanced handling of evolution vs. intelligent design), and interactive entertainment. Given the present structure of American administrative law, it seems unlikely that AI will be treated comprehensively in the near term. Nevertheless, it is possible to enumerate broad categories of legal and policy issues that AI tends to raise in various contexts. Privacy Private information about an individual can be revealed through decisions and predictions made by AI. While some of the ways that AI implicates privacy mirror those of technologies such as computers and the internet, other issues may be unique to AI. For example, the potential of AI to predict future behavior based on previous patterns raises challenging questions. Companies already use machine learning to predict credit risk. And states run prisoner details through complex algorithms to predict the likelihood of recidivism when considering parole. In these cases, it is a technical challenge to ensure that factors such as race and sexual orientation are not being used to inform AI-based decisions. Even when such features are not directly provided to the algorithms, they may still correlate strongly with seemingly innocuous features such as zip code. Nonetheless, with careful design, testing, and deployment, AI algorithms may be able to make less biased decisions than a typical person. Anthropomorphic interfaces increasingly associated with AI raise novel privacy concerns. Social science research suggests people are hardwired to respond to anthropomorphic technology as though it were human. Subjects in one study were more likely to answer when they were born if the computer first stated when it was built.

132 - In another, they skipped sensitive questions when posed by an anthropomorphic interface.

133 - At a basic level lies the question: Will humans continue to enjoy the prospect of solitude in a world permeated by apparently social agents “living” in our houses, cars, offices, hospital rooms, and phones?

134 - Innovation policy Early law and policy decisions concerning liability and speech helped ensure the commercial viability of the Internet. By contrast, the software industry arguably suffers today from the decision of firms to pivot from open and free software to the more aggressive pursuit of intellectual property protections, resulting in what some have termed patent “thickets.” Striking the proper balance between incentivizing innovation in AI while promoting cooperation and protection against third party harm will prove a central challenge. Liability (civil) As AI is organized to directly affect the world, even physically, liability for harms caused by AI will increase in salience. The prospect that AI will behave in ways designers do not expect challenges the prevailing assumption within tort law that courts only compensate for foreseeable injuries. Courts might arbitrarily assign liability to a human actor even when liability is better located elsewhere for reasons of fairness or efficiency. Alternatively, courts could refuse to find liability because the defendant before the court did not, and could not, foresee the harm that the AI caused. Liability would then fall by default on the blameless victim.

The role

132 - Youngme Moon, “Intimate Exchanges: Using Computers to Elicit Self-Disclosure from Consumers,” Journal of Consumer Research 26, no. 4 (March 2000): 323–339.

133 - Lee Sproull, Mani Subramani, Sara Kiesler , Janet H. Walker, and Keith Waters, “When the Interface is a Face,” Human-Computer Interaction 11, no. 2 (1996):

97–124. 134 - M. Ryan Calo, “People Can Be So Fake: A New Dimension to Privacy and Technology Scholarship,” Penn State Law Review 114, no. 3 (2010):

809–855. - As AI applications engage in behavior that, were it done by a human, would constitute a crime, courts and other legal actors will have to puzzle through whom to hold accountable and on what theory. AI applications could increasingly shift investment from payroll and income to capital expenditure. Depending on a state budget’s reliance on payroll and income tax, such a shift could be destabilizing.

48 49 - privacy as their responsibility. Broad legal mandates encouraged companies to develop a professional staff and processes to enforce privacy controls, engage with outside stakeholders, and to adapt their practices to technology advances. Requiring greater transparency enabled civil society groups and media to become credible enforcers both in court and in the court of public opinion, making privacy more salient to corporate boards and leading them to further invest in privacy protection. In AI, too, regulators can strengthen a virtuous cycle of activity involving internal and external accountability, transparency, and professionalization, rather than narrow compliance. As AI is integrated into cities, it will continue to challenge existing protections for values such as privacy and accountability. Like other technologies, AI has the potential to be used for good or nefarious purposes. This report has tried to highlight the potential for both. A vigorous and informed debate about how to best steer AI in ways that enrich our lives and our society, while encouraging creativity in the field, is an urgent and vital need. Policies should be evaluated as to whether they democratically foster the development and equitable sharing of AI’s benefits, or concentrate power and benefits in the hands of a fortunate few. And since future AI technologies and their effects cannot be foreseen with perfect clarity, policies will need to be continually re-evaluated in the context of observed societal challenges and evidence from fielded systems. As this report documents, significant AI-related advances have already had an impact on North American cities over the past fifteen years, and even more substantial developments will occur over the next fifteen. Recent advances are largely due to the growth and analysis of large data sets enabled by the Internet, advances in sensory technologies and, more recently, applications of “deep learning.” In the coming years, as the public encounters new AI applications in domains such as transportation and healthcare, they must be introduced in ways that build trust and understanding, and respect human and civil rights. While encouraging innovation, policies and processes should address ethical, privacy, and security implications, and should work to ensure that the benefits of AI technologies will be spread broadly and fairly. Doing so will be critical if Artificial Intelligence research and its applications are to exert a positive influence on North American urban life in 2030 and beyond. “habits” than people, resulting in still fewer revenue sources.

The many municipalities relying on income from speeding or parking tickets will have to find alternatives if autonomous cars can drop people off and find distance parking, or if they are programmed not to violate the law. As a result, government bodies trying to balance their budgets in light of advances in AI may pass legislation to slow down or alter the course of the technology. Politics AI technologies are already being used by political actors in gerrymandering and targeted “robocalls” designed to suppress votes, and on social media platforms in the form of “bots.”

138 - They can enable coordinated protest as well as the ability to predict protests, and promote greater transparency in politics by more accurately pinpointing who said what, when. Thus, administrative and regulatory laws regarding AI can be designed to promote greater democratic participation or, if ill-conceived, to reduce it. This list is not exhaustive and focuses largely on domestic policy in the United States, leaving out many areas of law that AI is likely to touch. One lesson that might be drawn concerns the growing disconnect between the context-specific way in which AI is governed today and a wider consideration of themes shared by AI technologies across industries or sectors of society. It could be tempting to create new institutional configurations capable of amassing expertise and setting AI standards across multiple contexts. The Study Panel’s consensus is that attempts to regulate “AI” in general would be misguided, since there is no clear definition of AI (it isn’t any one thing), and the risks and considerations are very different in different domains. Instead, policymakers should recognize that to varying degrees and over time, various industries will need distinct, appropriate, regulations that touch on software built using AI or incorporating AI in some way. The government will need the expertise to scrutinize standards and technology developed by the private and public sector, and to craft regulations where necessary. Guidelines for the future Faced with the profound changes that AI technologies can produce, pressure for “more” and “tougher” regulation is inevitable. Misunderstanding about what AI is and is not, especially against a background of scare-mongering, could fuel opposition to technologies that could benefit everyone. This would be a tragic mistake. Regulation that stifles innovation, or relocates it to other jurisdictions, would be similarly counterproductive. Fortunately, principles that guide successful regulation of current digital technologies can be instructive. A recent multi-year study comparing privacy regulation in four European countries and the United States, for example, yielded counter-intuitive results.

139 - Those countries, such as Spain and France, with strict and detailed regulations bred a “compliance mentality” within corporations, which had the effect of discouraging both innovation and robust privacy protections. Rather than taking responsibility for privacy protection internally and developing a professional staff to foster it in business and manufacturing processes, or engaging with privacy advocates or academics outside their walls, these companies viewed privacy as a compliance activity. Their focus was on avoiding fines or punishments, rather than proactively designing technology and adapting practices to protect privacy. By contrast, the regulatory environment in the United States and Germany, which combined more ambiguous goals with tough transparency requirements and meaningful enforcement, were more successful in catalyzing companies to view

138 - Political Bots, accessed August 1, 2016.

http://politicalbots.org/

139 - Kenneth A. Bamberger and Deirdre K. Mulligan, Privacy on the Ground: Driving Corporate Behavior in the United States and Europe (Cambridge, Massachusetts: MIT Press, 2015). Like other technologies, AI has the potential to be used for good or nefarious purposes. A vigorous and informed debate about how to best steer AI in ways that enrich our lives and our society is an urgent and vital need.

50 - 51 - Newell and Simon pioneered the foray into heuristic search, an efficient procedure for finding solutions in large, combinatorial spaces. In particular, they applied this idea to construct proofs of mathematical theorems, first through their Logic Theorist program, and then through the General Problem Solver.

149 - In the area of computer vision, early work in character recognition by Selfridge and colleagues

150 - laid the basis for more complex applications such as face recognition.

151 - By the late sixties, work had also begun on natural language processing.

152 - “Shakey”, a wheeled robot built at SRI International, launched the field of mobile robotics. Samuel’s Checkers-playing program, which improved itself through self-play, was one of the first working instances of a machine learning system.

153 - Rosenblatt’s Perceptron,

154 - a computational model based on biological neurons, became the basis for the field of artificial neural networks. Feigenbaum and others advocated 155the case for building expert systems—knowledge repositories tailored for specialized domains such as chemistry and medical diagnosis.156 Early conceptual progress assumed the existence of a symbolic system that could be reasoned about and built upon. But by the 1980s, despite this promising headway made into different aspects of artificial intelligence, the field still could boast no significant practical successes. This gap between theory and practice arose in part from an insufficient emphasis within the AI community on grounding systems physically, with direct access to environmental signals and data. There was also an overemphasis on Boolean (True/False) logic, overlooking the need to quantify uncertainty. The field was forced to take cognizance of these shortcomings in the mid-1980s, since interest in AI began to drop, and funding dried up. Nilsson calls this period the “AI winter.” A much needed resurgence in the nineties built upon the idea that “Good OldFashioned AI”

157 - was inadequate as an end-to-end approach to building intelligent systems. Rather, intelligent systems needed to be built from the ground up, at all times solving the task at hand, albeit with different degrees of proficiency.

158 - Technological progress had also made the task of building systems driven by real-world data more feasible. Cheaper and more reliable hardware for sensing and actuation made robots easier to build. Further, the Internet’s capacity for gathering large amounts of data, and the availability of computing power and storage to process that data, enabled statistical techniques that, by design, derive solutions from data. These developments have allowed AI to emerge in the past two decades as a profound influence on our daily lives, as detailed in Section II.

149 - Allen Newell, John C Shaw, and Herbert A Simon, “Report on a general problem-solving program,” Proceedings of the International Conference on Information Processing, UNESCO, Paris 15-20 June 1959 (Unesco/Oldenbourg/Butterworths, 1960),

256–264. - 150 O. G. Selfridge, “Pandemonium: A paradigm for learning,” Proceedings of the Symposium on Mechanization of Thought Processes (London: H. M. Stationary Office, 1959):

511–531. 151 - Woodrow W. Bledsoe and Helen Chan, “A Man-Machine Facial Recognition System: Some Preliminary Results,” Technical Report PRI 19A (Palo Alto, California: Panoramic Research, Inc., 1965).

152 - D. Raj Reddy, “Speech Recognition by Machine: A Review,” Proceedings of the IEEE 64, no.4 (April 1976),

501–531. 153 - Arthur Samuel, “Some Studies in Machine Learning Using the Game of Checkers, IBM Journal of Research and Development 3, no. 3 (1959):

210—229. 154 - Frank Rosenblatt, “The Perceptron—A Perceiving and Recognizing Automaton,” Report 85-460-1, (Buffalo, New York: Cornell Aeronautical Laboratory, 1957).

155 - “Shakey the robot,” Wikipedia, last modified July 11, 2016, accessed August 1, 2016,

https://en.wikipedia.org/wiki/Shakey_the_robot

156 - Edward A. Feigenbaum and Bruce G. Buchanan, “DENDRAL and Meta-DENDRAL: Roots of Knowledge Systems and Expert System Applications,” Artificial Intelligence 59, no. 1-2 (1993),

233–240. 157 - John Haugeland, Artificial Intelligence: The Very Idea, (Cambridge, Massachusetts: MIT Press, 1985).

158 - Rodney A. Brooks, “Elephants Don’t Play Chess,” Robotics and Autonomous Systems 6, no. 1-2 (June 1990):

3–15. - APPENDIX I: A SHORT HISTORY OF AI This Appendix is based primarily on Nilsson’s book

140 - and written from the prevalent current perspective, which focuses on data intensive methods and big data. However important, this focus has not yet shown itself to be the solution to all problems. A complete and fully balanced history of the field is beyond the scope of this document. The field of Artificial Intelligence (AI) was officially born and christened at a 1956 workshop organized by John McCarthy at the Dartmouth Summer Research Project on Artificial Intelligence. The goal was to investigate ways in which machines could be made to simulate aspects of intelligence—the essential idea that has continued to drive the field forward. McCarthy is credited with the first use of the term “artificial intelligence” in the proposal he co-authored for the workshop with Marvin Minsky, Nathaniel Rochester, and Claude Shannon.

141 - Many of the people who attended soon led significant projects under the banner of AI, including Arthur Samuel, Oliver Selfridge, Ray Solomonoff, Allen Newell, and Herbert Simon. Although the Dartmouth workshop created a unified identity for the field and a dedicated research community, many of the technical ideas that have come to characterize AI existed much earlier. In the eighteenth century, Thomas Bayes provided a framework for reasoning about the probability of events.

142 - In the nineteenth century, George Boole showed that logical reasoning—dating back to Aristotle—could be performed systematically in the same manner as solving a system of equations.

143 - By the turn of the twentieth century, progress in the experimental sciences had led to the emergence of the field of statistics,

144 - which enables inferences to be drawn rigorously from data. The idea of physically engineering a machine to execute sequences of instructions, which had captured the imagination of pioneers such as Charles Babbage, had matured by the 1950s, and resulted in the construction of the first electronic computers.

145 - Primitive robots, which could sense and act autonomously, had also been built by that time.

146 - The most influential ideas underpinning computer science came from Alan Turing, who proposed a formal model of computing. Turing’s classic essay, Computing Machinery and Intelligence,

147 - imagines the possibility of computers created for simulating intelligence and explores many of the ingredients now associated with AI, including how intelligence might be tested, and how machines might automatically learn. Though these ideas inspired AI, Turing did not have access to the computing resources needed to translate his ideas into action. Several focal areas in the quest for AI emerged between the 1950s and the 1970s.148

140 - Nilsson, The Question for Artificial Intelligence.

141 - J. McCarthy, Marvin L. Minsky, Nathaniel Rochester, and Claude E. Shannon, “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence,” August 31, 1955, accessed August 1, 2016,

http://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html

142 - Thomas Bayes, “An Essay towards Solving a Problem in the Doctrine of Chances,” Philosophical Transactions of the Royal Society of London 53 (January 1, 1763): 370–418, accessed August 1, 2016,

143 - George Boole, An Investigation of the Laws of Thought on Which are Founded the Mathematical Theories of Logic and Probabilities, (Macmillan, 1854, reprinted with corrections, Dover Publications, New York, NY, 1958, and reissued by Cambridge University Press, 2009), accessed August 1, 2016,

144 - “History of statistics,” Wikipedia, Last modified June 3, 2016, accessed August 1, 2016,

https://en.wikipedia.org/wiki/History_of_statistics

145 - Joel N. Shurkin, Engines of the Mind: The Evolution of the Computer from Mainframes to Microprocessors (New York: W. W. Norton & Company, 1996).

146 - William Grey Walter, “An Electromechanical Animal,” Dialectica 4 (1950): 42–49.

147 - A. M. Turing, “Computing Machinery and Intelligence,” Mind 59, no. 236 (1950):

433–460. 148 - Marvin Minsky, “Steps toward Artificial Intelligence,” MIT Media Laboratory, October 24, 1960, accessed August 1, 2016,

http://web.media.mit.edu/~minsky/papers/steps.html

The field of Artificial Intelligence (AI) was officially born and christened at a 1956 workshop. The goal was to investigate ways in which machines could be made to simulate aspects of intelligence—the essential idea that has continued to drive the field forward. Although the separation of AI into sub-fields has enabled deep technical progress along several different fronts, synthesizing intelligence at any reasonable scale invariably requires many different ideas to be integrated.

52 - In summary, following is a list of some of the traditional sub-areas of AI. As described in Section II, some of them are currently “hotter” than others for various reasons. But that is neither to minimize the historical importance of the others, nor to say that they may not re-emerge as hot areas in the future. • Search and Planning deal with reasoning about goal-directed behavior. Search plays a key role, for example, in chess-playing programs such as Deep Blue, in deciding which move (behavior) will ultimately lead to a win (goal). • The area of Knowledge Representation and Reasoning involves processing information (typically when in large amounts) into a structured form that can be queried more reliably and efficiently. IBM’s Watson program, which beat human contenders to win the Jeopardy challenge in 2011, was largely based on an efficient scheme for organizing, indexing, and retrieving large amounts of information gathered from various sources.

159 • - Machine Learning is a paradigm that enables systems to automatically improve their performance at a task by observing relevant data. Indeed, machine learning has been the key contributor to the AI surge in the past few decades, ranging from search and product recommendation engines, to systems for speech recognition, fraud detection, image understanding, and countless other tasks that once relied on human skill and judgment. The automation of these tasks has enabled the scaling up of services such as e-commerce. • As more and more intelligent systems get built, a natural question to consider is how such systems will interact with each other. The field of Multi-Agent Systems considers this question, which is becoming increasingly important in on-line marketplaces and transportation systems. • From its early days, AI has taken up the design and construction of systems that are embodied in the real world. The area of Robotics investigates fundamental aspects of sensing and acting—and especially their integration—that enable a robot to behave effectively. Since robots and other computer systems share the living world with human beings, the specialized subject of Human Robot Interaction has also become prominent in recent decades. • Machine perception has always played a central role in AI, partly in developing robotics, but also as a completely independent area of study. The most commonly studied perception modalities are Computer Vision and Natural Language Processing, each of which is attended to by large and vibrant communities. • Several other focus areas within AI today are consequences of the growth of the Internet. Social Network Analysis investigates the effect of neighborhood relations in influencing the behavior of individuals and communities.

Crowdsourcing is yet another innovative problem-solving technique, which relies on harnessing human intelligence (typically from thousands of humans) to solve hard computational problems. Although the separation of AI into sub-fields has enabled deep technical progress along several different fronts, synthesizing intelligence at any reasonable scale invariably requires many different ideas to be integrated. For example, the AlphaGo program

160 161 - that recently defeated the current human champion at the game of Go used multiple machine learning algorithms for training itself, and also used a sophisticated search procedure while playing the game.

159 - David A. Ferrucci, “Introduction to ‘This is Watson,’” IBM Journal of Research and Development, 56, no. 3-4 (2012): 1.

160 - David Silver et al., “Mastering the Game of Go with Deep Neural Networks and Tree Search.”

161 - Steven Borowiec and Tracey Lien, “AlphaGo beats human Go champ in milestone for artificial intelligence,” Los Angeles Times, March 12, 2016, accessed August 1, 2016.

asimo_run71cdjcpqqpl506px-Asimoaibo-sony

 @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@

L'INTELLIGENCE ARTIFICIELLE ET LA VIE EN 2030 UNE ÉTUDE DE L'ANNÉE SUR L'INTELLIGENCE ARTIFICIELLE | RAPPORT DU GROUPE D'ÉTUDES DE 2015 | PREMIERE SEPTEMBRE 2016

L'étude de cent ans sur l'intelligence artificielle, lancée à l'automne 2014, est une enquête à long terme sur le domaine de l'Intelligence Artificielle (IA) et ses influences sur les personnes, leurs communautés et la société. Il considère la science, l'ingénierie et le déploiement des systèmes informatiques basés sur l'IA. En tant qu'activité de base, le Comité permanent qui supervise l'étude de cent ans forme un comité d'étude tous les cinq ans pour évaluer l'état actuel de l'IA. Le comité d'étude passe en revue les progrès réalisés par IA dans les années qui suivent le rapport immédiat, envisage les progrès potentiels à venir et décrit les défis et les opportunités techniques et sociétales que ces avancées augmentent, y compris dans des domaines tels que l'éthique, l'économie et la conception des systèmes compatible avec la cognition humaine.

Le but général de l'examen périodique de l'étude de cent ans est de fournir un ensemble de réflexions collectées et connectées sur l'IA et ses influences au fur et à mesure que le champ avance. On s'attend à ce que les études développent des synthèses et des évaluations qui fournissent des orientations éclairées par des experts pour les orientations de la recherche, du développement et de la conception des systèmes d'IA, ainsi que des programmes et des politiques visant à garantir que ces systèmes bénéficient largement aux individus et à la Société.

1 - L'étude de cent ans est basée sur un effort antérieur informellement connu sous le nom de «Étude AAAI Asilomar». En 2008-2009, le président de l'Association pour l'avancement de l'intelligence artificielle (AAAI), Eric Horvitz, a réuni un groupe d'experts de l'IA de multiples institutions et domaines du domaine, ainsi que des spécialistes des sciences cognitives, de la philosophie et du droit. En travaillant dans des sous-groupes distribués, les participants ont abordé les développements à long terme de l'IA, les possibilités à long terme et les préoccupations juridiques et éthiques, puis se sont réunis lors d'une réunion de trois jours à Asilomar pour partager et discuter de leurs résultats. Un bref rapport écrit sur les discussions approfondies des réunions, amplifié par les discussions ultérieures des participants avec d'autres collègues, a généré un large intérêt et un débat sur le terrain et au-delà.

L'impact de la réunion d'Asilomar et les progrès importants dans l'IA qui comprenaient les algorithmes et les technologies de l'IA qui commencent à entrer dans la vie quotidienne dans le monde entier ont stimulé l'idée d'une étude récurrente à long terme sur l'IA et son influence sur les personnes et la société. L'étude de cent ans a ensuite été dotée d'une université pour permettre une «étude de cent ans sur l'intelligence artificielle (AI100)», Université de Stanford, accessible le 1er août 2016.

https://ai100.stanford.edu

Le but général de l'examen périodique de l'étude de cent ans est de fournir un ensemble de réflexions collectées et connectées sur l'IA et ses influences au fur et à mesure que le champ avance.

2-3 - Comme une conséquence de la décision de se concentrer sur la vie dans les villes nord-américaines, les applications militaires ont été considérées comme hors du champ d'application de ce rapport initial. Il ne s'agit pas de minimiser l'importance d'une surveillance minutieuse et d'une réflexion sur les implications des progrès de l'IA pour la défense et la guerre, y compris les développements et les déploiements potentiellement déstabilisants. Le rapport est conçu pour répondre à quatre publics visés. Pour le grand public, il vise à fournir une représentation accessible, scientifiquement et technologiquement précise de l'état actuel de l'IA et de son potentiel. Pour l'industrie, le rapport décrit les technologies pertinentes et les défis juridiques et éthiques, et peut aider à guider l'allocation des ressources. Le rapport s'adresse également aux gouvernements locaux, nationaux et internationaux pour les aider à mieux planifier l'IA dans la gouvernance.

Enfin, le rapport peut aider les chercheurs de l'IA, ainsi que leurs institutions et leurs bailleurs de fonds, à définir des priorités et à examiner les problèmes éthiques et juridiques soulevés par la recherche sur l'IA et ses applications. Compte tenu de la nature unique de l'étude de cent ans sur l'IA, nous nous attendons à ce que les générations futures de comités permanents et de comités d'études, ainsi que les chercheurs, les experts en politiques, les dirigeants du secteur privé et public et le grand public réfléchiront à cette évaluation car elles font de nouvelles évaluations de l'avenir d'IA. Nous espérons que ce premier effort de la série qui s'étend avant nous sera utile à la fois pour ses échecs et ses succès dans la prédiction précise de la trajectoire et des influences de l'IA.

Le Comité permanent remercie les membres du comité d'étude d'avoir investi leur expertise, leurs points de vue et leur temps considérable pour la création de ce rapport inaugural. Nous remercions particulièrement le professeur Peter Stone d'avoir accepté de servir de Président de l'étude et de son leadership judicieux, habile et dévoué du panel, de ses discussions et de la création du rapport ainsi que le Comité permanent de l'étude de cent ans sur l'intelligence artificielle Barbara J. Grosz, président Russ Altman Eric Horvitz Alan Mackworth Tom Mitchell Deidre Mulligan Yoav Shoham qui a approfondi les réflexions approfondies et les recherches scientifiques interdisciplinaires qui pourraient inspirer l'innovation et fournir des conseils intelligents aux agences gouvernementales et industrie.

Ce rapport est le premier de la série d'études prévue qui se poursuivra pendant au moins cent ans. Le Comité permanent a défini une charge du Panel d'étude pour le Comité d'étude inaugural à l'été 2015 et a recruté le professeur Peter Stone, à l'Université du Texas à Austin, pour présider le panel. Le comité d'étude de dix-sept membres, composé d'experts en IA du milieu universitaire, des laboratoires d'entreprise et de l'industrie, et des savants expérimentés en droit, en science politique, en politique et en économie, a été lancé au milieu de l'automne 2015. Les participants représentent diverses spécialités des régions géographiques, les genres et les étapes de carrière. Le Comité permanent a examiné de façon approfondie les moyens d'encadrer la charge du Panel d'étude pour tenir compte à la fois des progrès récents en matière d'IA et des impacts sociétaux potentiels sur les emplois, l'environnement, les transports, la sécurité publique, les soins de santé, l'engagement communautaire et le gouvernement.

Le Comité a examiné diverses façons de concentrer l'étude, y compris les sous-domaines de l'enquête et leur statut, en examinant une technologie particulière telle que l'apprentissage par machine ou le traitement du langage naturel, et l'étude de domaines d'application particuliers tels que les soins de santé ou le transport. Le comité a finalement choisi une approche thématique sur «L'IA et la vie en 2030 » afin de reconnaître que les diverses utilisations et impacts de l'IA ne se produiront pas indépendamment les uns des autres ou d'une multitude d'autres développements sociétaux et technologiques. Reconnaissant le rôle central que les villes ont joué dans la plus grande partie de l'expérience humaine, l'accent a été concentré sur les grandes zones urbaines où vivent la plupart des personnes.

Le Comité permanent a également atténué l'attention sur une ville typiquement nord-américaine en reconnaissance de la grande variabilité des milieux et des cultures urbaines dans le monde et limite les efforts du premier groupe d'étude. Le Comité permanent s'attend à ce que les projections, les évaluations et les conseils proactifs découlant de l'étude auront une pertinence mondiale plus large et prévoient des études futures afin d'élargir la portée du projet à l'échelle internationale.

Panneau d'étude : Peter Stone, Université du Texas à Austin, président Rodney Brooks, Rethink Robotics Erik Brynjolfsson, Institut de technologie de Massachussets Ryan Calo, Université de Washington Oren Etzioni, Institut Allen pour AI Greg Hager, Université Johns Hopkins Julia Hirschberg, Université de Columbia Shivaram Kalyanakrishnan , Institut indien de technologie Bombay Ece Kamar, Microsoft Research Sarit Kraus, Université Bar Ilan Kevin Leyton-Brown, Université de la Colombie-Britannique David Parkes, Université de Harvard William Press, Université du Texas à Austin Anna Lee (Anno) Saxenian, Université de Californie, Berkeley Julie Shah, Institut de technologie de Massachussets Milind Tambe, Astro Teller de l'Université de Californie du Sud, X Remerciements: Les membres du Panel d'étude reconnaissent avec gratitude le soutien et la contribution précieuse du Comité permanent, en particulier la présidente, Barbara Grosz, qui a traité avec suprême remerciement le rôle peu enviable de la médiation entre deux grands, très passionnés C.Ommittes.

Nous remercions également Kerry Tremain pour sa contribution infatigable et perspicace sur le produit écrit pendant le vaste processus d'édition et de polissage, ce qui a incontestablement renforcé le rapport considérablement.

TABLE DES MATIÈRES PRÉFACE RÉSUMÉ ANALYTIQUE 4 APERÇU 6 SECTION I: QUEL EST L'INTELLIGENCE ARTIFICIELLE 

12 - Définir IA -  12 Tendances de recherche IA - ​​14 SECTION II: AI PAR DOMAINE - 18 Transport - 18 Robots domestiques / de services - 24 Soins de santé - 25 Éducation - 31 Collectivités à faibles ressources - 35 Sécurité publique - 36 Emploi et lieu de travail - 38 Divertissement - 40 SECTION III: PERSPECTIVES ET RECOMMANDATIONS POUR L'IA POLITIQUE PUBLIQUE - 42 AI Policy, Now and in the Future - 42 APPENDICE I: UNE HISTOIRE COURTE DE L'IA

- 50 4 5 - Le travail est augmenté ou remplacé par l'IA, créant de nouveaux défis pour l'économie et la société de manière plus générale. La conception de l'application et les décisions stratégiques prises à court terme sont susceptibles d'avoir des influences durables sur la nature et les orientations de ces développements, ce qui rend important pour les chercheurs, les développeurs, les spécialistes des sciences sociales et les décideurs de l'IA de mettre en équilibre l'impératif d'innover avec les mécanismes. Il faut s'assurer que les avantages économiques et sociaux d'IA sont largement partagés dans toute la société. Si la société s'approche de ces technologies principalement avec crainte et suspicion, des erreurs ralentiront le développement de l'IA ou la pousseront sous-terre en entraînant un travail important pour assurer la sécurité et la fiabilité des technologies de l'IA.

D'autre part, si la société s'approche de l'IA avec un esprit plus ouvert, les technologies émergeant du domaine pourraient transformer profondément la société en mieux dans les décennies à venir.

Panneau d'étude: Peter Stone, président, Université du Texas à Austin, Rodney Brooks, Rethink Robotics, Erik Brynjolfsson, Massachussets Institute of Technology, Ryan Calo, Université de Washington, Oren Etzioni, Allen Institute for AI, Greg Hager, Université Johns Hopkins, Julia Hirschberg, Université de Columbia, Shivaram Kalyanakrishnan, Indian Institute of Technology Bombay, Ece Kamar, Microsoft Research, Sarit Kraus, Université Bar Ilan. Kevin Leyton-Brown, Université de la Colombie-Britannique, David Parkes, Université de Harvard, William Press, Université du Texas à Austin, AnnaLee (Anno) Saxenian, Université de Californie, Berkeley, Julie Shah, Massachussets Institute of Technology, Milind Tambe, Université de Southern California, Astro Teller, X Comité permanent de l'étude de cent ans de l'intelligence artificielle: Barbara J. Grosz, président, Russ Altman, Eric Horvitz, Alan Mackworth, Tom Mitchell, Deidre Mulligan, Yoav Shoham.

RÉSUMÉ ANALYSE ARTIFICIELLE (IA) est une science et un ensemble de technologies informatiques qui s'inspirent, mais généralement fonctionnent de manière tout à fait différente dans les façons dont les gens utilisent leurs systèmes et leurs corps nerveux pour détecter, apprendre, raisonner et agir.

Bien que le rythme de progression de l'IA ait été inégalé et imprévisible, il y a eu des avancées considérables depuis la création du champ il y a soixante ans. Une fois que le domaine d'étude est essentiellement académique, l'IA du XXIe siècle permet une constellation de technologies traditionnelles qui ont un impact important sur la vie quotidienne. La vision par ordinateur et la planification de l'IA, par exemple, conduisent les jeux vidéo qui sont maintenant une plus grande industrie du divertissement que Hollywood. L'apprentissage approfondi, une forme d'apprentissage machine basée sur des représentations en couches de variables appelées réseaux neuronaux, a rendu la compréhension de la parole pratique sur nos téléphones et dans nos cuisines, et ses algorithmes peuvent être largement appliqués à un éventail d'applications reposant sur un modèle de reconnaissance vocale.

Le traitement du langage naturel (NLP) et la représentation et le raisonnement du savoir ont permis à une machine de battre le champion Jeopardy et apporte un nouveau pouvoir aux recherches sur le Web. Bien que impressionnantes, ces technologies sont très adaptées aux tâches particulières. Chaque application nécessite généralement des années de recherche spécialisée et une construction unique . Dans les applications ciblées de manière similaire, on peut s'attendre à des augmentations substantielles des utilisations futures des technologies de l'IA, y compris dans d'autres domaines comme les voitures autonomes, les diagnostics de soins de santé et les traitements ciblés, et une assistance physique pour les soins aux personnes âgées.

L'IA et la Robotique seront également appliquées dans le monde entier dans les industries qui s'efforcent d'attirer des travailleurs plus jeunes, tels que l'agriculture, la transformation des aliments, les centres d'approvisionnement et les usines. Ils faciliteront la livraison d'achats en ligne à travers des drones volants( comme chez AMAZONE), des camions auto-conduisant ou des robots qui peuvent monter les escaliers jusqu'à la porte d'entrée. Ce rapport est le premier d'une série à émettre à intervalles réguliers dans le cadre de l'étude de cent ans sur l'intelligence artificielle (IA 100). À partir d'une charge donnée par le Comité permanent IA 100 pour tenir compte des influences probables de l'IA dans une ville typiquement nord-américaine d'ici 2030, le Panel d'étude 2015, composé d'experts en IA et d'autres domaines pertinents, a concentré son attention sur huit domaines qu'ils considéraient le plus important : transport; Robots de service; soins de santé; éducation; Communautés à faible ressources; Sécurité publique; Emploi et lieu de travail et le divertissement. Dans chacun de ces domaines, le rapport reflète à la fois les progrès réalisés au cours des quinze dernières années et prévoit des développements au cours des quinze prochaines années.

Bien que tiré d'une source de recherche commune, chaque domaine reflète différentes influences et défis de l'IA, tels que la difficulté de créer un matériel sûr et fiable (robots de transport et de service), la difficulté d'interagir en douceur avec les experts humains (santé et éducation), le défi de gagner la confiance du public (communautés à faible ressources et sécurité publique), le défi de surmonter les peurs de marginaliser les humains (emploi et lieu de travail) et le risque social et sociétal de diminuer les interactions interpersonnelles (divertissement). Le rapport commence par une réflexion sur ce qui constitue l'Intelligence Artificielle, et se termine par des recommandations concernant les politiques liées à l'IA.

Ces recommandations incluent une expertise technique acquise sur l'IA dans le gouvernement et consacrent plus de ressources - et éliminant les obstacles - à la recherche sur l'équité, la sécurité, la vie privée et les impacts sociétaux des systèmes d'IA. Contrairement aux prédictions plus fantastiques de l'IA dans la presse populaire, le comité d'étude n'a pas été préoccupant que l'IA soit une menace imminente pour l'humanité. Aucune machine ayant des objectifs et des intentions durables à long terme n'a été développée et ne devrait pas non plus être développée dans un proche avenir. Au lieu de cela, des applications de plus en plus utiles de l'IA, avec des impacts positifs potentiellement profonds sur notre société et notre économie, vont probablement apparaître d'ici 2030, la période considérée par ce rapport.

Dans le même temps, bon nombre de ces développements stimuleront les perturbations dans la façon dont les augmentations substantielles dans les utilisations futures des applications d'IA, y compris plus de voitures autonomes, des diagnostics de soins de santé et des traitements ciblés et une assistance physique pour les soins aux aînés. Tout en s'appuyant sur des recherches et des technologies communes, les systèmes d'IA sont spécialisés pour accomplir des tâches particulières. Chaque application nécessite des années de recherche ciblée et une construction soigneuse et unique.

6 7 - Une image précise et sophistiquée d'AI-one qui est en concurrence avec sa représentation populaire - est entravée au début par la difficulté de fixer une définition précise de l'intelligence artificielle. Dans les approches examinées par le Comité d'étude, aucun ne suggère qu'il existe actuellement une IA «à usage général».

Tout en s'appuyant sur des recherches et des technologies communes, les systèmes d'IA sont spécialisés pour accomplir des tâches particulières carchaque application nécessite des années de recherche ciblée et une construction minutieuse et unique. Par conséquent, les progrès sont inégaux parmi ces huit domaines. Un exemple exemplaire est le transport, où quelques technologies clés ont catalysé l'adoption généralisée de l'IA avec une vitesse étonnante. Le transport autonome sera bientôt un lieu commun, comme la première expérience de la plupart des personnes avec les systèmes d'IA physiquement incorporés, influencera fortement la perception de l'IA par le public.

Comme les voitures deviennent de meilleurs conducteurs que les gens, les citadins posséderont moins de voitures, vivront plus loin du travail et passeront le temps différemment, ce qui entraînera une organisation urbaine entièrement nouvelle. Dans la ville typiquement nord-américaine en 2030, les applications de l'IA physiquement incorporées ne seront pas limitées aux voitures, mais sont susceptibles d'inclure des camions, des véhicules volants et des robots personnels. Les améliorations apportées à un matériel sûr et fiable stimuleront l'innovation au cours des quinze prochaines années, comme cela sera le cas chez Home / Service Robots, qui sont déjà entrés dans les maisons des gens, principalement sous forme d'aspirateurs.

Les meilleures puces, les capteurs 3D à faible coût, l'apprentissage par machine en nuage et les progrès dans la compréhension de la parole amélioreront les services futurs des robots et leurs interactions avec les gens. Les robots spéciaux proposent des paquets, nettoient des bureaux et améliorent la sécurité. Mais les contraintes techniques et les coûts élevés des dispositifs mécaniques fiables continueront à limiter les opportunités commerciales aux applications étroitement définies dans un avenir prévisible. Dans les soins de santé, il y a eu un énorme bond en avant dans la collecte de données utiles à partir de dispositifs de surveillance personnelle et d'applications mobiles, à partir de dossiers de santé électroniques (DSE) en milieu clinique et, dans une moindre mesure, de robots chirurgicaux conçus pour aider aux procédures et aux services médicaux avec des robots qui soutiennent les opérations hospitalières. Les applications basées sur l'IA pourraient améliorer les résultats sanitaires et la qualité de vie pour des millions de personnes dans les années à venir.

Bien que les applications cliniques aient été lentes à passer du laboratoire d'informatique au monde réel, des signes d'espoir montrent que le rythme de l'innovation s'améliorera. Les progrès dans les soins de santé peuvent être encouragés par le développement d'incitations et de mécanismes de partage de données et pour éliminer les obstacles à la réglementation et aux obstacles commerciaux. Pour de nombreuses applications, les systèmes IA devront travailler en étroite collaboration avec les fournisseurs de soins et les patients pour gagner leur confiance. Les progrès dans la façon dont les machines intelligentes interagissent naturellement avec les soignants, les patients et les familles des patients sont cruciaux. Permettre des interactions plus fluides entre les personnes et les technologies prometteuses de l'IA est également un défi majeur dans l'éducation, qui a considérablement progressé dans la même période.

Bien que l'éducation de qualité exige toujours un engagement actif des enseignants humains, IA promet d'améliorer l'éducation à tous les niveaux, en particulier en fournissant une personnalisation à l'échelle humaine. Les enseignants interactifs de machines sont maintenant adaptés aux étudiants pour l'enseignement des sciences, des mathématiques, de la langue et d'autres disciplines. Le traitement du langage naturel, l'apprentissage par machine et le "crowdsourcing" ont stimulé l'apprentissage en ligne et ont permis aux enseignants de l'enseignement supérieur de multiplier la taille de leurs salles de classe tout en répondant aux besoins et aux styles d'apprentissage des élèves. Au cours des quinze prochaines années, dans une ville typiquement nord-américaine, l'utilisation de ces technologies dans la salle de classe et dans la maison risque de se développer de manière significative, à condition qu'elles puissent être intégrées de manière significative à l'apprentissage en face à face.

Au-delà de l'éducation, de nombreuses possibilités existent pour les méthodes d'IAI pour aider les communautés à faible ressources en fournissant des mesures d'atténuation et des solutions à divers problèmes sociaux. Traditionnellement, les bailleurs de fonds ont sous-investi dans une recherche sur l'IA sans demande commerciale. Avec des incitations ciblées et des priorités de financement.

APERÇU : Les représentations effrayantes et futuristes de l'Intelligence Artificielle qui dominent les films et les romans et façonnent l'imagination populaire sont fictives. En réalité, l'IA change déjà notre vie quotidienne, presque entièrement dans le but d'améliorer la santé humaine, la sécurité et la productivité.

Contrairement aux films, il n'y aura pas de race de robots surhumains à l'horizon ou probablement même possible bien que le potentiel d'abus des technologies de l'IA soit reconnu et abordé, leur plus grand potentiel est, entre autres choses, de rendre la conduite plus sûre, d'aider les enfants à apprendre et d'étendre et d'améliorer la vie des gens.

En fait, les applications de l'IA de manière positive dans les écoles, les maisons et les hôpitaux connaissent déjà une croissance accélérée. Les grandes universités de recherche consacrent les départements aux études d'IAI, et des entreprises technologiques telles que Apple, Facebook, Google, IBM et Microsoft dépensent beaucoup pour explorer les applications d'IA qu'elles considèrent comme critiques pour leur avenir. Même Hollywood utilise les technologies de l'IA pour amener ses fantasmes dystopiques de l'IA à l'écran. Les innovations reposant sur la vision par ordinateur, la reconnaissance de la parole et le traitement du langage naturel ont influencé ces changements, tout comme les progrès scientifiques et technologiques concurrents dans les domaines connexes. L'IA modifie également la façon dont les gens interagissent avec la technologie. Beaucoup de gens se sont déjà habitués à toucher et à parler à leurs téléphones intelligents. Les relations futures avec les machines deviennent de plus en plus nuancées, fluides et personnalisées à mesure que les systèmes d'IA apprennent à s'adapter aux personnalités et aux objectifs individuels. Ces applications d'IA aideront à surveiller le bien-être des gens, à les alerter sur les risques à venir et à fournir des services si nécessaire ou souhaité. Par exemple, en seulement une quinzaine d'années dans une ville typiquement nord-américaine - le calendrier et la portée de ce rapport -, les applications d'IA sont susceptibles de transformer le transport vers les véhicules autonomes avec le ramassage et la livraison des personnes et des forfaits sur place. Cela permettra de reconfigurer le paysage urbain, car les embouteillages et les défis de stationnement deviennent obsolètes. La focalisation de cette étude sur une ville nord-américaine typique est délibérée et vise à mettre en évidence des changements spécifiques affectant la vie quotidienne des millions de personnes qui les habitent.

Le panel d'étude a encore rétréci son enquête dans huit domaines où l'IA a déjà ou a pour objectif d'avoir le plus grand impact: le transport, les soins de santé, l'éducation, les communautés à faible ressources, la sécurité et la sécurité publique, les emplois et les lieux de travail, les robots de la maison et du service et le divertissement. Bien que tirés d'une source commune de recherche, les technologies de l'IA ont influencé et continueront à influencer ces domaines de manière différente. Chaque domaine est confronté à des défis variés liés à l'IA, y compris la difficulté de créer un matériel sûr et fiable pour détecter et effectuer (robots de transport et de service), la difficulté d'interagir en douceur avec les experts humains (santé et éducation), le défi de gagner la confiance du public ( les communautés à faible ressources et la sécurité et la sécurité publique), le défi de surmonter les craintes de marginalisation des humains (emploi et lieu de travail) et le risque de diminuer l'interaction interpersonnelle (divertissement).

Certains domaines sont principalement des secteurs d'activité, tels que les transports et les soins de santé, tandis que d'autres sont plus orientés vers les consommateurs, tels que les robots de divertissement et de service à la maison. Certaines entreprises touchent des secteurs tels que l'emploi / le milieu de travail et les communautés à faible ressources. Dans chaque domaine, même si l'IA continue de fournir des avantages importants, elle soulève également des problèmes éthiques et sociaux importants, y compris des problèmes de confidentialité. Les robots et d'autres technologies de l'IA ont déjà commencé à déplacer des emplois dans certains secteurs. En tant que société, nous sommes maintenant à un moment crucial pour déterminer comment déployer des technologies basées sur l'IA de manière à promouvoir, sans entraver, des valeurs démocratiques comme la liberté, l'égalité et la transparence. Pour les individus, la qualité de la vie que nous menons et la façon dont nos contributions sont évaluées sont susceptibles de changer progressivement, mais de manière marquée. Au cours des prochaines années, la recherche sur l'IA, le développement de systèmes et les cadres sociaux et réglementaires détermineront comment les avantages de l'IA sont pris en compte par rapport à ses coûts et à leurs risques, et à quel point ces avantages se propagent largement. Beaucoup ont déjà eu l'habitude de toucher et de parler à leurs téléphones intelligents. ( Smartphones portables).

Les relations futures entre les gens avec les machines deviendront de plus en plus nuancées, fluides et personnalisées. La société est maintenant à un moment crucial pour déterminer comment déployer des technologies basées sur l'IA de manière à promouvoir plutôt que d'entraver les valeurs démocratiques telles que la liberté, l'égalité et la transparence.

-8 9 - Les marchés pour les produits axés sur les données et les incitations économiques pour trouver de nouveaux produits et marchés, ont également stimulé les progrès de la recherche. Maintenant, en devenant une force centrale dans la société, le domaine de l'IA évolue vers la construction de systèmes intelligents qui peuvent collaborer efficacement avec les gens et qui sont plus généralement conscients de l'humanité, y compris des façons créatives de développer des moyens interactifs et évolutifs pour les Robots enseignants. Ces tendances conduisent les domaines actuellement «chauds» de la recherche sur les méthodes fondamentales et les domaines d'application: l'apprentissage par machine à grande échelle concerne la conception d'algorithmes d'apprentissage, ainsi que la mise à l'échelle des algorithmes existants, afin de fonctionner avec des ensembles de données extrêmement importants. L'apprentissage approfondi, une classe de procédures d'apprentissage, a facilité la reconnaissance des objets dans les images, l'étiquetage vidéo et la reconnaissance d'activité, et entraîne des percées significatives dans d'autres domaines de la perception, tels que l'audio, la parole et le traitement du langage naturel.

L'apprentissage par renforcement est un cadre qui met l'accent sur l'apprentissage par machine de la reconnaissance de formes à la prise de décision séquentielle axée sur l'expérience. Il promet de transmettre les demandes d'IA vers des actions dans le monde réel. Tout en étant largement confiné au milieu universitaire au cours des dernières décennies, il voit maintenant des succès pratiques et réels. Robotics s'intéresse actuellement à la formation d'un robot pour interagir avec le monde autour de lui de manière généralisée et prévisible, comment faciliter la manipulation d'objets dans des environnements interactifs et comment interagir avec les gens. Les progrès de la robotique s'appuieront sur des progrès proportionnels pour améliorer la fiabilité et la généralité de la vision par ordinateur et d'autres formes de perception de la machine. La vision par ordinateur est actuellement la forme la plus importante de la perception de la machine.

Il a été la sous-zone de l'IA la plus transformée par la montée de l'apprentissage approfondi. Pour la première fois, les ordinateurs sont en mesure d'effectuer certaines tâches de vision mieux que les gens. La recherche en cours est axée sur l'image automatique et le sous-titrage vidéo. Le traitement du langage naturel, souvent associé à la reconnaissance automatique de la parole, devient rapidement une marchandise pour les langues largement parlées avec de grands ensembles de données. La recherche se déplace maintenant pour développer des systèmes raffinés et capables qui peuvent interagir avec les personnes à travers le dialogue, et non seulement réagir aux demandes stylisées. De grands progrès ont également été réalisés dans la traduction automatique entre différentes langues, avec plus d'échanges personne-à-personne en temps réel sur l'horizon proche. La recherche de systèmes collaboratifs étudie des modèles et des algorithmes pour aider à développer des systèmes autonomes qui peuvent fonctionner en collaboration avec d'autres systèmes et avec des humains. La recherche sur le "crowdsourcing" et le calcul humain étudie les méthodes pour augmenter les systèmes informatiques en effectuant des appels automatiques à l'expertise humaine pour résoudre les problèmes que les ordinateurs ne peuvent pas résoudre correctement.

La théorie du jeu algorithmique et le choix social des calculs attirent l'attention sur les dimensions informatiques économiques et sociales de l'IA, comme la façon dont les systèmes peuvent gérer les incitations potentiellement désalignées, y compris les participants ou les entreprises humaines intéressés et les agents automatisés basés sur l'IA qui les représentent. La recherche Internet of Things (IoT) ( l'internet des objets ) est consacrée à l'idée qu'un large éventail de périphériques, y compris les appareils, les véhicules, les bâtiments et les caméras, peuvent être interconnectés pour collecter et partager leur abondante information sensorielle à des fins intelligentes.

L'informatique neuromorphe est un ensemble de technologies qui cherchent à imiter les réseaux de neurones biologiques afin d'améliorer l'efficacité matérielle et la robustesse des systèmes informatiques, en remplaçant souvent un accent plus ancien sur les modules distincts pour l'entrée / sortie, le traitement des instructions et la mémoire. Les technologies de l'IA pourraient aider à répondre aux besoins des communautés à faible ressources et les efforts en mutation sont prometteurs.

En utilisant l'exploration de données et l'apprentissage par machine, par exemple, l'IA a été utilisée pour créer des modèles prédictifs pour aider les organismes gouvernementaux à résoudre des problèmes tels que la prévention de l'empoisonnement au plomb chez les enfants à risque et la distribution de nourriture efficacement. Ces efforts naissants suggèrent qu'on pourrait faire plus, en particulier si les organismes et les organisations peuvent s'engager et créer de la confiance avec ces communautés. Gagner de la confiance publique est également un défi pour l'utilisation de l'IA par les professionnels de la sécurité publique et de la sécurité. Les villes nord-américaines et les agences fédérales ont déjà commencé à déployer des technologies de l'IA dans l'administration des frontières et l'application de la loi. D'ici 2030, ils s'appuieront fortement sur eux, y compris des caméras améliorées et des drones pour la surveillance, des algorithmes pour détecter la fraude financière et la police prédictive. Ce dernier soulève le spectre des personnes innocentes surveillées de façon injustifiable, et il faut veiller à éviter de systématiser les biais humains et de protéger les libertés civiles.

Les outils de prédiction de l'IA bien déployés ont le potentiel de fournir de nouveaux types de transparence sur les données et les inférences, et peuvent être appliqués pour détecter, supprimer ou réduire les biais humains, plutôt que de le renforcer. Les décisions sociales et politiques sont également en jeu dans les influences d'IA sur les tendances en matière d'emploi et de travail, telles que les filets de sécurité nécessaires pour protéger les gens contre les changements structurels dans l'économie. L'IA est prêt à remplacer les personnes dans certains types d'emplois, comme dans la conduite des taxis et des camions. Cependant, dans de nombreux domaines, l'IA remplacera probablement les tâches plutôt que les emplois à court terme et créera également de nouveaux types d'emplois. Mais les nouveaux emplois qui apparaîtront sont plus difficiles à imaginer à l'avance que les emplois existants qui seront probablement perdus. L'IA réduira également le coût de nombreux biens et services, ce qui rendra plus d'efficacité tout le monde. À plus long terme, l'IA peut être considérée comme un mécanisme radicalement différent pour la création de richesse dans lequel tout le monde devrait avoir droit à une partie des trésors produits par l'IA dans le monde entier. Ce n'est pas trop tôt pour le débat social sur la manière dont les fruits économiques des technologies de l'IA devraient être partagés.

Entertainment a été transformé par des réseaux sociaux et d'autres plateformes pour le partage et la navigation sur les blogs, les vidéos et les photos, qui s'appuient sur des techniques développées activement en PNL, la récupération d'informations, le traitement d'image, le "crowdsourcing" et l'apprentissage par machine.

Certaines sources traditionnelles de divertissement ont également embrassé IA pour composer de la musique, créer des performances de scène et même pour générer des scènes 3D à partir de textes en langage naturel. L'enthousiasme avec lequel les gens ont déjà répondu au divertissement piloté par l'IA a été surprenant. Comme pour de nombreux aspects de l'IA, il y a un débat en cours sur la mesure dans laquelle la technologie remplace ou améliore la sociabilité. L'IA facilitera de plus en plus le divertissement plus interactif, personnalisé et plus engageant. La recherche devrait viser à comprendre comment tirer parti de ces attributs pour les bénéfices des individus et de la Société.

Quelle est la prochaine étape pour la recherche sur l'IA?

La recherche qui alimente la révolution de l'IA a également connu des changements rapides. Le plus important parmi eux est la maturation de l'apprentissage par machine, stimulée en partie par la montée de l'économie numérique, qui fournit et exploite de nombreuses quantités de données.

D'autres facteurs comprennent l'augmentation des ressources en "cloud computing" et la demande des consommateurs pour un accès généralisé aux services tels que la reconnaissance de la parole et le support de la navigation. L'apprentissage par machine a été propulsé de manière dramatique par des succès empiriques impressionnants de réseaux de neurones artificiels, qui peuvent maintenant être formés avec des ensembles de données énormes et des calculs à grande échelle. Cette approche est appelée «apprentissage approfondi». Le saut dans la performance des algorithmes de traitement de l'information s'est accompagné de progrès significatifs dans la technologie du matériel pour les opérations de base telles que la détection, la perception et la reconnaissance d'objets. De nouvelles plateformes et à plus long terme, l'IA peut être considérée comme un mécanisme radicalement différent pour la création de richesse dans lequel tout le monde devrait avoir droit à une partie des trésors produits par l'IA dans le monde entier.

Le domaine de l'IA évolue vers la construction de systèmes intelligents qui peuvent collaborer efficacement avec les gens, y compris des façons créatives de développer des moyens interactifs et évolutifs pour les personnes qui enseignent les robots.

10 11 - les capacités et l'efficacité des personnes qui y ont accès. Les politiques devraient être évaluées quant à savoir s'ils favorisent les valeurs démocratiques et le partage équitable des bénéfices de l'IA, ou concentrent le pouvoir et les avantages dans les mains d'un nombre heureux. Au fur et à mesure que ce rapport documente, des progrès significatifs liés à l'IA ont déjà eu un impact sur les villes nord-américaines au cours des quinze dernières années, et des développements encore plus importants se produiront au cours des quinze prochaines années. Les progrès récents sont en grande partie attribuables à la croissance et à l'analyse de grands ensembles de données activés par Internet, des progrès dans les technologies sensorielles et, plus récemment, des applications d'apprentissage approfondi.

Au cours des prochaines années, au fur et à mesure que le public rencontrera de nouvelles applications de l'IA dans des domaines tels que les transports et les soins de santé, ils doivent être introduits de manière à renforcer la confiance et à la compréhension et à respecter les droits humains et civils. Tout en encourageant l'innovation, les politiques et les processus devraient aborder les implications éthiques, de la vie privée et de la sécurité, et devraient veiller à ce que les avantages des technologies de l'IA soient répartis de façon large et équitable. Cela sera essentiel si la recherche sur l'Intelligence Artificielle et ses applications exercent une influence positive sur la vie urbaine nord-américaine en 2030 et au-delà avec la politique d'IA, maintenant et à l'avenir La mesure du succès pour les applications d'IA est la valeur qu'elles créent pour des vies humaines. Dans cette optique, ils devraient être conçus pour permettre aux gens de comprendre avec succès les systèmes d'IA, de participer à leur utilisation et de créer leur confiance.

Les politiques publiques devraient contribuer à faciliter l'adaptation de la société aux applications de l'IA, étendre leurs avantages et atténuer leurs erreurs et leurs échecs inévitables. On devrait encourager les débats sur la façon dont l'IA est déployée, y compris les préoccupations concernant la protection de la vie privée et les avantages d'IA. Compte tenu de la rapidité avec laquelle les technologies de l'IA sont réalisées et des préoccupations concomitantes concernant leurs implications, le Panel d'étude recommande que toutes les couches du gouvernement acquièrent une expertise technique en IA. De plus, la recherche sur l'équité, la sécurité, la vie privée et les implications sociétales des systèmes d'IA devraient être encouragées en supprimant les obstacles et en augmentant les dépenses publiques et privées pour l'appuyer. Actuellement aux États-Unis, au moins seize agences distinctes gèrent des secteurs de l'économie liés aux technologies de l'IA. Les progrès rapides dans la recherche sur l'IA et, en particulier, ses applications exigent des experts dans ces secteurs pour développer de nouveaux concepts et métaphores pour le droit et la politique.

Qui est responsable lorsqu'une voiture auto-auto s'arrête ou qu'un dispositif médical intelligent échoue?

Comment les applications de l'IA peuvent-elles être empêchées de promulguer une discrimination raciale ou une tricherie financière?

Qui devrait profiter des gains d'efficience permis par les technologies de l'AI et quelles protections devraient être accordées aux personnes dont les compétences sont rendues obsolètes?

Au fur et à mesure que les gens intègrent plus largement et profondément les processus industriels et les produits de consommation, les pratiques exemplaires doivent être réparties et les régimes réglementaires sont adaptés. Bien que le Panel d'étude ne juge pas probable que les systèmes d'IA à court terme choisissent de manière autonome de porter préjudice aux personnes, il sera possible aux personnes d'utiliser des systèmes basés sur l'IA pour des fins dangereuses et utiles. Et bien que les algorithmes d'IA puissent être capables de prendre des décisions moins biaisées qu'une personne typique, il reste un défi technique profond pour s'assurer que les données qui éclairent les décisions basées sur l'IA peuvent être exemptes de biais qui pourraient conduire à une discrimination fondée sur l'orientation de la race, la sexualité ou d'autres facteurs. Face aux changements profonds que les technologies de l'IA peuvent produire, la pression pour une réglementation «plus» et de «plus en plus difficile» est probablement inévitable. Les malentendus sur ce qu'est l'IA et ce ne sont pas susceptibles d'alimenter l'opposition aux technologies susceptibles de bénéficier à tous. Une activité réglementaire inappropriée serait une erreur tragique.

Une réglementation mal informée qui étouffe l'innovation, ou la relocalise dans d'autres juridictions, serait contre-productive.

2 - Heureusement, les principes qui guident la réglementation réussie des technologies numériques actuelles constituent un point de départ. En matière de réglementation de la vie privée, de larges mandats légaux couplés à des exigences de transparence rigoureuses et à une application significative, plutôt que des contrôles stricts, encouragent les entreprises à développer des processus et du personnel professionnel pour faire respecter les contrôles de la vie privée, s'engager avec des intervenants externes et adapter leurs pratiques aux avancées technologiques. Cela, à son tour, appuie le développement d'associations commerciales professionnelles et de comités de normes qui diffusent les meilleures pratiques. Dans l'IA, les régulateurs peuvent également renforcer un cycle d'activité vertueux impliquant la responsabilisation interne, externe, la transparence et la professionnalisation plutôt que la conformité étroite. Un débat vigoureux et éclairé sur la meilleure façon de diriger l'IA de manière à enrichir nos vies et notre société, tout en encourageant la créativité sur le terrain, est un besoin urgent et vital.

Les technologies de l'IA pourraient élargir les inégalités d'opportunité existantes si l'accès à celles-ci - ainsi que le calcul à haute puissance et les données à grande échelle qui alimentent beaucoup d'entre elles - sont distribués de manière injustifiée dans toute la Société. Ces technologies amélioreront la vie de tous.

2 - Kate Crawford, «Problème de Guy White Intelligence», The New York Times, 25 juin 2016, accessible le 1er août 2016,

Les malentendus sur ce qu'est l'IA et ce ne sont pas susceptibles d'alimenter l'opposition aux technologies susceptibles de bénéficier à tous. Une réglementation mal informée qui étouffe l'innovation serait une erreur tragique.

12 - 13 - La mesure humaine Notamment, la caractérisation de l'intelligence comme spectre n'accorde aucun statut spécial au cerveau humain. Mais à ce jour, l'intelligence humaine n'a pas de correspondance dans les mondes biologiques et artificiels pour une polyvalence extrême, avec les capacités "à raisonner, à atteindre les objectifs, à comprendre et à générer du langage, à percevoir et à répondre aux entrées sensorielles, à prouver des théorèmes mathématiques, à jouer à des jeux difficiles, à synthétiser et à Résumer l'information, créer de l'art et de la musique, et même écrire des histoires. "

6 - Cela rend l'intelligence humaine à un choix naturel pour comparer les progrès de l'IA. Il peut même être proposé, en règle générale, que toute activité que les ordinateurs puissent effectuer et les personnes une fois exécutées doit être compté comme une instance d'intelligence. Mais associer n'importe quelle capacité humaine n'est qu'une condition suffisante, pas nécessaire. Il existe déjà de nombreux systèmes qui dépassent l'intelligence humaine, au moins en vitesse, comme la planification des arrivées quotidiennes et des départs de milliers de vols dans un aéroport. La longue quête d'IA - et le succès final - pour battre les joueurs humains au jeu d'échecs a offert une instance de haut niveau pour comparer l'intelligence humaine à l'intelligence mécanique. L'échec a fasciné les gens pendant des siècles. Lorsque la possibilité de construire des ordinateurs est devenue imminente, Alan Turing, dont beaucoup considèrent le père de l'informatique, "a mentionné l'idée que les ordinateurs montrent l'intelligence avec l'échec comme paradigme".

7 - Sans accès à des ordinateurs puissants, "Turing a joué un jeu dans lequel il a simulé l'ordinateur, prenant environ une demi-heure par mouvement". Mais ce n'est qu'après une longue série d'améliorations dans les années soixante et soixante-dix - apporté par les groupes chez Carnegie Mellon , Stanford, MIT, l'Institut de physique théorique et expérimentale à Moscou et Northwestern University, que les programmes d'équitation ont commencé à gagner leur compétence. La dernière poussée a traversé un projet de longue durée à IBM, qui a culminé avec le programme Deep Blue battant Garry Kasparov, alors champion du monde d'échecs, d'un score de 3,5-2,5 en 1997. Curieusement, dès que l'IA a rattrapé son Une cible évasive que Deep Blue a été décrite comme une collection de "méthodes de force brute" qui n'était pas une "intelligence réelle".

8 - En fait, la publication ultérieure d'IBM sur Deep Blue, qui donne des détails détaillés sur ses procédures de recherche et d'évaluation, ne mentionne pas le mot «intelligent» même une seule fois.

9 - Was Deep Blue était-il intelligent ou non? Une fois de plus, la frontière avait déménagé. Une définition opérationnelle IA peut également être définie par ce que font les chercheurs d'IA. Ce rapport considère IA principalement comme une branche de l'informatique qui étudie les propriétés de l'intelligence en synthétisant l'intelligence.

10 - Bien que l'avènement d'IA ait dépendu de la progression rapide des ressources informatiques matérielles, l'accent ici sur les logiciels reflète une tendance dans la communauté de l'IA. Plus récemment, les progrès réalisés dans la construction de matériel adapté à l'informatique basée sur le réseau neuronal ont créé

6 - Nilsson, The Question for Artificial Intelligence.

7 - Nilsson, The Question for Artificial Intelligence,

89. 8 - McCorduck, Machines Who Think, 433.

9 - Murray Campbell, A. Joseph Hoane Jr. et Feng-hsiung Hsu, «Deep Blue», Intelligence artificielle 134, n. 1 et 2 (2002): 57-83.

10 - Herbert A. Simon, «Intelligence artificielle: une science empirique», Intelligence artificielle, 77, no. 2 (1995):

95-127. 11 - Paul Merolla John V. Arthur, Rodrigo Alvarez-Icaza, Andrew S. Cassidy, Jun Sawada, Filipp Akopyan, Bryan L. Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, Bernard Brezzo, Ivan Vo, Steven K. Esser, Rathinakumar Appuswamy, Brian Taba, Arnon Amir, Myron D. Flickner, William P. Risk, Rajit Manohar et Dharmendra S. Modha, «Un circuit intégré de Million Spiking-Neuron avec un réseau et une interface de communication évolutifs», consulté le 1er août 2016.

SECTION I: QU'EST-CE QUE L'INTELLIGENCE ARTIFICIELLE ?

Cette section décrit comment les chercheurs et les praticiens définissent «Intelligence artificielle» et les domaines de la recherche et de l'application de l'IA qui se développent actuellement. Il profite des définitions de ce que l'IA est et n'est pas, et décrit certaines des zones actuellement «chaudes» de la recherche sur l'IA. Cette section pose les bases de la section II, qui décrit les impacts de l'IA et son avenir dans huit domaines et la section III, qui décrit les problèmes liés à la conception de l'IA et à la politique publique et fait des recommandations pour encourager l'innovation de l'IA tout en protégeant les valeurs démocratiques.

DÉFINIR L'INTELLIGENCE ARTIFICIELLE :

Curieusement, l'absence d'une définition précise et universellement acceptée de l'IA a probablement aidé le secteur à grandir, à s'épanouir et à progresser à un rythme toujours accéléré. Les praticiens, les chercheurs et les développeurs de l'IA sont plutôt guidés par un sens brutal de la direction et un impératif de «s'entendre avec lui». Pourtant, une définition reste importante et Nils J. Nilsson a fourni une utilité: « L'intelligence artificielle est celle l'activité consacrée à la fabrication de machines intelligentes, et l'intelligence est cette qualité qui permet à une entité de fonctionner de manière appropriée et avec une prévoyance dans son environnement ".

3 - Dans cette perspective, la caractérisation de l'IA dépend du crédit qui est prêt à donner des logiciels et du matériel synthétisés pour fonctionner «de manière appropriée» et avec «la prévoyance». Une simple calculatrice électronique effectue des calculs beaucoup plus rapide que le cerveau humain et ne se confie presque jamais .

4 - Une calculatrice est-elle intelligente? Comme Nilsson, le Panel d'étude considère que le renseignement repose sur un spectre multidimensionnel. Selon cette vue, la différence entre un calculateur arithmétique et un cerveau humain n'est pas unique, mais d'échelle, de vitesse, de degré d'autonomie et de généralité. Les mêmes facteurs peuvent être utilisés pour évaluer toute autre instance de logiciel de reconnaissance de la parole, des cerveaux, des systèmes de contrôle de croisière dans les voitures, des programmes de lecture, des thermostats et de les placer à un emplacement approprié dans le spectre. Bien que notre large interprétation place le calculateur dans le spectre du renseignement, de tels dispositifs simples ressemblent peu à l'IA d'aujourd'hui. La frontière de l'IA s'est bien avancée et les fonctions de la calculatrice ne représentent qu'un parmi les millions que les smartphones d'aujourd'hui peuvent effectuer. Les développeurs d'IA travaillent maintenant sur l'amélioration, la généralisation et la mise à l'échelle de l'intelligence actuellement trouvée sur les smartphones. En fait, le domaine de l'IA est une tentative continue de repousser la frontière de l'intelligence de la machine. Ironiquement, IA souffre du destin perpétuel de perdre ses revendications, ce qui, finalement et inévitablement, est tiré à l'intérieur de la frontière, un motif répétitif connu sous le nom de «l'effet IA» ou le «paradoxe impair» - L'IA apporte une nouvelle technologie dans le pli commun , les gens s'habituent à cette technologie, il cesse d'être considéré comme une IA et une nouvelle technologie émerge.

5 - Le même modèle se poursuivra dans le futur. L'IA ne "livre" aucun produit qui change la vie d'un boulon en bleu. Au contraire, les technologies IA continuent à s'améliorer de manière continue et progressive.

3 - Nils J. Nilsson, La quête de l'intelligence artificielle: une histoire d'idées et de réalisations (Cambridge, UK: Cambridge University Press, 2010).

4 - Wikimedia Images, consulté le 1er août 2016.

5 - Pamela McCorduck, machines qui pensent: une enquête personnelle sur l'histoire et les perspectives de l'intelligence artificielle, 2 e éd. (Natick, MA: A. K. Peters, Ltd., 2004; San Francisco: W. H. Freeman, 1979), les citations portent sur l'édition de Peters. Le renseignement repose sur un spectre multidimensionnel. Selon cette vue, la différence entre un calculateur arithmétique et un cerveau humain n'est pas unique, mais d'échelle, de vitesse, de degré d'autonomie et de généralité. Une image précise et sophistiquée d'IA-One qui est en concurrence avec sa représentation populaire - est entravée par la difficulté de fixer une définition précise de l'intelligence artificielle.

14 15 - l'étiquetage, la reconnaissance d'activité, et plusieurs variantes de celles-ci. L'apprentissage approfondi fait également des progrès importants dans d'autres domaines de la perception, tels que l'audio, la parole et le traitement du langage naturel. Apprentissage de renforcement alors que l'apprentissage par machine traditionnel a principalement porté sur l'extraction de motifs, l'apprentissage par renforcement met l'accent sur la prise de décision qui est une technologie qui aidera l'IA à progresser plus profondément dans le domaine de l'apprentissage et de l'exécution des actions dans le monde réel. Il existe depuis plusieurs décennies comme un cadre pour la prise de décision séquentielle axée sur l'expérience, mais les méthodes n'ont pas connu un grand succès dans la pratique, principalement en raison de problèmes de représentation et de mise à l'échelle.

Cependant, l'avènement de l'apprentissage en profondeur a permis de renforcer l'apprentissage avec un "tir dans le bras". Le succès récent d'AlphaGo, un programme informatique développé par Google Deepmind qui a battu le champion Go humain dans un match de cinq matchs, était dû en grande quantité de la partie à l'apprentissage de renforcement.

AlphaGo a été formé en initialisant un agent automatisé avec une base de données d'experts humains, mais a ensuite été raffiné en jouant un grand nombre de jeux contre lui-même et en appliquant l'apprentissage de renforcement. Robotique. La navigation robotique, au moins dans des environnements statiques, est largement résolue. Les efforts actuels considèrent comment former un robot pour interagir avec le monde qui l'entoure de manière généralisable et prévisible. Une exigence naturelle qui se pose dans des environnements interactifs est la manipulation, un autre sujet d'intérêt actuel. La révolution de l'apprentissage profond ne fait que commencer à influencer la Robotique, en grande partie parce qu'il est beaucoup plus difficile d'acquérir les grands ensembles de données marquées qui ont orienté d'autres domaines d'apprentissage axés sur l'apprentissage. L'apprentissage par renforcement (voir ci-dessus), qui évite l'exigence de données marquées, peut aider à combler cette lacune, mais nécessite que les systèmes puissent explorer en toute sécurité un espace politique sans commettre d'erreurs qui nuisent au système lui-même ou à d'autres. Les progrès dans la perception de la machine fiable, y compris la vision par ordinateur, la force et la perception tactile, dont une grande partie sera guidée par l'apprentissage par machine, continueront d'être des facteurs clés pour améliorer les capacités de la Robotique.

Vision informatique :

La vision par ordinateur est actuellement la forme la plus importante de la perception de la machine. Il a été la sous-zone de l'IA la plus transformée par la montée de l'apprentissage approfondi. Jusqu'à il y a quelques années, les machines à vecteurs de soutien étaient la méthode de choix pour la plupart des tâches de classification visuelle. Mais la confluence de l'informatique à grande échelle, en particulier sur les GPU, la disponibilité de gros ensembles de données, en particulier via Internet, et les raffinements des algorithmes de réseaux neuronaux ont entraîné des améliorations spectaculaires dans les performances sur les tâches de référence (par exemple, la classification sur ImageNet17).

Pour la première fois, les ordinateurs sont en mesure d'exécuter certaines tâches de classification visuelle (étroitement définies) mieux que les personnes. La recherche en cours est axée sur l'image automatique et le sous-titrage vidéo.

Traitement du langage naturel :

En général associé à la reconnaissance automatique de la parole, le traitement du langage naturel est une autre partie très active de la perception de la machine. Il devient rapidement une marchandise pour les langues traditionnelles avec de grands ensembles de données. Google a annoncé que 20% des requêtes mobiles actuelles sont effectuées par voix,

18 - et des démonstrations récentes ont démontré la possibilité d'une traduction en temps réel. La recherche se déplace maintenant vers le développement de systèmes raffinés et capables qui peuvent interagir avec les personnes à travers le dialogue, et non seulement réagir aux demandes stylisées.

17 - ImageNet, Stanford Vision Lab, Université de Stanford, Université de Princeton, 2016, consulté le 1er août 2016, 

http://www.image-net.org/

18 - Greg Sterling, "Google dit que 20% des requêtes mobiles sont des recherches vocales", Search Engine Land, le 18 mai 2016, a eu accès au 1er août 2016.

Un couplage plus étroit entre le matériel et les logiciels dans l'avancement de l'IA :

L'«intelligence» reste un phénomène complexe dont les aspects variés ont attiré l'attention de plusieurs domaines d'études, y compris la psychologie, l'économie, les neurosciences, la biologie, l'ingénierie, les statistiques et la linguistique. Naturellement, le domaine de l'IA a bénéficié des progrès réalisés par tous ces champs alliés. Par exemple, le réseau neuronal artificiel, qui a été au cœur de plusieurs solutions basées sur l'IA12 13, a été initialement inspiré par des réflexions sur le flux d'informations dans les neurones biologiques.

14 - TENDANCES DE RECHERCHE D'IA : 

Jusqu'au début du millénaire, l'appel d'IA était largement dans sa promesse de livrer de la construction de systèmes intelligents, mais au cours des quinze dernières années, une grande partie de cette promesse a été rachetée.

15. - Les technologies de l'IA ont déjà pénétré nos vies. À mesure qu'ils deviennent une force centrale dans la Société, le domaine passe de la construction de systèmes intelligents pour la construction de systèmes intelligents qui sont conscients de l'homme et de la confiance. Plusieurs facteurs ont alimenté la révolution de l'IA. Le plus important d'entre eux est la maturation de l'apprentissage par machine, soutenue en partie par les ressources de "cloud computing" et la collecte de données Web largement répandue. L'apprentissage par machine a été propulsé de manière dramatique par «apprentissage approfondi», une forme de réseaux de neurones artificiels adaptatifs formés à l'aide d'une méthode appelée backpropagation.

16 - Ce saut dans la performance des algorithmes de traitement de l'information s'est accompagné de progrès significatifs dans la technologie matérielle pour les opérations de base comme détection, perception et reconnaissance d'objets. Les nouvelles plates-formes et marchés pour les produits axés sur les données et les incitations économiques pour trouver de nouveaux produits et marchés ont également contribué à l'avènement de la technologie axée sur l'IA. Toutes ces tendances conduisent les domaines de recherche «chauds» décrits ci-dessous.

Cette compilation vise simplement à refléter les domaines qui, d'une métrique ou autre, reçoivent actuellement une plus grande attention que d'autres. Ils ne sont pas nécessairement plus importants ou précieux que d'autres. En effet, certaines des régions actuellement «chaudes» étaient moins populaires au cours des dernières années, et il est probable que d'autres régions ressembleront de manière similaire à l'avenir.

L'apprentissage en machine à grande échelle :

Beaucoup des problèmes de base de l'apprentissage par machine (comme l'apprentissage supervisé et non surveillé) sont bien compris. L'effort majeur des efforts actuels consiste à mettre à l'échelle les algorithmes existants pour fonctionner avec des ensembles de données extrêmement importants. Par exemple, alors que les méthodes traditionnelles pouvaient faire plusieurs passages sur l'ensemble de données, les modernes sont conçus pour ne faire qu'une seule passe. Dans certains cas, seules les méthodes sublinaires (celles qui ne regardent qu'une fraction des données) peuvent être admises.

Apprentissage approfondi :

La capacité de former avec succès des réseaux de neurones convolutionnels a le plus bénéficié du domaine de la vision par ordinateur, avec des applications telles que la reconnaissance d'objets, la vidéo

12 - Gerald Tesauro, «Problèmes pratiques dans l'apprentissage de la différence temporelle», Apprentissage automatique, no. 8 (1992):

257-77. 13 - David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel et Demis Hassabis, "Maîtriser le jeu de Go avec des réseaux de neurones profonds et la recherche d'arbres", Nature 529 (2016):

484-489. 14 - W. McCulloch et W. Pitts, W., "Un calcul logique des idées immanentes dans l'activité nerveuse", Bulletin de Biophysique Mathématique, 5 (1943):

115-133. 15 - L'annexe I offre une courte histoire de l'IA, y compris une description de certains des domaines de recherche traditionnellement essentiels, qui ont changé au cours des six dernières décennies.

16 - La rétroprojection est une abréviation de "propagation vers l'arrière des erreurs", une méthode commune de formation de réseaux neuronaux artificiels utilisés en conjonction avec une méthode d'optimisation telle que la descente en gradient. La méthode calcule le dégradé d'une fonction de perte par rapport à tous les poids du réseau. L'intelligence humaine n'a pas de correspondance dans les mondes biologiques et artificiels pour une polyvalence extrême, avec les capacités «à raisonner, à atteindre les objectifs, à comprendre et à générer du langage ... créer de l'art et de la musique, et même écrire des histoires».

Les technologies de l'IA absorbent déjà nos vies. Comme ils deviennent une force centrale dans la société, le domaine évolue de la construction simple de systèmes intelligents pour construire des systèmes intelligents qui sont conscients de l'homme et de la confiance.

16 17 - Poursuivant activement des modèles alternatifs d'informatique - en particulier ceux qui s'inspirent de ce qu'on sait sur les réseaux de neurones biologiques - dans le but d'améliorer l'efficacité matérielle et la robustesse des systèmes informatiques. À l'heure actuelle, de tels ordinateurs « neuromorphes » n'ont pas encore clairement démontré de grands gains et commencent tout juste à devenir commercialement viables. Mais il est possible qu'ils deviennent monnaie courante (même si ce n'est que des ajouts à leurs cousins ​​de Von Neumann) dans un proche avenir.

Les réseaux nerveux profonds ont déjà créé un éclaboussement dans le paysage de l'application. Une onde plus grande peut toucher lorsque ces réseaux peuvent être formés et exécutés sur du matériel neuromorphe dédié, par opposition à simuler sur les architectures standard de Von Neumann, comme elles le sont aujourd'hui. Les tendances générales et l'avenir de la recherche sur l'IA Le succès retentissant du paradigme axé sur les données a déplacé les paradigmes traditionnels de l'IA. Les procédures telles que la démonstration du théorème et la représentation et le raisonnement de la connaissance basée sur la logique reçoivent une attention réduite, en partie en raison du défi continu de se connecter aux mises à la terre du monde réel. La planification, qui a été le fondement de la recherche sur l'IA dans les années soixante-dix et quatre-vingts, a également reçu moins d'attention en retard en partie grâce à sa forte dépendance à l'égard des hypothèses de modélisation difficiles à satisfaire dans les applications réalistes.

Les approches basées sur les modèles, telles que les approches basées sur la physique de la vision et le contrôle et la cartographie traditionnels en Robotique, ont en grande partie donné lieu à des approches axées sur les données qui ferment la boucle pour détecter les résultats des actions dans la tâche en cours. Le raisonnement bayésien et les modèles graphiques, très populaires même tout récemment, semblent également sortir de la faveur, avoir été noyés par le déluge de données et le succès remarquable de l'apprentissage en profondeur. Au cours des quinze prochaines années, le Comité d'étude s'attend à ce que l'accent soit mis sur le développement de systèmes qui soient conscients de l'Humanité, ce qui signifie qu'ils spécifient spécifiquement et sont spécifiquement conçus pour les caractéristiques des personnes avec lesquelles ils sont censés interagir. Il y a beaucoup d'intérêt à essayer de trouver de nouvelles façons créatives de développer des moyens interactifs et évolutifs pour enseigner des Robots. En outre, les systèmes de type IoT - les appareils et le Cloud - deviennent de plus en plus populaires, tout comme la pensée des dimensions sociales et économiques de l'IA.

Au cours des prochaines années, de nouvelles capacités de reconnaissance de la perception / objet et des plates-formes robotiques sécuritaires pour l'homme augmenteront, tout comme les produits basés sur les données et leurs marchés. Le Comité d'étude s'attend également à une réémergence de certaines des formes traditionnelles de l'IA, car les praticiens se rendent compte des limites inévitables des approches d'approfondissement purement de bout en bout. Nous encourageons les jeunes chercheurs à ne pas réinventer la roue, mais plutôt à prendre conscience des progrès significatifs dans de nombreuses régions de l'IA au cours des cinquante premières années du domaine et dans des domaines connexes tels que la théorie du contrôle, la science cognitive et la psychologie.

Systèmes collaboratifs :

La recherche sur des systèmes collaboratifs étudie des modèles et des algorithmes pour aider à développer des systèmes autonomes qui peuvent fonctionner en collaboration avec d'autres systèmes et avec des humains. Cette recherche repose sur l'élaboration de modèles formels de collaboration et étudie les capacités nécessaires pour que les systèmes deviennent des partenaires efficaces. On s'intéresse de plus en plus aux applications qui peuvent utiliser les forces complémentaires des humains et des machines - pour les humains pour aider les systèmes d'IA à surmonter leurs limites et pour que les agents augmentent les capacités et les activités humaines. Le crowdsourcing et le calcul humain, étant donné que les capacités humaines sont supérieures aux méthodes automatisées pour accomplir de nombreuses tâches, la recherche sur le crowdsourcing et le calcul humain étudie les méthodes pour augmenter les systèmes informatiques en utilisant l'intelligence humaine pour résoudre les problèmes que les ordinateurs ne peuvent pas résoudre correctement.

Introduit il y a environ une quinzaine d'années, cette recherche a maintenant une présence établie dans l'IA. L'exemple le plus connu du crowdsourcing est Wikipedia, un référentiel de connaissances qui est maintenu et mis à jour par les internautes et dépasse de loin les sources d'information compilées traditionnellement, telles que les encyclopédies et les dictionnaires, à l'échelle et à la profondeur. Crowdsourcing se concentre sur l'élaboration de moyens novateurs pour exploiter l'intelligence humaine. Les plates-formes de science citoyenne dynamisent les bénévoles pour résoudre des problèmes scientifiques, tandis que les plates-formes de crowdsourcing payées telles que Amazon Mechanical Turk fournissent un accès automatisé à l'intelligence humaine à la demande. Les travaux dans ce domaine ont facilité les progrès dans d'autres sous-domaines de l'IA, y compris la vision par ordinateur et la PNL, en permettant de recueillir de nombreuses quantités de données de formation marquées et / ou de données d'interaction humaines dans un court laps de temps. Les efforts de recherche actuels explorent les divisions idéales des tâches entre les humains et les machines en fonction de leurs capacités et coûts différents.

La théorie du jeu algorithmique et le choix social du calcul :

Une attention particulière est portée sur les dimensions informatiques économiques et sociales de l'IA, y compris les structures d'incitation. L'IA distribuée et les systèmes multi-agents ont été étudiés depuis le début des années 1980, ont gagné en évidence à partir de la fin des années 1990 et ont été accélérés par Internet. Une exigence naturelle est que les systèmes traitent des incitations potentiellement désalignées, y compris des participants ou des entreprises humaines intéressés, ainsi que des agents automatisés basés sur l'IA qui les représentent. 

Les sujets qui reçoivent l'attention comprennent la conception du mécanisme de calcul (une théorie économique de la conception d'incitation, la recherche de systèmes compatibles avec les incitatifs où les intrants sont réellement signalés), le choix social de calcul (une théorie sur la façon d'agréger les ordres de classement sur les alternatives), l'incitation à l'information , les règles de notation, la prédiction par les pairs) et la théorie des jeux algorithmiques (équilibre des marchés, des jeux en réseau et des jeux de société tels que Poker, un jeu où des progrès significatifs ont été réalisés ces dernières années grâce à des techniques d'abstraction et à l'apprentissage sans regret).

Internet of Things (IoT) ( l'Internet des Objets ) :

Un nombre croissant de recherches est consacré à l'idée qu'un large éventail de périphériques peuvent être interconnectés pour collecter et partager leurs informations sensorielles. De tels appareils peuvent inclure des appareils, des véhicules, des bâtiments, des caméras et d'autres choses. Bien qu'il soit question de technologie et de réseau sans fil pour connecter les périphériques, L'IA peut traiter et utiliser les énormes quantités de données qui en résultent à des fins intelligentes et utiles. 

À l'heure actuelle, ces appareils utilisent une gamme déconcertante de protocoles de communication incompatibles. L'IA pourrait aider à dompter cette Tour de Babel.

Informatique Neuromorphique :

Les ordinateurs traditionnels mettent en œuvre le modèle d'informatique de Von Neumann, qui sépare les modules d'entrée / sortie, de traitement d'instructions et de mémoire. Avec le succès des réseaux de neurones profonds sur un large éventail de tâches, les fabricants sont de plus en plus nombreux. Une recherche croissante est consacrée à l'idée qu'un large éventail de périphériques peuvent être interconnectés pour collecter et partager leurs informations sensorielles. De tels appareils peuvent inclure des appareils, des véhicules, des bâtiments, des caméras et d'autres choses. Le traitement du langage naturel est un domaine très actif de la perception de la machine. La recherche se déplace maintenant vers le développement de systèmes capables d'interagir avec les personnes à travers le dialogue, et non seulement de réagir aux demandes stylisées.

18 19 - Ils avaient déjà un certain nombre de fonctionnalités qui ont combiné la détection en temps réel avec la perception et la prise de décision telles que les systèmes de freinage antiblocage (ABS), le contrôle des airbags, les systèmes de contrôle de traction (TCS) et le contrôle de stabilité électronique (ESC).

22 - Les capacités automatisées ont été introduites progressivement dans les voitures commerciales depuis 2003, comme résumé dans le tableau suivant. Contexte Fonctionnalité automatisée Date de sortie Stationnement Système de stationnement intelligent depuis 2003.

23 - Invocation du stationnement depuis 2016.

24 - Système de départ de l'autoroute et de l'autoroute depuis 2004 en Amérique du Nord.

25 - Contrôle de la vitesse de l'aviation Arterial & Highway Depuis 2005 en Amérique du Nord.

26 - Surveillance des spots aveugles depuis 2007

27 - Changement d'autoroute depuis 2015.

28 - Ces fonctionnalités assistent des conducteurs ou reprennent complètement les activités bien définies pour une sécurité et un confort accrus.

Les voitures actuelles peuvent se garer elles-mêmes, effectuer un contrôle de croisière adapté sur les autoroutes, se diriger pendant le trafic stop-and-go et alerter les pilotes sur les objets dans des endroits obscènes pendant les changements de voie. La technologie de la vision et du radar a été exploitée pour développer des systèmes de pré-collision qui permettent aux voitures de se freiner de manière autonome lorsque le risque de collision est détecté. L'apprentissage approfondi a également été appliqué pour améliorer la capacité des automobiles à détecter les objets dans l'environnement et à reconnaître le son.

29 - Véhicules à conduite autonome depuis les années 1930, les écrivains de science-fiction ont rêvé d'un avenir avec des voitures auto-conduisant, et leur construction a été un défi pour la communauté de l'IA depuis les années 1960. Au cours des années 2000, le rêve des véhicules autonomes est devenu une réalité dans la mer et le ciel, et même sur Mars, mais les voitures auto-conduisant n'existaient que comme prototypes de recherche dans les laboratoires.

Conduire dans une ville a été considéré comme un problème trop complexe pour l'automatisation en raison de facteurs comme les piétons, le trafic lourd et les nombreux événements imprévus qui peuvent se produire en dehors du contrôle de la voiture. Bien que les composants technologiques sont nécessaires à la conduite d'un véhicule.

22 - Carl Liersch, "Vehicle Technology Timeline: From Automated to Driverless", Robert Bosch (Australie) Pty. Ltd., 2014, a accédé au 1er août 2016,

23 - "Intelligent Parking Assist System", Wikipedia, dernière modification le 26 juillet 2016, accessible le 1er août 2016,

https://en.wikipedia.org/wiki/Intelligent_Parking_Assist_System

24 - L'équipe Tesla Motors, «Invoquez votre Tesla à partir de votre téléphone», Tesla, le 10 janvier 2016, a eu accès au 1er août 2016,

https://www.teslamotors.com/blog/summon-your-tesla-your-phone

25 - Système d'alerte de départ de voie ", Wikipedia, dernière modification le 24 juillet 2016, accessible le 1er août 2016,

https://en.wikipedia.org/wiki/Lane_departure_warning_system

26 - "Système autonome de contrôle de la croisière", Wikipedia, dernière modification du 30 juillet 2016, accessible le 1er août 2016,

https://en.wikipedia.org/wiki/Autonomous_cruise_control_system

27 - "écran aveugle", Wikipedia, dernière modification le 20 avril 2016, accessible le 1er août 2016,

https://en.wikipedia.org/wiki/Blind_spot_monitor

28 - Dana Hull, "Tesla démarre les fonctionnalités du pilote automatique", Boomberg Technology, le 14 octobre 2015, a eu accès au 1er août 2016,

29 - Aaron Tilley, "Nouvelle puce Qualcomm apporte un apprentissage approfondi aux voitures", Forbes, le 5 janvier 2016, a eu accès au 1er août 2016.

SECTION II: AI PAR DOMAINE :

Bien que différents exemples de recherche et de pratique de l'IA partagent des technologies communes, comme l'apprentissage par machine, elles varient également considérablement dans les différents secteurs de l'économie et de la société. Nous appelons ces secteurs «domaines», et dans cette section décrivons les différents états de la recherche et de la mise en œuvre de l'IA, ainsi que les impacts et les défis distincts, dans huit d'entre eux: le transport; la Robotique à domicile / service; soins de santé; éducation; Communautés à faible ressources; Sécurité publique; Emploi et lieu de travail; tt le divertissement. Sur la base de ces analyses, nous prévoyons également les tendances dans une ville typiquement nord-américaine au cours des quinze prochaines années. Contrairement à la représentation typique d'IA dans la Culture populaire, nous cherchons à proposer un aperçu équilibré des façons dont IA commence déjà à transformer la vie quotidienne et à la façon dont ces transformations devraient augmenter d'ici 2030.

LES TRANSPORTS : 

Le transport est susceptible d'être un des premiers domaines dans lesquels le grand public sera invité à faire confiance à la fiabilité et à la sécurité d'un système d'IA pour une tâche critique. Le transport autonome sera bientôt un lieu commun et, comme la première expérience de la plupart des personnes avec les systèmes d'IA physiquement incorporés, influencera fortement la perception de l'IA par le public.

Une fois que le matériel physique est rendu suffisamment sûr et robuste, son introduction à la vie quotidienne peut se produire si soudainement pour surprendre le public, ce qui nécessitera un temps d'ajustement. Comme les voitures deviendront de meilleurs conducteurs que les gens, les citadins posséderont moins de voitures, vivront plus loin de leur lieu de travail et passeront le temps différemment, ce qui entraînera une organisation urbaine entièrement nouvelle. De plus, dans la ville typiquement nord-américaine en 2030, les changements ne seront pas limités aux voitures et aux camions, mais ils sont susceptibles d'inclure des véhicules volants et des Robots personnels, et augmenteront les problèmes sociaux, éthiques et politiques. Quelques technologies clés ont déjà catalysé l'adoption généralisée de l'IA dans les transports. Par rapport à 2000, l'ampleur et la diversité des données sur le transport personnel et le transport à la population disponibles aujourd'hui, grâce à l'adoption de smartphones, à la diminution des coûts et à l'amélioration de la précision de la variété des capteurs, sont étonnantes. Sans la disponibilité de ces données et de la connectivité, des applications telles que la détection en temps réel et la prédiction du trafic, des calculs d'itinéraires, des déplacements de pair à point et des voitures auto-conduisant ne seraient pas possibles.

LES GPS INTELLIGENTS : 

Les GPS plus intelligents ont été introduits dans les véhicules personnels en 2001 avec des dispositifs de navigation dans la voiture et sont depuis devenu une partie fondamentale de l'infrastructure de transport.19 Le GPS aide les conducteurs tout en fournissant des informations à grande échelle aux entreprises technologiques et aux villes sur les modes de transport. L'adoption généralisée des smartphones avec la technologie GPS accroît encore la connectivité et la quantité de données de localisation partagée par les particuliers. Les véhicules actuels sont également équipés d'une large gamme de capacités de détection. Une automobile moyenne aux États-Unis devrait contenir soixante dix capteurs, y compris les gyroscopes, les accéléromètres, les capteurs de lumière ambiante et les capteurs d'humidité20. Les capteurs ne sont pas nouveaux pour les véhicules. Les automobiles construites avant 2000 avaient des capteurs pour l'état interne du véhicule, comme sa vitesse, son accélération et sa position de roue.21 19 Mark Sullivan, «Une brève histoire du GPS», PCWorld, le 9 août 2012, a eu accès au 1er août 2016.

http://www.pcworld.com/article/2000276/a-brief-history-of-gps.html

20 - William J. Fleming, «Nouveaux capteurs automobiles - Une revue», IEEE Sensors Journal 8, no 11, (2008): 1900-1921. 21 Jean Jacques Meneu, éd., "Automotive Sensors: Now and in the Future", Arrow, le 24 septembre 2015, a eu accès au 1er août 2016.

https://www.arrow.com/fr/research-and-events/articles/automotivesensors-now-and-in-the-future 

Comme les voitures deviendront de meilleurs conducteurs que les gens, les habitants de la ville détiendront moins de voitures, vivront plus loin du travail et passeront le temps différemment, ce qui entraînera une organisation urbaine entièrement nouvelle. Le transport autonome sera bientôt un lieu commun et, comme la première expérience de la plupart des personnes avec les systèmes d'IA physiquement incorporés, influencera fortement la perception de l'IA par le public.

20 21 - Les passagers aux États-Unis passent vingt-cinq minutes à parcourir chaque trajet.

37 - Avec la technologie auto-conduite, les gens auront plus de temps à travailler ou se divertiront pendant leurs déplacements. Et le confort accru et la diminution de la charge cognitive avec les voitures auto-conduisant et le transport partagé peuvent affecter l'endroit où les gens choisissent de vivre. Le besoin réduit de stationnement peut affecter la façon dont les villes et les espaces publics sont conçus. Les voitures auto-conduisant peuvent également servir à accroître la liberté et la mobilité des différents sous-groupes de la population, y compris les jeunes, les personnes âgées et les handicapés. Les voitures auto-conduisant et les services de transport par les pairs peuvent éliminer la nécessité de posséder un véhicule.

L'effet sur l'utilisation totale de l'automobile est difficile à prévoir. Les déplacements de véhicules vides et la volonté accrue des gens de voyager peuvent conduire à des kilomètres plus élevés. Alternativement, les véhicules autonomes partagés - les personnes qui utilisent les voitures comme un service plutôt que de posséder les leurs - peuvent réduire le nombre total de kilomètres , surtout si elles sont combinées avec des incitations bien construites, comme les péages ou les remises, pour répartir la demande de voyage, partager les déplacements et réduire les congestions . La disponibilité du transport partagé peut déplacer le besoin de transport public - ou les transports publics peuvent changer de forme vers le transit personnel rapide, déjà disponible dans quatre villes qui utilise des véhicules de petite capacité pour transporter des personnes sur demande de point à point entre de nombreuses stations.

39 - Au fur et à mesure que les véhicules autonomes se généraliseront, des questions surgiront sur leur sécurité, y compris sur la façon de s'assurer que les technologies sont sécurisées et testées correctement dans différentes conditions routières avant leur sortie. Les véhicules autonomes et l'infrastructure de transport connectée créeront un nouveau lieu pour les pirates informatiques afin d'exploiter les vulnérabilités aux attaques. Des questions d'éthique sont également impliquées dans la programmation de voitures pour agir dans des situations où les blessures humaines ou la mort sont inévitables, surtout quand il y a des choix partagés pour nous mettre en danger. 

Les systèmes juridiques dans la plupart des États aux États-Unis n'ont pas de règles concernant les voitures auto-conduisant. En 2016, quatre États aux États-Unis (Nevada, Floride, Californie et Michigan), l'Ontario au Canada, le Royaume-Uni, la France et la Suisse ont adopté des règles pour l'essai de voitures auto-conduisant sur la voie publique. Même ces lois ne traitent pas des questions concernant la responsabilité et l'attribution de sanctions pour un accident pour les auto-conduire et les autos semi-autonomes.

40 - Planification des transports :

En 2005, les villes ont commencé à investir dans l'infrastructure de transport pour développer les capacités de détection pour le trafic de véhicules et de piétons.

41 - Les capteurs actuellement utilisés comprennent les boucles inductives, les caméras vidéo, les capteurs hyperfréquences à distance, les radars et le GPS.

42 - Par exemple, en 2013, la ville de New-York a commencé à utiliser une combinaison de capteurs hyperfréquences, un réseau de caméras et des lecteurs passants pour détecter le trafic des véhicules.

43 37 - Brian McKenzie et Melanie Rapino, «Commutant aux États-Unis: 2009», American Community Survey Reports, Bureau du recensement des États-Unis, septembre 2011, a eu accès au 1er août 2016. 

https://www.census.gov/prod/2011pubs/acs-15.pdf

38 - Morgantown, Virginie-Occidentale; Masdar City, EAU; Londres, Angleterre; Et Suncheon, Corée du Sud.

39 - "Transitoire personnel rapide", Wikipedia, dernière modification le 18 juillet 2016, accessible le 1er août 2016. 

https://en.wikipedia.org/wiki/Personal_rapid_transit

40 - Patrick Lin, «L'éthique des voitures autonome», L'Atlantique, le 8 octobre 2013, a eu accès au 1er août 2016,

http://www.theatlantic.com/technology/archive/2013/10/the-ethics-ofautonomous-cars/280360/

41 - Steve Lohr, «Améliorer l'efficacité de l'infrastructure», The New York Times, le 29 avril 2009, a eu accès au 1er août 2016, 

42 - "Système de transport intelligent", Wikipedia, dernière modification le 28 juillet 2016, accessible le 1er août 2016.

https://en.wikipedia.org/wiki/Intelligent_transportation_system

43 - Access Science Editors, "Gestion du trafic actif: contrôle de signalisation adaptative", Access Science, 2014, consulté le 1er août 2016,

http://www.accessscience.com/content/active-traffic-managementadaptive-traffic-signal-control/BR0106141

Il est possible que cette conduite autonome est été disponible à partir de l'an 2000 - et en fait, des prototypes de voitures autonomes existaient - peu ont prédit que les entreprises traditionnelles développeront et déploieront des voitures autonomes d'ici 2015. Au cours du premier défi de la Defense Advanced Research Projects Agency (DARPA) sur la conduite autonome en 2004, les équipes de recherche n'ont pas réussi à compléter le défi dans un contexte désertique limité. Mais en huit années courtes, à partir des années  2004-2012, des progrès rapides et surprenants se sont produits tant dans le milieu universitaire que dans l'industrie. Les progrès de la technologie de détection et de l'apprentissage par machine pour les tâches de perception ont progressé et, par conséquent, les véhicules autonomes de Google et les voitures semi-autonomes de Tesla circulent aujourd'hui dans les rues de villes. Les autos automobiles de Google, qui ont enregistré plus de 1 500 000 kilomètres (300 000 kilomètres sans accident), sont totalement autonomes, pas d'intrants humains nécessaires. Tesla a largement libéré la capacité de conduire soi-même aux voitures existantes avec une mise à jour logicielle. Leurs voitures sont semi-autonomes, les conducteurs humains devraient rester engagés et prendre en charge s'ils détectent un problème potentiel.

Il n'est pas encore clair si cette approche semi-autonome est durable, car à mesure que les gens sont plus confiants dans les capacités des voitures, ils sont susceptibles d'accorder moins d'attention à la route et de devenir moins fiables lorsqu'ils sont les plus nécessaires. Le premier accident de la circulation impliquant une voiture autonome, qui s'est produit en juin 2016, a amené cette question à un accent plus net.

35 - Dans un proche avenir, les algorithmes de détection permettront d'obtenir des performances super humaines pour les capacités requises pour conduire. La perception automatisée, y compris la vision, est déjà proche ou à niveau humain pour des tâches bien définies telles que la reconnaissance et le suivi. Les progrès de la perception seront suivis d'améliorations algorithmiques dans des capacités de raisonnement de niveau supérieur telles que la planification. Un rapport récent prédit que les voitures auto-conduisant seront largement adoptées d'ici 2020.

36 - Et l'adoption de capacités de conduite autonome ne sera pas limitée au transport personnel. Nous verrons des véhicules de manutention auto-conducteurs et télécommandés, des véhicules volants et des camions. Les services de transport par les pairs, tels que le partage de vitesse, sont également susceptibles d'utiliser des véhicules autonomes.

Au-delà des voitures autonomes, les progrès de la robotique faciliteront la création et l'adoption d'autres types de véhicules autonomes, y compris les Robots et les Drones. Il n'est pas encore clair combien de meilleures voitures autonomes doivent devenir pour encourager une large acceptation. La collaboration requise dans les voitures semi-auto-conduisant et ses implications pour la charge cognitive des conducteurs humains n'est pas bien comprise. Mais si les voitures futures auto-conduisant sont adoptées avec la vitesse prévue elles risquent de dépasser les performances au niveau humain dans la conduite alors d'autres changements sociétaux importants devront suivre. Les voitures automotrices élimineront l'une des plus grandes causes de décès et de blessures accidentelles aux États-Unis et allongeront l'espérance de vie des gens. En moyenne, un "Navlab", Wikipedia, mis à jour le 4 juin 2016, a eu accès au 1er août 2016. 

31 - "Navlab: Laboratoire de navigation de l'Université Carnegie Mellon," Université Carnegie Mellon, consulté le 1er août 2016. 

http://www.cs.cmu.edu/afs/cs/project/alv/www/ 

32 - "Eureka Prometheus Project", Wikipedia, dernière modification le 12 février 2016, accessible le 1er août 2016.

https://en.wikipedia.org/wiki/Eureka_Prometheus_Project

33 - "Google Self-Driving Car Project," Google, consulté le 1er août 2016 

33 - Molly McHugh, "Les voitures de Tesla se conduisent maintenant, un peu," Wired, le 14 octobre 2015, ont eu accès au 1er août 2016

http://www.wired.com/2015/10/tesla-self-driving-overair-update-live/ 

34 - Molly McHugh, "Les voitures de Tesla se conduisent maintenant, un peu," Wired, le 14 octobre 2015, ont eu accès au 1er août 2016. h

http://www.wired.com/2015/10/tesla-self-driving-over-air-update-live/

35 - Anjali Singhvi et Karl Russell, "À l'intérieur de l'accident suicidaire de Tesla auto-conduisant", The New York Times, Dernière mise à jour le 12 juillet 2016, accessible le 1er août 2016.

36 - John Greenough, «10 millions de voitures auto-conduisant seront sur la route d'ici 2020», Business Insider, le 15 juin 2016, a eu accès au 1er août 2016.

http://www.wired.com/2015/10/tesla-self-driving-over-air-update-live/

Le transport partagé peut déplacer le besoin de transport public - ou les transports en commun peuvent changer de forme pour un transport en commun personnel qui utilise des véhicules de petite capacité pour transporter des personnes sur demande. Nous verrons des véhicules de manutention auto-conducteurs et télécommandés, des véhicules volants et des camions. Les services de transport par les pairs, tels que les visites guidées, sont également susceptibles d'utiliser des véhicules autonomes.

22 23 - L'IA est susceptible d'avoir un impact croissant sur l'infrastructure de la ville. Des modèles prédictifs précisés des mouvements des individus, de leurs préférences et de leurs objectifs sont susceptibles d'apparaître avec une plus grande disponibilité des données. Les questions éthiques concernant une telle émergence sont discutées dans la section III de ce rapport. Le Département des Transports des États-Unis a publié un appel à propositions en 2016 demandant aux villes moyennes d'imaginer l'infrastructure intelligente de la ville pour les transports. Cette initiative prévoit d'accorder quarante millions de dollars à une ville pour démontrer comment la technologie et les données peuvent être utilisées pour réimaginer les mouvements de personnes ainsi que les biens.

Une vision est un réseau de véhicules connectés qui peuvent atteindre un niveau élevé de sécurité dans la conduite avec une communication auto-voiture. Si cette vision devient réalité, nous prévoyons que les progrès de la coordination, de la collaboration et de la planification multi-agents auront une importance Impact sur les voitures futures et joueront un rôle dans la sécurité du système de transport. Les Robots sont également susceptibles de participer au transport en transportant des individus et des forfaits (c.f., robot Segway). Pour le transport de marchandises, l'intérêt pour les Drones a augmenté, et Amazon teste actuellement un système de livraison qui les utilise , bien qu'il subsiste des questions concernant les règles et les règles de sécurité appropriées. L'augmentation des capacités de détection, l'adoption de Drones et l'infrastructure de transport connexe susciteront également des préoccupations concernant la vie privée des individus et la sécurité des données privées. Au cours des prochaines années, ces problèmes de transports connexes devront être abordés soit par une action préventive de la part de l'industrie, soit dans le cadre juridique. Comme indiqué dans la discussion sur la politique de la section III, la façon dont cela est fait affectera le rythme et la portée des avances liées à l'IA dans le secteur des transports.

Transport à la demande :

Les services de transport à la demande tels que Uber et Lyft sont apparus comme une autre application essentielle de la détection, de la connectivité et de l'IA, avec des algorithmes d'adaptation des conducteurs aux passagers par localisation et adéquation (modélisation de la réputation). Grâce à la tarification dynamique , ces services rationnent l'accès par la volonté de payer, avec des prix dynamiques encourageant également une augmentation de l'offre de conducteurs qui sont devenus une méthode populaire pour le transport dans les villes. Avec leur avance rapide, de multiples problèmes politiques et juridiques, comme la concurrence avec les services de taxi existants et les préoccupations concernant le manque de réglementation et de sécurité. Les services de transport à la demande semblent être une force majeure pour les voitures auto-conduisant. Le co-voiturage et les déplacements ont longtemps été considérés comme une approche prometteuse pour réduire l'encombrement du trafic et utiliser les ressources de transport personnel. Des services tels que Zimride et Nuride rassemblent des personnes partageant des itinéraires similaires pour un voyage en commun. Mais cette approche du co-voiturage n'a pas réussi à prendre de l'ampleur.

55 - "US Le ministère des Transports lance un défi Smart City pour créer une ville de l'avenir, «Transport.gov, ministère des Transports des États-Unis, 7 décembre 2015, accessible le 1er août 2016.

https://www.transportation.gov/briefing-room/us-department-transportationlaunches-smart-city-challenge-create-city-future 

56 - Will Knight, "Communication voiture-voiture: une technologie sans fil simple promet de rendre la conduite beaucoup plus sûre", "MIT Technology Review, consulté le 1er août 2016,

57 -"Amazon Prime Air", Amazon, a eu accès au 1er août 2016,

http://www.amazon.com/b?node=8037720011  

58 - Jared Meyer, "Uber et Lyft changent la façon dont les Américains se déplacent sur leur pays", a examiné National Review, le 7 juin 2016, a eu accès au 1er août 2016,

59 - Alexander Howard, «Comment les plateformes numériques aiment LinkedIn, Uber et TaskRabbit changent l'économie à la demande», le Huffington Post, le 14 juillet 2015, a eu accès au 1er août 2016,

http://www.huffingtonpost.com/entry/online-talent-platforms_us_55a03545e4b0b8145f72ccf6 

60 - "Annonce d'UberPool", Uber Newsroom, 5 août 2014, accessible le 1er août 2016.

Les villes utilisent les méthodes IA pour optimiser les services de plusieurs façons, telles que les horaires de bus et de métro, et le suivi des conditions de circulation pour ajuster dynamiquement les limites de vitesse ou appliquer des prix intelligents sur les autoroutes, les ponts et les voies.

HOV.44 45 46 - Utiliser des capteurs et des caméras dans le réseau routier peuvent également optimiser la synchronisation des feux de signalisation pour améliorer les flux de trafic et pour aider à l'application automatisée. Ces stratégies dynamiques visent à mieux utiliser les ressources limitées dans le réseau de transport et sont rendues possibles par la disponibilité des données et la connectivité généralisée de personnes.

Avant les années 2000, les planificateurs de transport étaient forcés de s'appuyer sur des stratégies de tarification statiques liées à des jours ou des heures particuliers, pour gérer la demande. À mesure que des stratégies dynamiques de tarification sont adoptées, cela soulève de nouvelles questions concernant la répartition équitable des biens publics, car les conditions du marché dans des situations à forte demande peuvent rendre les services indisponibles pour des segments du public. La disponibilité de données à grande échelle a également fait des transports un domaine idéal pour les applications d'apprentissage par machine. Depuis 2006, des applications telles que Mapquest, Google Maps et Bing Maps ont été largement utilisées par le public pour les déplacements de routage, l'utilisation des transports publics, la réception d'informations en temps réel et les prévisions sur les conditions de trafic, la recherche de services autour d'un emplacement.

51 52 - Des algorithmes de recherche optimale ont été appliqués au routage des véhicules et des piétons vers une destination donnée (c.-à-d. 53 54). Malgré ces progrès, l'application généralisée des techniques de détection et d'optimisation à l'infrastructure de la ville a été plus lente que l'application de ces techniques à des véhicules ou à des personnes individuelles. Bien que les villes individuelles aient implémenté des applications de détection et d'optimisation, il n'y a pas encore de standardisation de l'infrastructure de détection et des techniques d'IA utilisées. Les coûts d'infrastructure, les priorités différentes des villes et les coûts élevés de coordination entre les parties concernées ont ralenti l'adoption, tout comme les préoccupations du public concernant la vie privée liées à la détection. Pourtant, Kitae Jang, Koohong Chung et Hwasoo Yeo, «Une stratégie de tarification dynamique pour les voies d'accès à haute occupation», ont fait l'étude d'une recherche sur les transports en partie A:

"Politique et Pratique" :

67 - en (2014): 69-80. 45 "Seattle Variable Tolling Study", Département de Transports de la Ville de Seattle, mai 2009, consulté le 1er août 2016. 

46 - James F. Peltz, «Les prix dynamiques s'accrochent dans les secteurs public et privé», technologie gouvernementale, le 21 mars 2016, a eu accès au 1er août 2016.

http://www.govtech.com/budgetfinance/Dynamic-Pricing-Is-Catching-On-in-the-Public-and-Private-Sectors.html

47 - Arthur G Sims et Kenneth W. Dobinson. «La philosophie et les avantages du système de circulation adaptée de Sydney (« SCAT) », Transactions IEEE sur la technologie des véhicules 29, no. 2 (1980): 130-137.

48 - "New-York City" lance le système de gestion du trafic actif le plus sophistiqué de la Nation Powered by TransCore's TransSuite Traffic Management Software and RFID Technology", Business Wire, le 27 septembre 2009, a eu accès au 1er août 2016.

49 - Eric Horvitz, Johnson Apacible, Raman Sarin et Lin Liao, "Prévision, Attente et Surprise: Méthodes, Dessins et Étude d'un Service Déployé de Prévision du Trafic", Actes de la Vingt-Première Conférence sur l'Incertitude et l'Intelligence Artificielle (2005 ) (Arlington, Virginie: AUAI Press, juillet 2005).

275-283. 50 - Timothy Hunter, Ryan Herring, Pieter Abbeel et Alexandre Bayen, «Inférence du chemin et du temps de déplacement à partir des données du véhicule GPS Probe», NIPS Analyse des réseaux et apprentissage avec les graphiques 12, no. 1 (2009).

51 - John Krumm et Eric Horvitz, «Predestination: Infermer les destinations des trajectoires partielles», UbiComp 2006: Informatique omniprésente, Actes de la 8e Conférence internationale, septembre 2006 (Springer Berlin, Heidelberg, 2006).

243-260. 52 - Jill Duffy, "Organisez-vous: en utilisant les rappels basés sur l'emplacement", PC Magazine, le 30 juin 2014, a eu accès au 1er août 2016. 

http://www.pcmag.com/article2/0,2817,2460207,00.asp

 

53 - Robert J. Szczerba, Peggy Galkowski, I. Glicktein et Noah Ternullo. «Algorithme robuste pour la planification des itinéraires en temps réel», Transactions IEEE sur les systèmes aérospatiaux et électroniques 36, no. 3 ( année 2000):

869-878. 54 - Matt Duckham et Lars Kulik, «Simplest» Paths: Automated Route Selection for Navigation, «Spatial Information Theory». Fondements de la science de l'information géographique, Actes de la Conférence internationale, COSIT 2003, septembre 2003 (Springer-Verlag Berlin Heidelberg, 2003).

169-185. - Notre panel d'étude ne s'attend pas à ce que les drones puissent voler, nager et conduire, ou piloter des quadricorteurs pour devenir un moyen commun de transport d'ici 2030 (bien que les prototypes existent aujourd'hui). Des questions d'ordre éthique apparaissent lors de la programmation de voitures pour agir dans des situations où les blessures humaines ou la mort sont inévitables, en particulier lorsqu'il existe des choix partagés pour déterminer s'il y a lieu de mettre en danger et pour des fonctions telles que l'étiquetage d'image et la reconnaissance d'objets 3D, tout en étant fréquent dans les conférences d'IA, ne sont encore que quelques années de développement en tant que produits.

Robots domestiques en 2030 :

Malgré la lenteur de la croissance des Robots à la maison, il y a des signes que cela changera au cours des quinze prochaines années. Des sociétés telles que Amazon Robotics et Uber développent de grandes économies d'échelle en utilisant diverses technologies d'agrégation. Aussi: System in Module (SiM), avec beaucoup de sous-systèmes System On Chip (SoC), est maintenant mis en place par les fabricants de puces téléphoniques (SnapDragon de Qualcomm, Artik de Samsung, etc.). Ce sont mieux que les supercalculateurs de moins de dix ans avec huit cœurs de plus de six centièmes, et des silicones spécialisés pour la cryptographie, les pilotes de caméra, les DSP supplémentaires et le silicium dur pour certains algorithmes perceptifs. Cela signifie que les appareils à faible coût pourront supporter beaucoup plus d'IA à bord que ce que nous avons pu considérer au cours des quinze dernières années. Cloud («l'ordinateur de quelqu'un d'autre») permettra une libération plus rapide de nouveaux logiciels sur les Robots domestiques et plus de partage de jeux de données rassemblés dans de nombreuses maisons différentes, ce qui alimentera l'apprentissage par machine en nuage, puis des améliorations de puissance déjà déployé des Robots.

Les grands progrès dans la compréhension de la parole et l'étiquetage des images grâce à un apprentissage approfondi amélioreront les interactions des Robots avec les gens dans leurs maisons. Les capteurs 3D à faible coût, pilotés par des plates-formes de jeux, ont alimenté le travail sur les algorithmes de perception 3D de milliers de chercheurs dans le monde entier, ce qui accélérera le développement et l'adoption de Robots domestiques et de services. Au cours des trois dernières années, des armes à Robot à faible coût et sûres ont été introduites dans des centaines de laboratoires de recherche dans le monde entier, suscitant une nouvelle classe de recherche sur les manipulations qui seront éventuellement applicables à la maison, peut-être vers 2025. Plus d'une demi-douzaine de startups autour du monde développent des rRbots basés sur l'IA pour la maison, pour l'instant se concentrant principalement sur les interactions sociales. De nouveaux problèmes d'éthique et de confidentialité peuvent en résulter.

HEALTHCARE :

Pour les technologies de l'IA, les soins de santé ont longtemps été considérés comme un domaine prometteur. Les applications basées sur l'IA pourraient améliorer les résultats sanitaires et la qualité de vie pour des millions de personnes dans les années à venir, mais seulement si elles gagnent la confiance des médecins, des infirmières et des patients et si les obstacles politiques, réglementaires et commerciaux sont supprimés. Les applications principales comprennent le soutien à la décision clinique, la surveillance et le coaching des patients, les dispositifs automatisés pour aider à la chirurgie ou les soins aux patients et la gestion des systèmes de santé. Les réussites récentes, telles que les médias sociaux miniers pour inférer les risques possibles pour la santé, l'apprentissage automatique de la machine pour prédire les patients à risque et la Robotique pour soutenir la chirurgie, ont élargi le sens de la possibilité d'IA dans les soins de santé. L'amélioration des méthodes d'interaction avec les professionnels de la santé et les patients sera un défi majeur. Comme dans d'autres domaines, les données sont un facteur clé.

Il y a eu un énorme bond en avant dans la collecte de données utiles à partir de dispositifs de surveillance personnelle et d'applications mobiles, à partir de dossiers médicaux électroniques (DSE) en milieu clinique et, dans une moindre mesure, de Robots conçus pour faciliter les procédures médicales et les opérations hospitalières. Mais l'utilisation de ces données pour permettre des diagnostics et des traitements plus finis pour les patients et les populations de patients a été difficile. La recherche et le déploiement ont été ralenties par des réglementations périmées et des structures incitatives. Les mauvaises méthodes d'interaction homme-ordinateur et les difficultés et les risques inhérents à la mise en œuvre des technologies dans un système aussi vaste et complexe ont ralenti la réalisation de l'interaction de l'IA avec les gens. Depuis des décennies, les gens ont imaginé des véhicules de transport à la fine pointe de la technologie. Bien que les futures voitures soient plus intelligentes et que les Drones soient largement disponibles, il est peu probable qu'en 2030, nous aurions largement adopté des véhicules de transport qui fonctionnent différemment de ceux que nous avons aujourd'hui. Notre panel d'étude ne s'attend pas à ce que les Drones puissent voler, nager et conduire, ou piloter des quadricorteurs pour devenir un moyen commun de transport dans cet horizon (bien que les prototypes existent aujourd'hui). Nous nous attendons à ce que les humains deviennent des partenaires pour conduire des voitures et des Drones dans leur formation, leur exécution et leur évaluation. Ce partenariat se produira à la fois lorsque les humains ce seront familiarisées patiquement avec des machines. Nous prévoyons des progrès dans les algorithmes pour faciliter l'apprentissage par machine à partir d'intrants humains. Nous nous attendons également à des modèles d'algorithmes pour la modélisation de l'attention humaine, à la communication et à la coordination entre les humains et les machines. Cela fait partie intégrante du développement de véhicules futurs.

- ROBOTS D'ACCUEIL / SERVICE :

Les robots sont entrés dans les maisons des personnes au cours des quinze dernières années. Une croissance décevante lente dans la diversité des applications s'est produite simultanément avec une IA de plus en plus sophistiquée déployée sur les applications existantes. Les progrès de l'IA sont souvent inspirés par des innovations mécaniques qui, à leur tour, provoquent l'introduction de nouvelles techniques d'IA. Au cours des quinze prochaines années, les progrès coïncidents dans les technologies mécaniques et de l'IA promettent d'accroître l'utilisation sûre et fiable et l'utilité des robots domestiques dans une ville nord-américaine typique. Les Robots spéciaux proposent des paquets, nettoient des bureaux et améliorent la sécurité, mais les contraintes techniques et les coûts élevés des dispositifs mécaniques fiables continueront à limiter les opportunités commerciales à des applications étroitement définies dans un avenir prévisible. Comme pour les voitures auto-conduisant et les autres machines de transport, la difficulté de créer un matériel fiable et prêt pour le marché ne doit pas être sous-estimée. Aspirateurs en 2001, après de nombreuses années de développement, le Electrolux Trilobite, un robot de nettoyage à vide est devenu le premier robot domestique. Il avait un système de contrôle simple pour éviter les obstacles, et une certaine navigation. Un an plus tard, iRobot a introduit Roomba, qui était un dixième du prix du Trilobite et, avec seulement 512 octets de RAM, utilisait un contrôleur basé sur le comportement. La chose la plus intelligente était d'éviter de tomber dans les escaliers. Depuis, seize millions de Roombas ont été déployés partout dans le monde et plusieurs autres marques concurrentes existent maintenant. À mesure que la puissance de traitement et la capacité de RAM des processeurs intégrés à faible coût se sont améliorés par rapport à leur lâche en l'an 2000, les capacités d'IA de ces robots se sont également considérablement améliorées. La navigation simple, l'auto-charge et les actions pour traiter des bacs à poussière complets ont été ajoutés, suivis de la capacité de traiter les cordons électriques et les glands de tapis, grâce à une combinaison d'améliorations mécaniques et de perception par capteur. 

Plus récemment, l'ajout de VSLAM complet (Visual Simultaneous Location and Mapping) - une technologie d'IA qui existe depuis vingt ans - a permis aux robots de construire un modèle mondial 3D complet d'une maison à mesure qu'ils se nettoient et deviennent plus efficaces dans leur couverture de nettoyage. Les attentes anticipées que de nombreuses nouvelles applications seraient trouvées pour les robots domestiques ne se sont pas concrétisées. Les aspirateurs à robot sont limités aux zones plates localisées, tandis que les maisons réelles ont beaucoup d'étapes simples et souvent des escaliers. Il y a eu très peu de recherches sur la mobilité des robots dans les maisons réelles. Les plates-formes matérielles restent difficiles à créer, et il y a peu d'applications que les gens souhaitent acheter.

Algorithmes perceptifs :

Les robots spéciaux proposent des paquets, nettoient des bureaux et améliorent la sécurité, mais les contraintes techniques et les coûts élevés continueront à limiter les opportunités commerciales dans un avenir prévisible. Au cours des quinze prochaines années, les progrès coïncidents dans les technologies mécaniques et de l'IA promettent d'accroître l'utilisation sûre et fiable et l'utilité des robots domestiques dans une ville nord-américaine typique. En tant que procédures médicales et des enregistrements cliniques à vie pour des centaines de millions d'individus deviennent disponibles. De même, la capture automatisée des données environnementales personnelles à partir d'appareils portables augmentera la médecine personnalisée.

Ces activités deviennent plus commercialement viables à mesure que les fournisseurs découvrent des moyens d'engager de grandes populations (par exemple, ShareCare) et ensuite créer des données à l'échelle de la population qui peuvent être exploitées pour produire des analyses et des recommandations individualisées. Malheureusement, la FDA a tardé à approuver des logiciels de diagnostic innovants et il existe de nombreux obstacles à l'innovation rapide. Les exigences de HIPAA (Loi sur la transférabilité et la responsabilité de l'assurance maladie) pour protéger la vie privée des patients créent des obstacles juridiques au flux de données sur les patients à des applications qui pourraient utiliser les technologies de l'IA. Les effets négatifs imprévus des médicaments homologués pourraient apparaître de façon routinière, plus tôt et plus rigoureusement qu'aujourd'hui, mais les applications mobiles qui analysent les interactions médicamenteuses peuvent être empêchées de tirer les informations nécessaires des dossiers des patients. D'une manière plus générale, la recherche et l'innovation dans les soins de santé sont entravés par le manque de méthodes et de normes largement acceptées pour la protection de la vie privée. La FDA a tardé à approuver des logiciels innovants, en partie à cause d'une compréhension claire des compromis coût / bénéfice de ces systèmes. Si les régulateurs (principalement la FDA) reconnaissent que les rapports postérieurs à la commercialisation efficaces sont une couverture sûre contre certains risques de sécurité, une approbation initiale plus rapide des nouveaux traitements et des interventions peut devenir possible. L'interprétation automatisée de l'image a également été un sujet d'étude prometteur depuis des décennies. Les progrès réalisés dans l'interprétation de grandes archives d'images faiblement étiquetées, telles que les grandes archives de photos grattées sur le Web, ont été explosives. Au premier rougissement, il est surprenant qu'il n'y ait pas eu de révolution similaire dans l'interprétation des images médicales. La plupart des modalités d'imagerie médicale (CT, MR, échographie) sont intrinsèquement numériques, les images sont toutes archivées, et il existe de grandes entreprises établies avec une R & D interne (par exemple, Siemens, Philips, GE) consacrée à l'imagerie. Mais plusieurs barrières ont des progrès limités à ce jour. La plupart des archives d'images d'hôpitaux sont devenues numériques au cours de la dernière décennie. Plus important encore, le problème de la médecine n'est pas de reconnaître ce qui se trouve dans l'image (est-ce qu'il y a un foie ou un rein?), Mais plutôt de faire un jugement très fin à ce sujet (la tache légèrement plus sombre dans le foie suggère-t-elle un potentiel de Tumeur cancéreuse ?).

Des règles strictes régissent ces jugements à haut risque. Même avec des technologies de pointe, un radiologue devra probablement regarder les images, de sorte que la proposition de valeur n'est pas encore convaincante. En outre, la réglementation sanitaire exclut une fédération facile de données dans les établissements. Ainsi, seules de très grandes organisations de soins intégrés, tels que Kaiser Permanente, sont en mesure d'attaquer ces problèmes. Pourtant, l'interprétation automatisée / augmentée de l'image a commencé à prendre de l'ampleur. Les prochaines quinze années n'apporteront probablement pas de radiologie entièrement automatisée, mais les incursions initiales dans le contrôle du "triage" ou du second niveau vont probablement améliorer la rapidité et la rentabilité de l'imagerie médicale. Lorsqu'il est couplé aux systèmes électroniques d'enregistrement des patients, des techniques d'apprentissage par machine à grande échelle pourraient être appliquées aux données d'images médicales. Par exemple, plusieurs systèmes de soins de santé majeurs ont des archives de millions d'analyses de patients, dont chacun a un rapport radiologique associé, et la plupart ont un dossier de patient associé. Déjà, des documents apparaissent dans la littérature montrant que les réseaux de neurones profonds peuvent être formés pour produire des résultats radiologiques de base, avec une grande fiabilité, grâce à une formation à partir de ces données.

64 63 - Sharecare, consulté le 1er août 2016.

64 - Hoo-Chang Shin, Holger R. Roth, Mingchen Gao, Le Lu, Ziyue Xu, Isabella Nogues, Jianhua Yao, Daniel Mollura et Ronald M. Summers, "Deep Convolutional Neural Networks for Computer-aided Detection: CNN Architectures, Dataset Caractéristiques et apprentissage par transfert, «Transactions IEEE sur l'imagerie médicale 35, no. 5 (2016):

1285-1298. - Promis dans les soins de santé, la réduction ou l'élimination de ces obstacles, conjuguée aux innovations encore à l'horizon, permettront d'améliorer considérablement les résultats sanitaires et la qualité de vie de millions de personnes dans les années à venir.

Le cadre clinique :

Depuis des décennies, la vision d'un assistant clinicien de l'IA a été un cliché proche. Bien qu'il y ait eu des pilotes réussis de la technologie liée à l'IA dans les soins de santé , le système de prestation de soins de santé actuel reste mal structuré pour absorber et déployer des avancées rapides. Les incitations fournies par la Loi sur les soins abordables ont accéléré la pénétration des dossiers de santé électroniques (DSE) dans la pratique clinique, mais la mise en œuvre a été médiocre, ce qui a entravé la confiance des cliniciens dans leur utilité. Un petit groupe de sociétés contrôle le marché du DSE, et les interfaces utilisateur sont largement considérées comme de mauvaise qualité, y compris les "pop-ups" ennuyeuses que les médecins rejettent systématiquement. La promesse de nouvelles analyses utilisant les données des EHR, y compris l'IA, reste largement non réalisée en raison de ces obstacles et d'autres obstacles réglementaires et structurels. Dans l'attente des prochaines quinze années, les avancées d'IA, si elles sont associées à des données suffisantes et à des systèmes bien ciblés, promettent de changer les tâches cognitives assignées aux cliniciens humains. Les médecins sollicitent systématiquement des descriptions verbales des symptômes de la présentation des patients et, dans leur tête, mettent en corrélation les modèles par rapport à la présentation clinique des maladies connues. Avec une assistance automatisée, le médecin pourrait plutôt superviser ce processus, en appliquant son expérience et son intuition pour guider le processus d'intrants et évaluer la sortie de l'intelligence de la machine.

 

L'expérience littéraire "pratique" du médecin restera critique. Un défi majeur consiste à intégrer de manière optimale les dimensions humaines des soins aux processus automatisés de raisonnement. Pour atteindre les progrès futurs, les cliniciens doivent être impliqués et engagés dès le départ pour s'assurer que les systèmes sont bien conçus et fiables. Déjà, une nouvelle génération de médecins plus compétents en technologie utilise systématiquement des applications spécialisées sur les appareils mobiles. Dans le même temps, les charges de travail sur les cliniciens de soins primaires ont augmenté au point qu'ils sont reconnaissants de l'aide de n'importe quel trimestre. Ainsi, l'opportunité d'exploiter de nouvelles méthodes d'apprentissage, de créer des schémas structurés d'inférence en exploitant automatiquement la littérature scientifique et de créer de véritables assistants cognitifs en soutenant le dialogue sans forme, n'a jamais été aussi grande. À condition que ces avancées ne soient pas entravées par des obstacles réglementaires, juridiques et sociaux, des améliorations immenses de la valeur des soins de santé sont à notre portée.

Analyse des soins de santé :

Au niveau de la population, la capacité d'IA à réduire les résultats de millions d'enregistrements cliniques de patients promet un diagnostic et un traitement plus précis et plus personnalisés. La découverte automatisée des connexions génotype-phénotype sera également possible car le séquençage complet du génome une fois dans la vie devient routinier pour chaque patient. Une capacité connexe (et peut-être plus tôt) sera de trouver «les patients comme le mien» comme moyen d'informer les décisions de traitement en fonction de l'analyse d'une cohorte similaire.

Les données traditionnelles et non traditionnelles sur la santé, complétées par des plates-formes sociales, peuvent conduire à l'apparition de sous-populations autodéfinies, chacune gérée par un écosystème environnant de prestataires de soins de santé augmentée avec des recommandations automatisées et des systèmes de surveillance. Ces développements ont le potentiel de transformer radicalement les soins de santé.

61 - LeighAnne Olsen, Dara Aisner et J. Michael McGinnis, éditeurs, "Table ronde sur la médecine basée sur la preuve" de l'Institut de médecine (États-Unis), The Learning Healthcare System: Workshop Summary. (Washington (DC): National Academies Press (US); 2007), accessible le 1er août 2016.

62 - Katherine E. Henry, David N. Hager, Peter J. Pronovost et Suchi Saria, "Un score d'alerte précoce en temps réel ciblé (TREWScore) pour le choc septique", Science Translational Medicine 7. Un petit groupe de sociétés contrôle le marché du DSE, et les interfaces utilisateur sont largement considérées comme de mauvaise qualité, y compris les pop-ups ennuyeuses que les médecins rejettent systématiquement. Les applications basées sur l'IA pourraient améliorer les résultats sanitaires et la qualité de vie pour des millions de personnes dans les années à venir, mais seulement si elles gagnent la confiance des médecins, des infirmières et des patients et soit relativement simple une fois qu'un patient est debout sur un "marcheur" (mais ne sera certainement pas trivial pour les patients qui se remettent de la chirurgie et / ou des patients âgés, en particulier dans les couloirs remplis d'équipement et d'autres personnes).

Conduire une aiguille pour placer une suture est relativement simple une fois que l'aiguille est correctement placée. Cela implique que de nombreux systèmes futurs impliqueront une interaction intime entre les personnes et les machines et nécessiteront des technologies qui facilitent la collaboration entre elles. La croissance de l'automatisation permettra de mieux comprendre les processus de soins de santé. Historiquement, la Robotique n'a pas été une science fortement axée sur les données ou axée sur les données. Cela change à mesure que l'automatisation (semi) s'infiltre dans les soins de santé. Au fur et à mesure que les nouvelles plateformes de chirurgie, de livraison et de soins aux patients sont en ligne, les débuts de la quantification et de l'analyse prédictive sont intégrés aux données provenant de ces plates-formes. Ces données seront utilisées pour évaluer la qualité de la performance, identifier les lacunes, les erreurs ou bien des optimisations potentielles et seront utilisés comme rétroaction pour améliorer les performances. Bref, ces plates-formes faciliteront la connexion entre ce qui est fait et les résultats obtenus, ce qui rend le véritable médicament en boucle fermée d'une possibilité réelle. 

Santé mobile :

À ce jour, les analyses fondées sur des données probantes sur les soins de santé se sont appuyées sur les données traditionnelles sur les soins de santé, principalement les dossiers médicaux électroniques discutés ci-dessus. Dans le contexte clinique, il existe des tendances encourageantes à apporter de nouvelles données. Par exemple, Tele Language permet à un clinicien humain de mener des séances de thérapie linguistique avec plusieurs patients simultanément à l'aide d'un agent d'IA formé par le clinicien. Et Lifegraph, qui extrait des modèles comportementaux et crée des alertes à partir de données recueillies passivement à partir d'un téléphone intelligent d'un patient, a été adopté par des psychiatres en Israël pour détecter les signes précurseurs d'un comportement pénible chez les patients. À l'avenir, grâce à la révolution informatique mobile, la croissance étonnante de la «biométrie à l'état sauvage» - et l'explosion des plates-formes et des applications qui les utilisent - est une tendance anticipée et imprévue. Des milliers d'applications mobiles offrent maintenant des informations, introduisent une modification de comportement ou identifient des groupes de «personnes comme moi». Ceci, combiné à la tendance émergente de dispositifs de suivi de mouvement plus spécialisés tels que Fitbit et l'apparition (inter) entre à la maison Des dispositifs de surveillance de l'environnement et de la santé, ont créé un nouveau secteur d'innovation dynamique. En combinant les données sociales et de soins de santé, certaines applications de soins de santé peuvent effectuer l'exploration de données, l'apprentissage et la prédiction à partir des données capturées, bien que leurs prédictions soient relativement rudimentaires.

La convergence des données et des fonctionnalités dans toutes les applications suscitera probablement des produits nouveaux et même évidents, comme une application d'exercice qui propose non seulement un horaire d'exercice, mais suggère également le meilleur moment pour le faire et fournit un encadrement pour respecter ce calendrier.

72 - Azad Shademan, Ryan S. Decker, Justin D. Opfermann, Simon Leonard, Axel Krieger et Peter CW Kim, «Chirurgie autonome autonome de tissus molles robotisés», Science Translational Medicine 8, no. 337 (2016):

73 - Carolyn Chen, Lee White, Timothy Kowalewski, Rajesh Aggarwal, Chris Lintott, Bryan Comstock, Katie Kuksenok, Cecilia Aragon, Daniel Holst et Thomas Lendvay, «L'évaluation des compétences techniques de la foule: une nouvelle méthode pour évaluer la performance chirurgicale». Journal of Surgical Research 187, no. 1 (2014):

65-71. - Robotique de la santé :

Il y a quinze ans, la robotique de la santé était en grande partie la science-fiction. Une société appelée Robodoc, une spin-out d'IBM, a développé des systèmes Robotiques pour les chirurgies orthopédiques, comme les remplacements de la hanche et du genou. La technologie a fonctionné, mais la société a lutté commercialement et a finalement été fermée et acquise pour sa technologie.

66 - Plus récemment, cependant, la recherche et l'utilisation pratique de la Robotique chirurgicale ont explosé. En 2000, Intuitive Surgical a introduit le système Da Vinci, une nouvelle technologie initialement commercialisée pour soutenir la Chirurgie de pontage cardiaque minimalement invasive, puis a gagné une fraction importante sur le marché pour le traitement du Cancer de la prostate et a fusionné avec sa seule concurrence majeure, Computer Motion, en 2003.

Le Da Vinci, maintenant dans sa quatrième génération, offre une visualisation en 3D (par opposition à la laparoscopie monoculaire 2D) et des instruments à poignet dans une plate-forme ergonomique. Il est considéré comme la norme de soins dans de multiples procédures laparoscopiques et utilisé dans près des trois quarts d'un million de procédures par année, fournissant non seulement une plate-forme physique, mais aussi une nouvelle plate-forme de données pour l'étude du processus de chirurgie. Le Da Vinci prévoit un jour où une vision plus approfondie de la façon dont les professionnels de la santé mènent le processus de fourniture de soins médicaux interventionnels sera possible. La présence de l'opération quotidienne de Da Vinci a également ouvert les portes à de nouveaux types d'innovation, de la nouvelle instrumentation à la fusion d'images à de nouveaux biomarqueurs, créant ainsi son propre écosystème d'innovation. Le succès de la plate-forme a inspiré des concurrents potentiels en chirurgie Robotique, notamment le verbe spin-off Alphabet, en collaboration avec J & J / Ethicon.

69 - Il est probable qu'il y en ait beaucoup plus, chacun explorant un créneau ou un espace unique et construisant un écosystème de la détection, de l'analyse des données, de l'augmentation et de l'automatisation. L'automatisation intelligente dans les opérations hospitalières a été moins réussie. L'histoire n'est pas différente de la Robotique chirurgicale. Il y a vingt ans, une entreprise, Help Mate, a créé un Robot pour les accouchements hospitaliers, ainsi que pour les repas et les dossiers médicaux, mais a finalement fait faillite. Plus récemment, Aethon a introduit TUG Robots pour les livraisons de base, mais peu d'hôpitaux ont investi dans cette technologie à ce jour. Cependant, la Robotique dans d'autres industries de services telles que les hôtels et les entrepôts, y compris Amazon Robotics (anciennement Kiva), démontre que ces technologies sont pratiques et rentables dans au moins certains paramètres à grande échelle et peuvent finalement susciter une innovation supplémentaire dans les soins de santé. Dans l'avenir, de nombreuses tâches qui apparaissent dans les soins de santé seront susceptibles d'augmenter, mais ne seront pas entièrement automatisées. 

Par exemple, les Robots peuvent être en mesure de livrer des marchandises dans la bonne pièce d'un hôpital, mais exigent une personne pour les retirer et les placer dans leur emplacement final. Marcher pour un patient dans le couloir peut-être fait avec ROBODOC, consulté le 1er août 2016.

http://www.robodoc.com/professionals.html

66 - PENSEUR chirurgical, consulté le 1er août 2016, 

http://thinksurgical.com/history  

67 - Intuitive Surgical, consulté le 1er août 2016,

http://www.intuitivesurgical.com

68 - Trefis Team, "La chirurgie intuitive maintient son élan de croissance avec une forte croissance dans les volumes de procédure", Forbes, le 22 janvier 2016, a eu accès au 1er août 2016, 

69 - Evan Ackerman, "Google et Johnson & Johnson conjuguent pour créer des robots médicaux chirurgicaux, promettent des verbes," IEEE Spectrum, le 17 décembre 2015, a eu accès au 1er août 2016,

http://spectrum.ieee.org/automaton/robotics/medical-robots/google-verily-johnson-johnsonverb-surgical-medical-robots

70 - John M. Evans et Bala Krishnamurthy, "HelpMate®, le service de messagerie robotique sans piste: une perspective sur le développement d'un robot mobile autonome commercial", Notes de cours en matière de contrôle et de sciences de l'information 236, 18 juin 2005 (Springer-Verlag London Limited , 1998), 182-210, consulté le 1er août 2016.

http://link.springer.com/chapter/10.1007%2FBFb0030806

71 - Aethon, consulté le 1er août 2016.

http://www.aethon.com

Les dispositifs spécialisés de suivi des mouvements ... et la relation émergente (inter) entre l'environnement domestique et les dispositifs de surveillance de la santé ont créé un nouveau secteur dynamique de l'innovation. Le problème de la médecine n'est pas de reconnaître ce qui se trouve dans l'image: est-ce qu'il y a un foie ou un rein? Mais plutôt de faire un jugement à fond. Des règles strictes régissent ces jugements à haut risque.

30 31 - ÉDUCATION :

Au cours des quinze dernières années, les avancées de l'IA sont considérables dans l'éducation. Les applications sont largement utilisées par les éducateurs et les apprenants aujourd'hui, avec une certaine variation entre K-12 et les paramètres de l'université. Bien que l'éducation de qualité exige toujours un engagement actif des enseignants humains, l'IA promet d'améliorer l'éducation à tous les niveaux, en particulier en fournissant une personnalisation à l'échelle. À l'instar des soins de santé, la résolution de la meilleure intégration de l'interaction humaine et de l'apprentissage en face-à-face avec les technologies prometteuses de l'IA reste un défi majeur. Les robots ont longtemps été des dispositifs éducatifs populaires, en commençant par les premiers kits Lego Mindstorms développés avec le MIT Media Lab dans les années 1980. Les systèmes de tutorat intelligents (STI) pour les sciences, les mathématiques, la langue et d'autres disciplines correspondent aux étudiants avec des tuturistes interactifs.

Le traitement du langage naturel, en particulier lorsqu'il a été combiné avec l'apprentissage par machine et le crowdsourcing, a stimulé l'apprentissage en ligne et a permis aux enseignants de multiplier la taille de leurs salles de classe tout en abordant simultanément les besoins et les styles d'apprentissage des élèves. Les ensembles de données provenant de grands systèmes d'apprentissage en ligne ont alimenté une croissance rapide de l'analyse d'apprentissage. Pourtant, les écoles et les Universités ont été lentes dans l'adoption de technologies de l'IA principalement en raison du manque de fonds et du manque de preuves solides qu'elles aident les élèves à atteindre les objectifs d'apprentissage. Au cours des quinze prochaines années, dans une ville typiquement nord-américaine, l'utilisation de tuteurs intelligents et d'autres technologies de l'IA pour aider les enseignants dans la salle de classe et dans la maison est susceptible d'augmenter considérablement, tout comme l'apprentissage basé sur des applications de réalité virtuelle. Mais les systèmes informatiques d'apprentissage ne sont pas susceptibles de remplacer pleinement l'enseignement humain dans les écoles.

Enseigner des Robots :

Aujourd'hui, des kits plus sophistiqués et polyvalents pour les écoles K-12 sont disponibles auprès d'un certain nombre d'entreprises qui créent des Robots avec de nouvelles technologies de détection programmables dans une variété de langues. 

Ozobot est un Robot qui enseigne aux enfants à coder et à déduire déductivement tout en le configurant pour danser ou jouer en fonction de modèles codés par couleur. Les Cubelets aident à enseigner aux enfants la pensée logique en assemblant des blocs de robot pour penser, agir ou détecter, selon la fonction des différents blocs.

76 - Le Dash and Dot de Wonder Workshop s'étend sur une gamme de capacités de programmation. Les enfants de huit ans et plus peuvent créer des actions simples en utilisant un langage de programmation visuel, bloquer ou créer des applications iOS et Android en utilisant C ou Java.

PLEO rb est un robot animal qui aide les enfants à apprendre la biologie en enseignant au robot de réagir à différents aspects de l'environnement. Cependant, si amusant et engageant pour certains, pour que de tels kits deviennent répandus, il faudra démontrer de manière convaincante qu'ils améliorent le rendement scolaire des élèves.

Intelligent Tutoring Systems (ITS) et l'apprentissage en ligne ITS ont été développés à partir de projets de laboratoire de recherche tels que Why-2 Atlas, qui a soutenu le dialogue homme-machine pour résoudre les problèmes de physique au début de l'an 2000. La migration rapide des STI des étapes expérimentales de laboratoire à l'utilisation réelle est de Ozobot, accessible le 1er août 2016.

http://ozobot.com/

76 - "Cubelets", Modo Robotics, a eu accès au 1er août 2016.

http://www.modrobotics.com/cubelets

77 - "Meet Dash," Wonder Workshop, consulté le 1er août 2016. 

78 - "Pleo rb", Innvo Labs, consulté le 1er août 2016.

79 - Kurt VanLehn, Pamela W. Jordan, Carolyn P. Rosé, Dumisizwe Bhembe, Michael Böttner, Andy Gaydos, Maxim Makatchev, Umarani Pappuswamy, Michael Ringenberg, Antonio Roque, Stephanie Siler et Ramesh Srivastava, "L'architecture de Why2-Atlas : ont créé un entraîneur pour les soins aux personnes âgées qualitatives.

Au cours des quinze prochaines années, le nombre de personnes âgées aux États-Unis augmentera de plus de 50%. Le Bureau national des statistiques du travail prévoit que les agents de santé à domicile augmenteront de 38% au cours des dix prochaines années. Malgré les grandes opportunités dans ce domaine: les services de soutien social, d'interaction et de communication, la surveillance de la santé à domicile, une variété d'aides physiques simples à la maison, comme les marcheurs et la préparation légère des repas, il s'est produit peu au cours des quinze dernières années. Mais le changement de génération à venir accompagnera un changement dans l'acceptation de la technologie chez les personnes âgées. À l'heure actuelle, quelqu'un qui a soixante-dix ans est né en 1946 et a peut-être expérimenté une certaine forme d'informatique personnalisée à l'âge mûr ou plus tard, alors qu'un jeune de cinquante ans aujourd'hui est beaucoup plus perméable à l'informatique. En conséquence, il y aura un intérêt croissant et un marché pour les technologies déjà disponibles et en pleine maturité pour soutenir la santé physique, émotionnelle, sociale et mentale.

Voici quelques exemples probables par catégorie: Qualité de vie et indépendance • Le transport automatisé appuiera l'indépendance continue et l'élargissement des horizons sociaux. • Le partage de l'information aidera les familles à rester à distance, et les analyses prédictives peuvent être utilisées pour «pousser» les groupes familiaux vers des comportements positifs, tels que des rappels pour «appeler chez eux». • Les appareils intelligents à la maison aideront les activités de vie quotidienne au besoin, comme la cuisine et, si les capacités de manipulation du robot s'améliorent suffisamment, s'habiller et se laver. Santé et bien-être • Les applications mobiles qui surveillent les mouvements et les activités, associées aux plates-formes sociales, pourront faire des recommandations pour maintenir la santé mentale et physique. • La surveillance de la santé à domicile et l'accès à l'information sur la santé seront en mesure de détecter des changements d'humeur ou de comportement et alerter les soignants. • Une gestion personnalisée de la santé aidera à atténuer les complexités associées à de multiples problèmes de comorbidité et / ou à des interactions de traitement. Traitements et dispositifs • Une meilleure aide auditive et des dispositifs d'assistance visuelle atténueront les effets de la perte d'audition et de vision, améliorant ainsi la sécurité et la connexion sociale. • La réadaptation personnalisée et la thérapie à domicile réduiront le besoin de séjours dans les hôpitaux ou les établissements de soins. • Les appareils d'assistance physique (marcheurs intelligents, fauteuils roulants et exo-squelettes) étendront la gamme des activités d'une personne handicapée. Le Panel d'étude s'attend à une explosion de technologies de détection à faible coût qui peuvent fournir des capacités substantielles aux personnes âgées dans leurs maisons. En principe, les agents sociaux ayant une présence physique et des capacités physiques simples (par exemple un Robot mobile doté de capacités de communication de base) pourraient constituer une plate-forme pour de nouvelles innovations. Cependant, cela nécessitera une intégration dans de multiples domaines du traitement de l'intelligence artificielle, du raisonnement, de l'apprentissage, de la perception et de la Robotique pour créer un système qui soit utile et utilisable par les personnes âgées.

Ces innovations introduiront également des questions concernant la vie privée dans différents milieux, y compris les amis, la famille et les soignants, et créeront de nouveaux défis pour accueillir une population toujours plus active et engagée loin de la retraite.

74 - Jennifer M. Ortman, Victoria A. Velkoff et Howard Hogan, «Une nation vieillissante: la population âgée aux États-Unis: estimations et projections de la population», Current Population Reports, US Census Bureau (mai 2014), a eu accès au 1er août 2016. 

Bien que l'éducation de qualité exige toujours un engagement actif des enseignants humains, IA promet d'améliorer l'éducation à tous les niveaux, en particulier en fournissant une personnalisation à l'échelle. Une meilleure aide auditive et des dispositifs d'assistance visuelle atténueront les effets de la perte d'audition et de vision, améliorant la sécurité et la connexion sociale. La réadaptation personnalisée et la thérapie à domicile réduiront le besoin de séjours à l'hôpital. Seront également possibles, tels que ceux conçus pour évaluer le vocabulaire, wh (who / what / when / where / why) qui questionne et les questions à choix multiples en utilisant des ressources électroniques telles que WordNet, Wikipedia et des ontologies en ligne. Avec l'explosion des cours en ligne, ces techniques sont sûrement adoptées pour être utilisées dans l'éducation en ligne. Bien que l'impact à long terme de ces systèmes sur le système éducatif reste incertain, la communauté d'IA a beaucoup appris en très peu de temps.

Analyse d'apprentissage :

Les ensembles de données recueillis à partir de systèmes d'apprentissage en ligne à grande échelle, allant des MOOC à l'Académie Khan, ainsi que des programmes en ligne à plus petite échelle, ont alimenté la croissance rapide du domaine de l'analyse de l'apprentissage. Les cours en ligne ne sont pas seulement bons pour une livraison généralisée, mais sont des véhicules naturels pour la collecte de données et l'instrumentation expérimentale qui contribueront aux résultats scientifiques et amélioreront la qualité de l'apprentissage à l'échelle des organisations telles que la Society for Learning Analytics Research (SOLAR) et la montée des conférences, y compris la Conference sur l'analyse et le savoir d'apprentissage ainsi que la Conférence Learning at Scale (L @ S) qui reflètent cette tendance. Cette communauté applique l'apprentissage approfondi, le traitement du langage naturel et d'autres techniques d'IA pour analyser l'engagement, le comportement et les résultats des élèves.

Les projets actuels cherchent à modéliser les idées fausses communes des élèves, à prédire quels élèves risquent d'être défaillants et à fournir des commentaires en temps réel des élèves qui sont étroitement intégrés aux résultats d'apprentissage. Des travaux récents ont également été consacrés à la compréhension des processus cognitifs impliqués dans la compréhension, l'écriture, l'acquisition de connaissances et la mémoire, et d'appliquer cette compréhension à la pratique éducative en développant et en testant des technologies éducatives.

Défis et opportunités :

On pourrait s'attendre à une utilisation de plus en plus sophistiquée des technologies de l'IA dans les écoles, les collèges et les Universités à ce jour. Une grande partie de son absence s'explique par le manque de ressources financières de ces institutions ainsi que par le manque de données établissant l'efficacité des technologies. Ces problèmes sont abordés, bien que lentement, par des fondations privées et par de nombreux programmes pour former principalement des enseignants du secondaire dans les programmes d'été. Comme dans d'autres domaines de l'IA, un type excessif de promesses concernant les capacités des MOOC ont signifié que les attentes dépassent souvent la réalité.

Les expériences de certaines institutions, telles que l'expérience de l'Université d'Etat de San Jose avec Udacity , ont conduit à une évaluation plus sobre du potentiel des nouvelles technologies éducatives.

86 - Jonathan C. Brown, Gwen A. Frishkoff et Maxine Eskenazi, «Génération automatique de questions pour l'évaluation du vocabulaire», Conférence sur les procédures de la technologie des langues humaines et la Conférence sur les méthodes empiriques dans le traitement du langage naturel (HLT / EMNLP), Vancouver, octobre 2005, (Association for Computational Linguistics, 2005)

87 - Michael Heilman, «Génération de questions factuelles automatiques à partir du texte», thèse de doctorat

CMULTI-11-004 (Université Carnegie Mellon, 2011), accessible le 1er août 2016.

88 - Tahani Alsubait, Bijan Parsia et Uli Sattler, "Générer des questions de choix multiples des ontologies: combien pouvons-nous aller?", Dans les eds. P. Lambrix, E. Hyvönen. E. Blomqvist, V. Presutti, G. Qi, U. Sattler, Y. Ding et C. Ghidini, Ingénierie du savoir et gestion du savoir: EKAW en 2014 Satellite Events, VISUAL, EKM1 et ARCOE-Logic Linköping, Suède, 24 novembre 2014 - Papiers choisis révisés, (Suisse: Springer International Publishing, 2015).

89 - La 6ème Conférence internationale sur l'analyse et la connaissance de l'apprentissage, accessible le 1er août 2016.

http://www.lak16.solaresearch.org/ 

90 -Troisième conférence annuelle ACM sur l'apprentissage à l'échelle,

91 -Ry Rivard, "Udacity Project on 'Pause'," Inside Higher Ed, 18 juillet 2013, accessible le 1er août 2016,

https://www.insidehighered.com/news/2013/07/18/citing-disappointing-student-outcomessan-jose-state-pauses-work-udacity 

Des logiciels téléchargeables surprenants et bienvenus et des systèmes en ligne tels que Carnegie Speech ou Duolingo fournissent une formation en langue étrangère à l'aide des techniques de reconnaissance automatique de la parole (ASR) et de la PNL pour reconnaître les erreurs de langue et aider les utilisateurs à les corriger.

80 - Des systèmes de tutorat tels que Carnegie Cognitive Tutor81 ont été utilisés dans les écoles secondaires américaines pour aider les élèves à apprendre des mathématiques. D'autres ITS ont été développés pour la formation en géographie, circuits, diagnostic médical, informatique et programmation, génétique et chimie. Les tuteurs cognitifs utilisent un logiciel pour imiter le rôle d'un bon précepteur humain, par exemple, en fournissant des conseils lorsqu'un étudiant est coincé sur un problème de mathématiques. Sur la base de l'indice demandé et de la réponse fournie, le tuteur offre des commentaires contextuels spécifiques. Les applications augmentent dans l'enseignement supérieur. Un IER appelé SHERLOCK commence à être utilisé pour enseigner aux techniciens de la Force aérienne à diagnostiquer les problèmes de systèmes électriques dans les avions. Et l'Institut des sciences de l'information de l'Université de Californie du Sud a développé des modules de formation plus avancés basés sur l'Avatar pour former le personnel militaire envoyé à des postes internationaux dans des comportements appropriés lorsqu'il s'agit de personnes de milieux culturels différents. Les nouveaux algorithmes pour le tutorat personnalisé, tels que Bayesian Knowledge Tracing, permettent un apprentissage individualisé de la maîtrise et le séquençage des problèmes. Le plus surprenant a été l'explosion des Cours Massive Open Online (MOOC) et d'autres modèles d'éducation en ligne à tous les niveaux, y compris l'utilisation par des outils comme Wikipedia et Khan Academy ainsi que des systèmes sophistiqués de gestion de l'apprentissage qui développent des outils d'éducation et d'apprentissage adaptatifs synchrones et asynchrones.

Depuis la fin des années 1990, des entreprises telles que le Educational Testing Service et Pearson ont développé des outils d'évaluation automatique de la PNL pour les essais de qualité dans les tests normalisés. Beaucoup de MOOC devenus si populaires, y compris ceux créés par EdX, Coursera et Udacity, utilisent les techniques de PNL, d'apprentissage par machine et de crowdsourcing pour classer les questions de réponse courte et de discussion, ainsi que les affectations de programmation. Les systèmes d'éducation en ligne qui soutiennent l'éducation professionnelle et l'apprentissage tout au long de la vie professionnelle de troisième cycle se développent également rapidement. Ces systèmes sont très prometteurs car le besoin d'une interaction en face à face est moins important pour les professionnels du travail et les changeurs de carrière. Bien qu'ils ne soient pas les leaders dans les systèmes et les applications compatibles avec l'IA, ils deviendront des adopteurs précoces car les technologies sont testées et validées. On peut affirmer que l'IA est la sauce secrète qui a permis aux instructeurs, en particulier dans l'enseignement supérieur, de multiplier la taille de leurs salles de classe par quelques ordres de grandeur, les classes de quelques dizaines de milliers d'étudiants ne sont pas rares. Afin de tester continuellement de grandes classes d'étudiants, la génération automatisée des questions est la rédaction d'Essais de Physique, "Intelligent Tutoring Systems: Actes de la 6ème Conférence Internationale, (Springer Berlin Heidelberg, 2002) - VanLehn et al, «L'architecture de Why2-Atlas». - «Ressources et soutien», Carnegie Learning, consulté le 1er août 2016.

82 - Alan Lesgold, Suzanne Lajoie, Marilyn Bunzo et Gary Eggan, «SHERLOCK: un environnement de pratique entraîné pour un travail de dépannage électronique», dans J-H Larkin et RW Chabay, éd., Instruction assistée par ordinateur et systèmes de tutorat intelligent: objectifs partagés des approches complémentaires (Hillsdale, New Jersey: Lawrence Erlbaum Associates, 1988). Michael V. Yudelson, Kenneth R. Koedinger et Geoffrey J. Gordon, (2013). «Modèles de traçabilité bayesiens individualisés,« Intelligence artificielle dans l'éducation »(Springer Berlin Heidelberg, 2013).

84 - Jill Burstein, Karen Kukich, Susanne Wolff, Chi Lu, Martin Chodorow, Lisa Braden-Harder et Mary Dee Harris, «Marquage automatisé utilisant une technique d'identification des caractéristiques hybrides» dans les Actes de la réunion annuelle de l'Association de linguistique computationnelle de Montréal , Canada, août 1998, a eu accès au 1er août 2016.

https://www.ets.org/Media/Research/pdf/erater_acl98.pdf  

85 - EdX,

https://www.edx.org/

86 -Coursera,

https://www.coursera.org/

Udacité,

Tous ont eu accès au 1er août 2016. On peut affirmer que IA est la sauce secrète qui a permis aux instructeurs, en particulier dans l'enseignement supérieur, de multiplier la taille de leurs salles de classe par quelques ordres de grandeur-classe de quelques dizaines de milliers d'élèves sont pas rare. L'absence actuelle d'utilisation sophistiquée des technologies de l'IA dans les écoles, les collèges et les universités s'explique peut-être par le manque de ressources financières et le manque de données établissant l'efficacité des technologies, en fournissant des outils et des formations relativement simples dans leur utilisation. Par exemple, un grand nombre d'applications éducatives, dont beaucoup sont gratuites, sont en cours d'élaboration pour l'iPad. Du côté négatif, il existe déjà une tendance majeure parmi les étudiants à restreindre leurs contacts sociaux à ceux de l'électronique et à consacrer beaucoup de temps sans contact social, en interagissant avec les programmes en ligne.

Si l'éducation se fait de plus en plus en ligne, quel effet le manque de contacts réguliers et en face-à-face avec les pairs aura-t-il sur le développement social des élèves?

Certaines technologies ont même été créées pour créer des effets secondaires neurologiques d'autre part, les enfants autistes ont déjà bénéficié d'interactions avec les systèmes d'IA.

96 - COMMUNAUTÉS À FAIBLE RESSOURCES :

Il existe de nombreuses opportunités pour que l'IA améliore les conditions pour les personnes vivant dans les communautés à faible ressources dans une ville typiquement nord-américaine - et même, dans certains cas, elle l'a déjà fait. La compréhension de ces contributions directes en matière d'IA peut également indiquer des contributions potentielles dans les régions les plus pauvres du monde en voie de développement. Il n'y a pas eu de concentration significative sur ces populations dans les rassemblements d'IA et, traditionnellement, les bailleurs de fonds de l'IA ont sous-investi dans une recherche qui manque d'applications commerciales. Avec des incitations ciblées et des priorités de financement, les technologies de l'IA pourraient aider à répondre aux besoins des communautés à faible ressources. Les efforts fructueux sont prometteurs. En contrecarrant les craintes que l'IA puisse contribuer au chômage et à d'autres problèmes sociétaux, l'IA peut apporter des mesures d'atténuation et de solutions, en particulier si elle est mise en œuvre de manière à renforcer sa confiance dans les communautés touchées.

L'apprentissage par machine, les approches d'exploration de données sous la bannière de la «science des données pour le bien social», l'IA a été utilisée pour créer des modèles prédictifs pour aider les organismes gouvernementaux à utiliser plus efficacement leurs budgets limités et pour résoudre des problèmes tels que l'empoisonnement au plomb, un grave problème de santé publique majeur.  Cela a été dans les informations en raison des événements en cours à Flint, Michigan. Les enfants peuvent être testés pour des niveaux élevés de plomb, mais cela signifie malheureusement que le problème n'est détecté qu'après avoir été empoisonné. De nombreux efforts sont en cours pour utiliser des modèles prédictifs pour aider les organismes gouvernementaux à hiérarchiser les enfants à risque, y compris ceux qui n'ont pas encore été exposés. De même, le Département des services sociaux de l'Illinois (IDHS) utilise des modèles prédictifs pour identifier les femmes enceintes à risque pour les résultats néfastes à la naissance afin de maximiser l'impact des soins prénataux. La Ville de Cincinnati les utilise pour identifier de manière proactive et déployer des inspecteurs sur des propriétés à risque de violation de code.

Planification :

Les techniques de planification et de planification des tâches ont été appliquées par de nombreux groupes différents pour distribuer de la nourriture avant qu'elle ne se gâte pour ceux qui peuvent en avoir accès, comme les restaurants, les banques alimentaires, les centres communautaires et les particuliers. Les scientifiques ont étudié, par exemple la façon dont la dépendance au GPS peut entraîner des changements dans l'hypocampe. Kim Tingley, «The Secrets of the Wave Pilots», The New York Times, le 17 mars 2016, a eu accès au 1er août 2016.

96 - Judith Newman, "À Siri, avec amour: comment un garçon avec autisme est devenu BFF avec Siri d'Apple", le New York Times, le 17 octobre 2014, a eu accès au 1er août 2016.

97 - Eric Potash, Joe Brew, Alexander Loewi, Subhabrata Majumdar, Andrew Reece, Joe Walsh, Eric Rozier, Emile Jorgensen, Raed Mansour et Rayid Ghani, «Modélisation prédictive de la santé publique: prévention de l'empoisonnement au plomb chez les enfants». - Actes du 21e ACM SIGKDD Conférence internationale sur la découverte de connaissances et l'exploration de données (New York: Association for Computing Machinery, 2015), 2039-2047. 98 Data Science for Social Good, Université de Chicago, a eu accès au 1er août 2016.

http://dssg.uchicago.edu/

99 - Senay Solak, Christina Scherrer et Ahmed Ghoniem, «Le problème de l'arrêt et de la baisse dans les réseaux de distribution alimentaire sans but lucratif», Annals of Operations Research 221, no. 1 (octobre 2014): dans les quinze prochaines années, il est probable que les enseignants humains seront aidés par les technologies de l'IA avec une meilleure interaction humaine, tant en classe que dans la maison. Le Panel d'étude s'attend à ce que des scénarios de réalité virtuelle plus généraux et plus sophistiqués dans lesquels les élèves puissent se plonger dans des matières de toutes les disciplines seront développés. Certaines étapes dans cette direction sont prises maintenant en augmentant les collaborations entre les chercheurs d'IA et les chercheurs en Sciences Humaines et sociales, illustrés par le projet de correspondance Galileo de Stanford et le projet de fabrication de connaissance par Columbia.

93 - Ces efforts interdisciplinaires créent des expériences interactives avec des documents historiques et l'utilisation de la réalité virtuelle (VR) pour explorer les sites archéologiques interactifs. Les techniques VR sont déjà utilisées dans les sciences naturelles telles que la biologie, l'anatomie, la géologie et l'astronomie pour permettre aux élèves d'interagir avec des environnements et des objets difficiles à s'engager dans le monde réel. La récréation des mondes passés et des mondes fictifs deviendra aussi populaire pour les études des arts et d'autres sciences. Les techniques d'IA vont de plus en plus flouer la ligne entre l'éducation formelle, l'éducation en classe et l'apprentissage individuel et individuel. Les systèmes d'apprentissage adaptatifs, par exemple, deviennent une partie essentielle du processus d'enseignement dans l'enseignement supérieur en raison des pressions pour contenir les coûts tout en servant un plus grand nombre d'étudiants et en transférant les étudiants à travers l'école plus rapidement. Alors que l'éducation formelle ne disparaîtra pas, le Panel d'étude croit que les MOOC et d'autres formes d'éducation en ligne feront partie de l'apprentissage à tous les niveaux, de la K-12 à l'Université, dans une expérience de classe mixte. 

Ce développement facilitera des approches plus adaptées à l'apprentissage, dans lesquelles les élèves peuvent apprendre à leur propre rythme en utilisant des techniques éducatives qui leur permettent le mieux. Les systèmes d'éducation en ligne apprendront à mesure que les élèves apprendront, en soutenant des progrès rapides dans notre compréhension du processus d'apprentissage. L'analyse d'apprentissage, à son tour, accélérera le développement d'outils pour une éducation personnalisée. La transition actuelle des livres papier aux médias et aux textes numériques et audio est susceptible de devenir répandue dans l'éducation aussi. Les appareils de lecture numérique deviendront beaucoup plus intelligents, ce qui permettra aux étudiants d'accéder facilement à des informations supplémentaires sur les sujets abordés. La technologie de traduction automatique (MT) facilitera également la traduction du matériel didactique en différentes langues avec une bonne précision, tout comme il traduit actuellement des manuels techniques. Les services de traduction de manuels scolaires qui ne dépendent actuellement que des traducteurs humains incorporeront de plus en plus de méthodes automatiques pour améliorer la rapidité et l'accessibilité de leurs services aux systèmes scolaires. Les systèmes d'apprentissage en ligne élargiront également l'opportunité aux adultes et aux professionnels qui travaillent d'améliorer leurs connaissances et leurs compétences (ou pour réorganiser et apprendre un nouveau domaine) dans un monde où ces domaines évoluent rapidement. Cela comprendra l'expansion de diplômes professionnels entièrement en ligne ainsi que des certifications professionnelles basées sur des cours en ligne. Conséquences sociétales plus larges dans les pays où l'éducation est difficile à obtenir pour l'ensemble de la population, les ressources en ligne peuvent avoir un effet positif si la population a les outils nécessaires pour y accéder. 

Le développement des ressources pédagogiques en ligne devrait permettre aux fondations qui soutiennent les programmes éducatifs internationaux de fournir une très bonne qualité d'enseignement. Université de Stanford: projet de correspondance Galileo, consulté le 1er août 2016.

93- Le projet Making and Knowing: reconstruire l'atelier du 16ème siècle de BNF MS. FR. à l'Université Columbia, consulté le 1er août 2016.

http://www.makingandknowing.org

94 - Paul James, "3D déchirée HTC Vive apporte l'archéologie à la vie", Road to VR, 31 août 2015, accessible le 1er août 2016. 

http://www.roadtovr.com/3d-mapped-htc-vive-demo-bringsarchaeology-to-life/ 

Alors que l'éducation formelle ne disparaîtra pas, le Panel d'étude croit que les MOOC et d'autres formes d'éducation en ligne feront partie de l'apprentissage à tous les niveaux, de la K-12 à l'université, dans une expérience de classe mixte. Avec des incitations ciblées et des priorités de financement, les technologies de l'IA pourraient aider à répondre aux besoins des communautés à faible revenu. Les efforts fructueux sont prometteurs avec la preuve de faute professionnelle de la Police. Ces améliorations pourraient conduire à une surveillance encore plus répandue. Certaines villes ont déjà ajouté des Drones à des fins de surveillance et l'utilisation par la Police de Drones pour maintenir la sécurité des ports, des aéroports, des zones côtières, des voies navigables, des installations industrielles est susceptibles d'augmenter, suscitant des préoccupations concernant la vie privée, la sécurité et d'autres problèmes. Le CompStat du Département de police de New-York a été le premier outil qui pointe vers la police prédictive et de nombreux services de police l'utilisent maintenant.

105 - L'apprentissage automatique améliore considérablement la capacité de prédire où et quand les crimes sont plus susceptibles de se produire et qui peut les commettre. Comme cela a été dramatisé dans le film Minority Report, les outils de police prédictive soulèvent le spectre des personnes innocentes qui sont injustement ciblées. Mais les outils de prédiction de l'IA bien déployés ont le potentiel de supprimer ou de réduire les biais humains, plutôt que de le renforcer et les recherches de ressources devraient être orientées vers cet effet. Les techniques d'IA peuvent être utilisées pour développer des simulations intelligentes pour former le personnel de la loi à collaborer. Alors que les organisations criminelles internationales et les terroristes de différents pays sont en collusion, les forces de police de différents pays sont encore confrontées à des difficultés à unir leurs forces pour les combattre. La formation de groupes internationaux de personnel chargé de l'application de la loi pour travailler comme équipes est très difficile.

L'Union Européenne, par le biais du programme Horizon 2020, supporte actuellement de telles tentatives dans des projets tels que LawTrain.

106 - La prochaine étape consistera à passer de la simulation aux enquêtes réelles en fournissant des outils qui soutiennent ces collaborations. Il existe des outils pour numériser Twitter et autres flux pour rechercher certains types d'événements et leur impact sur la Sécurité. Par exemple, l'IA peut aider à l'analyse des réseaux sociaux afin d'éviter que les personnes risquées ne soient radicalisées par ISIS ( état islamique ) ou d'autres groupes violents. Les organismes d'application de la loi sont de plus en plus intéressés à essayer de détecter les plans d'événements perturbateurs des réseaux sociaux et à surveiller l'activité de rassemblements de personnes pour analyser la Sécurité. Il existe un travail important sur les simulations de foule pour déterminer comment les foules peuvent être contrôlées. Dans le même temps, des préoccupations légitimes ont été soulevées au sujet du potentiel pour les organismes d'application de la loi de surpasser et d'utiliser de tels outils pour violer la vie privée des personnes. L'Administration américaine de la sécurité des transports (TSA), la Garde côtière et les nombreuses autres agences de sécurité qui comptent actuellement sur l'IA augmenteront probablement leur confiance afin de permettre des améliorations significatives de l'efficience et de l'efficacité.

107 - Les techniques de l'IA, la lecture, l'analyse de la parole et l'analyse de la démarche peuvent aiderles enquêteurs , les interrogateurs et les gardiens de sécurité pour détecter les faiblesses possibles et les comportements criminels. Par exemple, la TSA a actuellement un projet ambitieux pour refaire la sécurité des aéroports à l'échelle nationale.

108 - Appelé DARMS, le système est conçu pour améliorer l'efficacité et l'efficience de la sécurité de l'aéroport en s'appuyant sur des informations personnelles pour adapter la sécurité en fonction de la catégorisation des risques d'une personne et des vols pris en cours.

La vision future de ce projet est un tunnel qui vérifie la sécurité des personnes pendant qu'ils la traversent. Une fois de plus, les développeurs de cette technologie devraient veiller à ne pas créer de biais (p. Ex. Sur la catégorie de niveau de risque d'une personne) grâce à l'utilisation de jeux de données qui reflètent un biais antérieur. - Walter L. Perry, Brian McInnis, Carter C. Price, Susan Smith , Et John S. Hollywood, «Le rôle de la prévision de la criminalité dans les opérations d'application de la loi», Rand Corporation Report 233 (2013).

105 - "CompStat", Wikipedia, dernière modification le 28 juillet 2016, accessible le 1er août 2016. 

106 - LAW-TRAIN, consulté le 1er août 2016.

http://www.law-train.eu/

107 - Milind Tambe, sécurité et théorie des jeux: Algorithmes, systèmes déployés, leçons apprises (New-York: Cambridge University Press, 2011). 108 Peter Neffenger, «Budget 2017 de TSA - Un engagement envers la sécurité (partie I),« Department of Homeland Security, 1er mars 2016, consulté le 1er août 2016.

109 - Crawford, «Problème de Guy Blanc d'AI».

Raisonnement avec les réseaux sociaux et maximisation de l'influence :

Les réseaux sociaux peuvent être exploités pour créer des interventions plus précoces et moins coûteuses impliquant de grandes populations. Par exemple, IA pourrait être en mesure d'aider à diffuser des informations relatives à la santé. À Los Angeles, il y a plus de 5 000 jeunes sans abri (âgés de treize à vingt-quatre ans). Les interventions individuelles sont difficiles et coûteuses, et la méfiance des jeunes à l'égard de l'autorité dicte que les messages clés sont mieux répandus par les pairs. Les programmes d'IA pourraient être en mesure d'exploiter les réseaux sociaux des jeunes sans-abri afin de sélectionner stratégiquement les leaders des pairs afin de diffuser des informations sur la santé, telles que la façon d'éviter la propagation du VIH. La nature dynamique et incertaine de ces réseaux pose des défis pour la recherche sur l'IA.

100 - Il faut également veiller à ce que les systèmes d'IA reproduisent des comportements discriminatoires, tels que l'apprentissage par machine qui identifie les personnes à travers des indicateurs raciaux illégaux ou par des facteurs de substitution hautement corrélés, Tels que les codes postaux. Mais si déployé avec beaucoup de soin, une dépendance accrue à l'IA peut entraîner une réduction de la discrimination globale, puisque les programmes d'IA sont intrinsèquement plus facilement audités que les humains. 

SÉCURITÉ PUBLIQUE ET SÉCURITÉ :

Les villes ont déjà commencé à déployer des technologies IA pour la sécurité et la sécurité publique. En 2030, la ville nord-américaine typique dépendra fortement d'eux. Ceux-ci incluent des caméras pour la surveillance qui peuvent détecter des anomalies indiquant un éventuel crime, des Drones et des applications de Police prédictive. Comme pour la plupart des problèmes, il y a des avantages et des risques. Gagner de l'intérêt public est crucial. Bien qu'il existe des préoccupations légitimes que la police qui intègre l'IA peut devenir dominante ou omniprésente dans certains contextes, le contraire est également possible. IA peut permettre aux policiers de devenir plus ciblés et utilisés uniquement lorsque cela s'avère nécessaire. En supposant un déploiement minutieux, l'IAI peut également aider à supprimer une partie du biais inhérent à la prise de décision humaine. L'une des utilisations les plus réussies de l'analyse de l'IA consiste à détecter la criminalité en col blanc, comme la fraude par carte de crédit.

101 - La cybersécurité (y compris le spam) est une préoccupation largement partagée et l'apprentissage par machine a un impact. Les outils d'IA peuvent également s'avérer utiles pour aider la police à gérer les scènes de crime ou les événements de recherche et de sauvetage en aidant les commandants à hiérarchiser les tâches et à allouer des ressources, même si ces outils ne sont pas encore prêts à automatiser de telles activités. L'amélioration de l'apprentissage en machine en général et le transfert de l'apprentissage en particulier - pour accélérer l'apprentissage dans de nouveaux scénarios en fonction de similitudes avec des scénarios passés - peuvent faciliter ces systèmes.

Les caméras déployées presque partout dans le monde aujourd'hui ont tendance à être plus utiles pour aider à résoudre les crimes que ne les empêchent. Cela est dû à la faible qualité de l'identification des événements à partir des vidéos et au manque de main-d'œuvre pour regarder les flux vidéo massifs. À mesure que l'IA pour ce domaine s'améliore, elle aidera mieux la prévention et la poursuite du crime grâce à une plus grande précision de la classification des événements et au traitement automatique efficace de la vidéo pour détecter les anomalies, y compris, potentiellement.

100 - Jordan Pearson, «L'intelligence artificielle pourrait aider à réduire le VIH chez les jeunes sans abri», Teamcore, Université de Californie du Sud, 4 février 2015, accessible le 1er août 2016.

101 - "RSA Adaptive Authentication", RSA, consulté le 1er août 2016, 

102 - Takeshi Arikuma et Yasunori Mochizuki, «Système de surveillance multimédia intelligent pour les villes plus sûres» Transactions APSIPA sur le traitement du signal et de l'information 5 (2016): 1-8. 103 "Big Op-Ed: décalage des avis sur les caméras de surveillance", Talk of the Nation, NPR, 22 avril 2013, consulté le 1er août 2016. 

http://www.npr.org/2013/04/22/178436355/big-op-ed-shiftingopinions-on-surveillance-camera/ 

L'une des utilisations les plus réussies de l'analyse de l'IA consiste à détecter la criminalité en col blanc, comme la fraude par carte de crédit. La cybersécurité (y compris le spam) est une préoccupation largement partagée, et l'apprentissage par machine a un impact. Comme cela a été dramatisé dans le film Minority Report, les outils de police prédictive soulèvent le spectre des personnes innocentes qui sont injustement ciblées. Mais les outils de prédiction de l'IA bien déployés ont le potentiel de supprimer ou de réduire les biais humains, marchés, qui ont souvent pour effet de réduire les obstacles à l'entrée et d'accroître la participation - des magasins d'applications à AirBnB à taskrabbit. Une communauté de recherche dynamique au sein d'IA étudie d'autres façons de créer de nouveaux marchés et de faire fonctionner plus efficacement les systèmes existants. Bien que le travail ait une valeur intrinsèque, la plupart des gens travaillent pour pouvoir acheter des biens et des services qu'ils apprécient. Étant donné que les systèmes IA exécutent un travail qui exigeait auparavant le travail humain, ils ont pour effet de réduire le coût de nombreux biens et services, ce qui rend le monde entier plus enrichi.

Mais, comme le montrent les débats politiques actuels, la perte d'emploi est plus marquée pour les personnes - en particulier celles qui sont directement touchées - que les gains économiques diffus, et l'IA est malheureusement souvent considérée comme une menace pour les emplois plutôt que pour une volonté de vivre. Il y a même peur dans certains milieux que les progrès de l'IA seront aussi rapides que pour remplacer tous les emplois humains - y compris ceux qui sont en grande partie cognitifs ou impliquent un jugement - au sein d'une seule génération. Ce scénario soudain est hautement improbable, mais IA va graduellement envahir presque tous les secteurs de l'emploi, ce qui nécessite un déplacement du travail humain que les ordinateurs peuvent prendre en charge. Les effets économiques de l'IA sur les emplois humains cognitifs seront analogues aux effets de l'automatisation et de la Robotique chez les humains dans les emplois manufacturiers. Beaucoup de travailleurs d'âge moyen ont perdu des emplois d'usine bien rémunérés et le statut socioéconomique de la famille et de la société qui ont traditionnellement travaillé avec de tels emplois. Une fraction encore plus importante de la main-d'œuvre totale peut, à terme, perdre des emplois «cognitifs» bien rémunérés. Comme le travail devient un facteur moins important dans la production par rapport au capital intellectuel, la majorité des citoyens peuvent trouver la valeur de leur travail insuffisante pour payer un niveau de vie socialement acceptable. Ces changements nécessiteront une réponse politique, plutôt que purement économique, à propos du type de filets de sécurité sociale qui devraient être mis en place pour protéger les gens des grands changements structurels de l'économie. En l'absence de politiques atténuantes, les bénéficiaires de ces changements peuvent être un petit groupe à la strate supérieure de la société. À court terme, l'éducation, la rééducation et l'invention de nouveaux biens et services peuvent atténuer ces effets. À plus long terme, le filet de sécurité sociale actuel doit devoir évoluer vers de meilleurs services sociaux pour tous, tels que les soins de santé et l'éducation, ou un revenu de base garanti. En effet, des pays comme la Suisse et la Finlande ont activement pris en compte ces mesures. 

L'IA peut être considérée comme un mécanisme radicalement différent de création de richesse dans lequel tout le monde devrait avoir droit à une partie du trésor mondial produit par l'IA. Il n'est pas trop tôt pour un débat social sur la manière dont les fruits économiques des technologies IA devraient être partagé. Comme les enfants dans les sociétés traditionnelles soutiennent leurs parents vieillissants, nos «enfants» artificiellement intelligents devraient nous soutenir, les «parents» de leur intelligence.

113 - Par exemple, Brynjolfsson et McAfee, Second Machine Age, ont deux chapitres consacrés à cela (Erik Brynjolfsson et Andrew McAfee, The Second Machine Age: travail, progrès et prospérité dans un temps de Brilliant Technologies, (New York: WW Norton & Company, Inc., 2014)) et Brynjolfsson, McAfee et Spence décrivent les réponses politiques pour la combinaison de la mondialisation et de la technologie numérique (Erik Brynjolfsson, Andrew McAfee et Michael Spence, Affaires étrangères, juillet / août 2014, ont eu accès au 1er août, 2016. 

114 - Le PIB ne fait pas un bon travail pour mesurer la valeur de nombreux produits numériques. Lorsque la société ne parvient pas à gérer ce qui n'est pas mesuré, il en découle une mauvaise décision politique. Une alternative est de regarder l'excédent du consommateur, pas seulement les flux de dollars. Comme IA est incorporé dans plus de biens, cette question devient plus saillante. Il se peut que le PIB baisse, mais les gens ont un meilleur bien-être grâce à l'accès à ces produits numériques. Voir Erik Brynjolfsson et Adam Saunders, «Ce que le PIB devient mal (pourquoi les gestionnaires devraient s'occuper)», Sloan Management Review, vol. 51, no. 1 (1er octobre 2009):

95-96 - EMPLOI ET LIEU DE TRAVAIL :

Alors que les technologies de l'IA sont susceptibles d'avoir un impact profond sur l'avenir des tendances de l'emploi et du lieu de travail dans une ville typiquement nord-américaine, il est difficile d'évaluer avec précision les impacts actuels, positifs ou négatifs. Au cours des quinze dernières années, l'emploi s'est déplacé en raison d'une récession majeure et d'une mondialisation croissante, en particulier avec l'introduction de la Chine à l'économie mondiale, ainsi que des changements énormes dans la technologie numérique non-IA. Depuis les années 1990, les États-Unis ont connu une croissance continue de la productivité et du PIB, mais le revenu médian a stagné et le ratio emploi / population a diminué. Il existe des exemples clairs d'industries dans lesquelles les technologies numériques ont eu des répercussions profondes, bonnes et mauvaises, et d'autres secteurs dans lesquels l'automatisation fera vraisemblablement des changements majeurs dans un proche avenir. Beaucoup de ces changements ont été fortement influencés par les technologies numériques «de routine», y compris la planification des ressources d'entreprise, la mise en réseau, le traitement de l'information et la recherche. La compréhension de ces changements devrait donner un aperçu de la façon dont l'IA affectera la demande de main-d'œuvre future, y compris le changement dans les demandes de compétences. À ce jour, les technologies numériques ont touché davantage les travailleurs du milieu qualifié, comme les agents de voyages, plutôt que le travail le plus qualifié dans l'Industrie.

110 - D'autre part, l'éventail des tâches que les systèmes numériques peuvent faire évolue au fur et à mesure que les systèmes IA s'améliorent, ce qui risque d'augmenter progressivement la portée de ce qui est considéré comme une routine. L'IA est également en train de se diriger vers le haut de gamme du spectre, y compris les services professionnels non historiquement réalisés par des machines. Pour réussir, les innovations d'IA devront surmonter les craintes humaines compréhensibles d'être marginalisées. L'IA remplacera probablement les tâches plutôt que les emplois à court terme et créera également de nouveaux types d'emplois. Mais les nouveaux emplois qui apparaîtront sont plus difficiles à imaginer à l'avance que les emplois existants qui seront probablement perdus. Les changements dans l'emploi se produisent généralement progressivement, souvent sans une transition nette, une tendance susceptible de se poursuivre alors que l'IA se déplace lentement vers le lieu de travail. Un éventail d'effets apparaîtra, allant de petites quantités de remplacement ou d'augmentation à un remplacement complet. Par exemple, bien que la plupart des emplois d'avocats ne soient pas encore automatisés l'IA a appliqué l'extraction de l'information juridique et la modélisation des sujets, il a aussi automatisé certaines parties des emplois des avocats de première année.

112 - Dans un avenir pas trop lointain, un éventail diversifié de titulaires d'emploi , des radiologues aux chauffeurs de camions aux jardiniers, peuvent être affectés. L'IA peut également influencer la taille et l'emplacement de la main-d'œuvre. De nombreuses organisations et institutions sont nombreuses parce qu'elles exercent des fonctions qui ne peuvent être mises à l'échelle qu'en ajoutant le travail humain, soit «horizontalement» dans les zones géographiques, soit «verticalement» dans les hiérarchies de gestion. Comme l'IA prend en charge de nombreuses fonctions, l'évolutivité n'implique plus les grandes organisations. Beaucoup ont noté le petit nombre d'employés de certaines entreprises Internet haut de gamme, mais pas d'autres. Il peut y avoir une échelle naturelle de l'entreprise humaine, peut-être où le P-DG peut connaître tout le monde dans l'entreprise. Grâce à la création de marchés de travail efficacement externalisés par l'IA, les entreprises peuvent avoir tendance à atteindre cette taille naturelle. L'IA créera également des emplois, en particulier dans certains secteurs, en rendant certaines tâches plus importantes et en créant de nouvelles catégories d'emplois en rendant possibles les nouveaux modes d'interaction. Des systèmes d'information sophistiqués peuvent être utilisés pour créer de nouveaux emplois.

110 - Jeremy Ashkenas et Alicia Parlapiano, «Comment la récession a transformé l'économie, en 255 graphiques», The New York Times, le 6 juin 2014, a eu accès au 1er août 2016. 

111 - R Dana Remus et Frank S. Levy, "Can Robots Be Lawyers? Les ordinateurs, les avocats et la pratique du droit, «Réseau de recherche en sciences sociales, modifié pour la dernière fois le 12 février 2016, a eu accès au 1er août 2016,

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2701092

112 - John Markoff, «Les armées d'avocats coûteux, remplacés par un logiciel moins cher», The New York Times, le 4 mars 2011, ont eu accès au 1er août 2016.

Comme le travail devient un facteur moins important dans la production par rapport au capital intellectuel, la majorité des citoyens peuvent trouver la valeur de leur travail insuffisante pour payer un niveau de vie socialement acceptable. L'IA remplacera probablement les tâches plutôt que les emplois à court terme et créera également de nouveaux types d'emplois. Mais les nouveaux emplois qui apparaîtront sont plus difficiles à imaginer à l'avance que les emplois existants qui seront probablement perdus.

Imaginer l'avenir :

Le succès de toute forme de divertissement est finalement déterminé par les individus et les groupes sociaux qui sont ses sujets. Les modes de divertissement que les gens trouvent attrayants sont diversifiés et changent avec le temps. Il est donc difficile de prévoir les formes que les spectacles auront dans les quinze prochaines années précisément. Néanmoins, les tendances actuelles suggèrent au moins quelques caractéristiques que le futur paysage de divertissement devrait contenir. À ce jour, la révolution de l'information s'est largement déroulée dans les logiciels. Cependant, avec la disponibilité croissante de capteurs et de dispositifs moins coûteux, une plus grande innovation dans le matériel utilisé dans les systèmes de divertissement est attendue. La réalité virtuelle et les haptiques pourraient entrer dans nos salons. Des Robots compagnons personnalisés sont déjà en cours de développement. Avec les améliorations complémentaires apportées à la reconnaissance automatique de la parole, le Panel d'étude s'attend à ce que l'interaction avec les robots et autres systèmes de divertissement soit axée sur le dialogue, peut-être contraint, mais qui commence a être progressivement plus humain. !

De même, on prévoit que les systèmes d'interaction développent de nouvelles caractéristiques telles que l'émotion, l'empathie et l'adaptation aux rythmes environnementaux tels que l'heure du jour. Aujourd'hui, un amateur avec une caméra vidéo et des outils logiciels facilement disponibles peut faire un film relativement bon. À l'avenir, des outils et des applications plus sophistiqués seront disponibles pour rendre encore plus facile la production de contenu de haute qualité, par exemple, pour composer de la musique ou pour chorégraphier la danse à l'aide d'un Avatar. La création et la diffusion du divertissement bénéficieront des progrès des technologies telles que l'ASR, le doublage et la traduction automatique, qui permettront de personnaliser le contenu à différents publics à peu de frais. Cette démocratisation et la prolifération des Médias créés par l'IA rendent difficile de prédire comment le goût du divertissement par les humains, qui sont déjà fluides, évoluera. Avec des contenus de plus en plus livrés numériquement, et de nombreuses quantités de données étant enregistrées sur les préférences des consommateurs et les caractéristiques d'utilisation, les puissances multimédias pourront micro-analyser et micro-servir des contenus à des segments de plus en plus spécialisés de la population-vers le bas à l'individu. Concevable la scène est prévue pour l'émergence de conglomérats médiatiques agissant comme «grands frères» qui sont capables de contrôler les idées et les expériences en ligne auxquelles des individus spécifiques sont exposés. Il reste à voir si une société plus large développera des mesures pour prévenir leur émergence. Ce sujet, ainsi que d'autres concernant la politique liée à l'IA, est traité plus en détail dans la section suivante :

Emoters, consulté le 1er août 2016. 

http://emoterbots.com/

124 - "Siri", Apple, Inc., a eu accès au 1er août 2016,

http://www.apple.com/in/ios/siri/

125 - Ryan Calo, «Manipulation du marché numérique», George Washington Law Review 82, no. 4 (2014): 995-1051.

DIVERTISSEMENT :

Avec la croissance explosive d'Internet au cours des quinze dernières années, peu d'entre nous peuvent imaginer leur vie quotidienne sans Internet. Alimenté par l'IA, Internet a établi le contenu généré par les utilisateurs comme une source viable d'informations et de divertissement. Les réseaux sociaux tels que Facebook, Google, Twitter... sont maintenant omniprésents et ils fonctionnent comme des canaux personnalisés d'interaction et de divertissement sociaux, parfois au détriment de l'interaction interpersonnelle. Des applications telles que WhatsApp et Snapchat permettent aux utilisateurs de smartphones de rester constamment «en contact» avec leurs pairs et de partager des sources de divertissement et d'informations. Dans les communautés en ligne telles que Second Life et les jeux de rôle tels que World of Warcraft, les gens imaginent une existence alternative dans un monde virtuel.

115 - Les appareils spécialisés, tels que le Kindle d'Amazon, ont également redéfini l'essentiel des passe-temps chers. Les livres peuvent maintenant être parcourus et achetés avec quelques balayages du doigt, stockés par milliers dans un appareil de poche, et lus de la même manière qu'un livre de poche. Des plates-formes de confiance existent maintenant pour partager et naviguer sur des blogs, des vidéos, des photos et des discussions d'actualité, en plus d'une variété d'autres informations générées par l'utilisateur. Pour fonctionner à l'échelle d'Internet, ces plates-formes doivent s'appuyer sur des techniques qui sont activement développées dans le traitement du langage naturel, la récupération d'informations, le traitement d'image, le crowdsourcing et l'apprentissage par machine. Des algorithmes tels que le filtrage collaboratif ont été développés, par exemple, pour recommander des films, des chansons ou des articles pertinents basés sur les détails démographiques de l'utilisateur et l'historique de navigation.

116 - Les sources traditionnelles de divertissement ont également embrassé l'IA pour suivre le rythme des temps. Comme le montre le livre et le film Moneyball, le sport professionnel est maintenant soumis à une analyse quantitative intensive. Au-delà des statistiques globales de performance, les signaux sur le terrain peuvent être surveillés à l'aide de capteurs et de caméras sophistiqués. Le logiciel a été créé pour composer de la musique et reconnaître les bandes sonores. Les techniques issues de la vision par ordinateur et de la PNL ont été utilisées pour créer des performances en scène. Même l'utilisateur isolé peut exercer sa créativité sur des plates-formes telles que WordsEye, qui génère automatiquement des scènes 3D d'origine naturelle. Le texte de la langue de l'IA a également été utilisé à l'aide de la recherche historique dans les arts il est largement utilisé dans la stylométrie et, plus récemment, dans l'analyse des tableaux. L'enthousiasme avec lequel les humains ont répondu au divertissement piloté par l'IA a été surprenant et a conduit à des inquiétudes qu'il réduise l'interaction interpersonnelle entre les êtres humains. On avait peu prédit que les gens passent des heures à se connecter avec un affichage. Les enfants semblent souvent être plus heureux de jouer à la maison sur leurs appareils plutôt que dehors avec leurs amis. Car l'IA facilitera de plus en plus le divertissement plus interactif, personnalisé et plus engageant. La recherche devrait viser à comprendre comment tirer parti de ces attributs pour les bénéfices des individus et de la société.

115 - Second Life, consulté le 1er août 2016 

http://secondlife.com

"World of  Warcraft", Blizzard Entertainment, Inc, a eu accès au 1er août 2016,

http://secondlife.com

116 - John S. Breese, David Heckerman et Carl Kadie, «Analyse empirique des algorithmes prédictifs pour le filtrage collaboratif», Actes de la 14e Conférence sur l'incertitude dans l'intelligence artificielle (juillet 1998), accessible le 1er août 2016,

http://arxiv.org/pdf/1301.7363.pdf

43-52. 117 - Michael Lewis, Moneyball: L'art de gagner un jeu déloyal (New York: W. W. Norton & Company, Inc., 2003):

http://www.imdb.com/title/tt1210166/ 

118 - MuseScore, consulté le 1er août 2016.

https://musescore.org/

119 - Shazam, consulté le 1er août 2016 

http://www.shazam.com/

120 - Annie Dorsen, consulté le 1er août 2016, 

http://www.anniedorsen.com/

121 - WordsEye, consulté le 1er août 2016,

https://www.wordseye.com 

122 - "Stylometry", Wikipedia, modifiée pour la dernière fois le 4 août 2016, accessible le 1er août 2016,

http://arxiv.org/pdf/1408.3218v1.pdf

http://arxiv.org/pdf/1408.3218v1.pdf

 

L'IA facilitera de plus en plus le divertissement plus interactif, personnalisé et plus engageant. La recherche devrait viser à comprendre comment tirer parti de ces attributs pour les bénéfices des individus et de la société. Des outils et des applications plus sophistiqués seront disponibles pour rendre encore plus facile de produire un contenu de haute qualité, par exemple, pour composer de la musique ou pour chorégraphier la danse à l'aide d'un Avatar.

42 43 - De même, l'IA pourrait élargir les inégalités d'opportunité existantes si l'accès aux technologies de l'IA, ainsi que le calcul à haute puissance et les données à grande échelle qui alimentent bon nombre d'entre elles, est injustement réparti dans toute la société. Ces technologies amélioreront les capacités et l'efficacité des personnes qui y ont accès. Une personne ayant accès à une technologie de traduction automatique précise sera mieux en mesure d'utiliser les ressources d'apprentissage disponibles dans différentes langues. De même, si la technologie de traduction vocale n'est disponible qu'en anglais, les personnes qui ne parlent pas anglais seront désavantagées. En outre, les applications IA et les données auxquelles elles s'appuient peuvent refléter les biais de leurs concepteurs et utilisateurs, qui spécifient les sources de données. Cela menace d'approfondir les préjugés sociaux existants et de concentrer inégalement les bénéfices d'IA entre différents sous-groupes de la société. Par exemple, certaines technologies de reconnaissance vocale ne fonctionnent pas bien pour les femmes et les personnes ayant des accents. Comme l'IA est de plus en plus utilisé dans les applications critiques, ces biais peuvent poser des problèmes d'équité à divers groupes dans la société. D'autre part, par rapport aux biais bien documentés dans la prise de décision humaine, les outils de prise de décision fondés sur l'IA ont le potentiel de réduire de manière significative le biais dans les décisions critiques telles que celui de prêter de l'argent à une personne ou l'envoyez en prison. Les préoccupations de protection de la vie privée concernant la surveillance de l'IA sont également répandues, en particulier dans les villes dotées d'une instrumentation omniprésente. Sous-surveillance, l'enregistrement d'une activité par un participant, généralement avec des dispositifs personnels portables, a également augmenté. Étant donné que les opinions sur le biais et la vie privée sont basées sur des jugements éthiques et de valeur personnels et sociétaux, les débats sur la façon de répondre à ces préoccupations vont probablement grandir et résister à une résolution rapide. De même, étant donné que l'IA génère des richesses importantes, les débats se développeront quant à la manière dont les fruits économiques des technologies de l'IA devraient être partagés, d'autant plus que l'expertise de l'IA et les ensembles de données sous-jacentes qui alimentent les applications sont concentrés dans un petit nombre de Grandes Entreprises. 

Pour aider à répondre à ces préoccupations concernant les implications individuelles et sociétales des technologies de l'IA en évolution rapide, le Panel d'étude propose trois recommandations générales:

1. Définir un chemin vers l'acquisition d'une expertise technique en IA à tous les niveaux de Gouvernement. Une gouvernance efficace nécessite plus d'experts qui comprennent et peuvent analyser les interactions entre les technologies de l'IA, les objectifs programmatiques et les valeurs générales de la société. En l'absence d'expertise technique suffisante pour évaluer la sécurité ou d'autres paramètres, les responsables nationaux ou locaux peuvent refuser d'autoriser une demande potentiellement prometteuse. Ou les fonctionnaires insuffisamment formés peuvent simplement prendre la parole des technologues de l'industrie de la lumière verte ( les lasers ), une application sensible qui n'a pas été correctement vérifiée. Sans comprendre comment les systèmes d'IA interagissent avec le comportement humain et les valeurs sociétales, les fonctionnaires seront mal positionnés pour évaluer l'impact de l'IA sur les objectifs programmatiques.

2. Supprimer les obstacles perçus et réels à la recherche sur l'équité, la sécurité, la vie privée et les impacts sociaux des systèmes d'IA. Certaines interprétations des lois fédérales telles que la Loi sur la fraude et l'abus en matière d'informatique et la disposition anti-contournement de la Loi sur le droit d'auteur numérique du millénaire actuel sont ambiguës quant à savoir si et comment les systèmes d'IA exclusifs peuvent être conçus et évalués par des universitaires, des journalistes et d'autres chercheurs. Une telle recherche est essentielle si les systèmes d'IA ayant des conséquences matérielles et autres matérielles doivent être correctement examinés et tenus responsables.

3. Augmenter le financement public et privé pour les études interdisciplinaires sur les impacts sociétaux de l'IA. En tant que société, nous subventionnons des ressources dans la recherche sur les implications sociétales des technologies de l'IA. Les fonds privés et publics devraient être orientés vers l'interdisciplinarité.

SECTION III: PERSPECTIVES ET RECOMMANDATIONS POUR LA POLITIQUE PUBLIQUE DE L'IA

L'objectif des applications de l'IA doit être de créer de la valeur pour la société. Nos recommandations politiques découlent de cet objectif et, alors que ce rapport se concentre sur une ville typiquement nord-américaine en 2030, les recommandations sont largement applicables à d'autres endroits au fil du temps. Les stratégies qui améliorent notre capacité à interpréter les systèmes d'IA et à participer à leur utilisation peuvent contribuer à renforcer la confiance et à éviter les échecs drastiques. Il faut veiller à accroître et à renforcer les capacités humaines et l'interaction, et à éviter la discrimination à l'égard des segments de la société. Les recherches visant à encourager cette direction et à informer les débats sur les politiques publiques devraient être soulignées.

Compte tenu de la réglementation actuelle sectorielle des industries américaines, des lois et des politiques nouvelles ou réorganisées seront nécessaires pour remédier aux effets généralisés que l'IA risque d'apporter. Plutôt que d'une réglementation «plus» ou «plus stricte», les politiques devraient être conçues pour encourager l'innovation, générer et transférer de l'expertise, et favoriser une large responsabilité de l'entreprise et de la citoyenneté pour aborder les problèmes sociétaux critiques soulevés par ces technologies. À long terme, l'IA permettra une nouvelle création de richesse qui nécessitera un débat social sur la manière dont les fruits économiques des technologies de l'IA devraient être partagés.

POLITIQUE D'AMI, MAINTENANT ET DANS LE FUTUR

Au fil de l'histoire, les humains se sont façonnés et adaptés aux nouvelles technologies. Ce rapport prévoit que les progrès des technologies de l'IA seront développés et mis en place de manière progressive, et non pas dans des sauts soudains et inattendus dans les techniques elles-mêmes en s'appuyant sur ce qui existe aujourd'hui, facilitant cette adaptation. D'autre part, de petites améliorations aux techniques, au pouvoir informatique ou à la disponibilité des données peuvent parfois conduire à de nouvelles applications qui changent de jeu. La mesure du succès pour les applications d'IA est la valeur qu'elles créent pour des vies humaines. Dans l'avenir, la facilité avec laquelle les gens utilisent et s'adaptent aux applications d'IA déterminera en grande partie leur succès. À l'inverse, les applications d'IA sont sensibles aux erreurs et aux défaillances, mais la réussite de ces dernières sera la façon dont les utilisateurs perçoivent et tolèrent leurs lacunes. À mesure que l'IA devient de plus en plus intégrée dans la vie quotidienne et utilisée pour des tâches plus critiques, les erreurs du système peuvent entraîner une réaction négative des utilisateurs et affecter négativement leur confiance. Bien que les accidents dans une voiture à auto-conduite soient moins probables que ceux entraînés par les humains, par exemple, ils attireront plus d'attention. Les stratégies de conception qui améliorent la capacité des humains à comprendre les systèmes et les décisions de l'IA (telles que l'explication explicite de ces décisions) et de participer à leur utilisation peuvent contribuer à renforcer la confiance et à prévenir les échecs drastiques. De même, les développeurs devraient aider à gérer les attentes des gens, ce qui affectera leur bonheur et leur satisfaction vis-à-vis des applications de l'IA. La frustration dans la réalisation des fonctions promises par un système diminue la confiance des gens et réduit leur volonté d'utiliser le système à l'avenir. Une autre considération importante est de savoir comment les systèmes d'IA qui assument certaines tâches affecteront les capacités des personnes. À mesure que les machines fournissent des performances super-humaines sur certaines tâches, la capacité des gens à les réaliser peut diminuer. Déjà, l'introduction de calculateurs dans les salles de classe a réduit la capacité des enfants à effectuer des opérations

Publicité
Publicité
Commentaires
Publicité