2
OVERVIEW OF CURRENT ARTIFICIAL INTELLIGENCE
James Fackler, Johns Hopkins Medicine, and Edmund Jackson, Hospital Corporation of America
INTRODUCTION
This chapter first acknowledges the roots of artificial intelligence (AI). We then briefly touch on areas outside of medicine where AI has had an impact and highlight where lessons from these other industries might cross into health care.
For decades, the attempt to capture knowledge in the form of a book has been challenging, as indicated by the adage “any text is out of date by the time the book is published.” However, in 2019, with what has been determined by some analyses as exponential growth in the field of computer science and AI in particular, change is happening at a pace that renders sentences in this chapter out of date almost immediately. To stay current, we can no longer rely on monthly updates from a stored PubMed search. Rather, daily news feeds from sources such as the Association for the Advancement of Artificial Intelligence or arXiv1 are necessary. As such, this chapter contains references to both historical publications as well as websites and web-based articles.
It is surpassingly difficult to define AI, principally because it has always been loosely spoken of as a set of human-like capabilities that computers seem about ready to replicate. Yesterday’s AI is today’s commodity computation. Within that caveat, we aligned this chapter with the definition of AI in Chapter 1. A formal definition of AI starts with the Oxford English Dictionary: “The capacity of computers or other machines to exhibit or simulate intelligent behavior; the field of study concerned with this,” or Merriam-Webster online: “1: a branch
___________________
of computer science dealing with the simulation of intelligent behavior in computers, 2: the capability of a machine to imitate intelligent human behavior.”
HISTORICAL PERSPECTIVE
If the term “artificial intelligence” has a birthdate, it is August 31, 1955, when John McCarthy, Marvin L. Minsky, Nathaniel Rochester, and Claude E. Shannon submitted “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.” The second sentence of the proposal reads, “The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it” (McCarthy et al., 2006). Naturally, the proposal and the resulting conference—the 1956 Dartmouth Summer Research Project on Artificial Intelligence—were the culmination of decades of thought by many others (Buchanan, 2005; Kline, 2011; Turing, 1950; Weiner, 1948). Although the conference produced neither formal collaborations nor tangible outputs, it clearly galvanized the field (Moor, 2006).
Thought leaders in this era saw the future clearly, although optimism was substantially premature. In 1960, J. C. R. Licklider wrote
The hope is that, in not too many years, human brains and computing machines will be coupled together very tightly, and that the resulting partnership will think as no human brain has ever thought and process data in a way not approached by the information-handling machines we know today. (Licklider, 1960)
Almost 60 years later, we are closer but not there yet.
Two major competing schools of thought developed in approaching AI: (1) symbolic representation and formal logic expressed as expert systems and advanced primarily with Lisp, a family of computer programming languages (and Prolog in Europe) by John McCarthy, and (2) conceptualization and mathematical frameworks for mirroring neurons in the brain, formalized as “perceptrons” by Frank Rosenblatt (1958; see also McCarthy, 1958). The latter was initially known as the connectionist school, but we now know the technique as artificial neural networks. The McCarthy school of formal logic was founded by the technical paper “Programs with Common Sense,” in which McCarthy defines and creates the first full AI program: Advice Taker (McCarthy, 1959). The major thrust of the paper is that “in order for a program to be capable of learning something it must first be capable of being told it,” and hence the formalization of declarative logic programming. By the 1970s, however, excitement gave way to disappointment because early successes that worked in well-structured narrow problems failed to
either generalize to broader problem solving or deliver operationally useful systems. The disillusionment, summarized in the Automatic Language Processing Advisory Committee report (NRC, 1966) and the Lighthill report (Lighthill, 1973), resulted in an “AI Winter” with shuttered projects, evaporation of research funding, and general skepticism about the potential for AI systems (McCarthy, 1974).
Yet, in health care, work continued. Iconic expert systems such as MYCIN (Shortliffe, 1974) and others such as Iliad, Quick Medical Reference, and Internist-1 were developed to assist with clinical diagnosis. AI flowered commercially in the 1980s, becoming a multibillion-dollar industry advising military and commercial interests (Miller et al., 1982; Sumner, 1993). However, all ultimately failed to reach the hype and lofty promises resulting in a second AI Winter from the late 1980s until the late 2000s.2
During the AI Winter, the various schools of computer science, probability, mathematics, and AI came together to overcome the initial failures of AI. In particular, techniques from probability and signal processing, such as hidden Markov models, Bayesian networks, stochastic search, and optimization, were incorporated into AI thinking, resulting in a field known as machine learning. The field of machine learning applies the scientific method to representing, understanding, and utilizing datasets, and, as a result, practitioners are known as data scientists. Popular machine learning techniques include random forests, boosting, support vector machines, and artificial neural networks. See Hastie et al. (2001) or Murphy (2013) for thorough reviews; these methods are also discussed in more detail in Chapter 5.
Around 2010, AI began its resurgence to prominence due to the success of machine learning and data science techniques as well as significant increases in computational storage and power. These advances fueled the growth of technology titans such as Google and Amazon.
Most recently, Rosenblatt’s ideas laid the groundwork for artificial neural networks, which have come to dominate the field of machine learning thanks to the successes of Geoffrey Hinton’s group, and later others, in solving computational problems in training expressive neural networks and the ubiquity of data necessary for robust training (Halevy et al., 2009; Krizhevsky et al., 2012). The resulting systems are called deep learning systems and showed significant performance improvements over prior generations of algorithms for some use cases. It is noteworthy that Hinton was awarded the 2018 Turing Prize alongside Yoshua Bengio and Yann LeCun for their work on deep learning (Metz, 2019).
Modern AI has evolved from an interest in machines that think to ones that sense, think, and act. It is important at this early stage to distinguish narrow from
___________________
2 See Newquist (1994) for a thorough review of the birth, development, and decline of early AI.
general AI. The popular conception of AI is of a computer, hyper capable in all domains, such as was seen even decades ago in science fiction with HAL 9000 in 2001: A Space Odyssey or aboard the USS Enterprise in the Star Trek franchise. These are examples of general AIs and, for now, are wholly fictional. There is an active but niche general AI research community represented by Deepmind, Cyc, and OpenAI, among others. Narrow AI, in contrast, is an AI specialized at a single task, such as playing chess, driving a car, or operating a surgical robot. Certain narrow AIs do exist and are discussed in further detail below.
As discussed above and more in Chapter 4, the word overhyped, however, should be mentioned again. The Gartner Hype Cycle is a good place to start understanding the current state of AI in the broad stages of innovation, inflated expectations, disillusionment, enlightenment, and finally productivity (Gartner, Inc., 2018). While we grow out of the AI Winter, it is crucial that we maintain this perspective. Research publications and marketing claims should not overstate utility. Cautious optimism is crucial (Frohlich et al., 2018).
AI IN NON–HEALTH CARE INDUSTRIES
There are many industries outside of health care that are further along in their adoption of AI into their workflows. The following section highlights a partial list of those industries and discusses aspects of AI use in these industries to be emulated and avoided.
Users
It is critical to consider AI primarily in terms of its relationship with users, particularly in the health care sector. Considerable concern exists about AIs replacing humans in the workforce once they are able to perform functions that previously required a human (Kolko, 2018; Zhang et al., 2019). However, when examined critically, it is usually the case that computer scientists design AI with human users in mind, and as such, AI usually extends the capacity, capability, and performance of humans, rather than replacing them (Topol, 2019). The self-driving car, our first example below, demonstrates how an AI and human might work together to achieve a goal, which enhances the human experience (Hutson, 2017). In other examples such as legal document review, the AI working with the human reviews more documents at a higher level of precision (Xu and Wang, 2019). This concept of a human and AI team is known in the AI literature as a “centaur” (Case, 2018) and in the anthropology literature as a “cyborg” (Haraway, 2000). To date, most of the focus on the use of AI has been to support physicians. However, patients, caregivers, and allied health clinicians of all types will also be
AI users. In the next sections we examine scenarios in which AI may influence health care. Again, it is important to note that, regardless of how extensive the deployment of the AI systems described below, care must be exercised in their translation into health care.
Automotive
Of all of the industries making headlines with the use of AI, the self-driving car has most significantly captured the public’s imagination (Mervis, 2017). In concept, a self-driving car is a motor vehicle that can navigate and drive its occupants without their interaction. Whether this should be the aspirational goal (i.e., “without their interaction”) is a subject of debate. For this discussion, it is more important to note that the component technologies have been evolving publicly for some years. Navigation has evolved from humans reading paper maps to satellite-based global positioning system (GPS)-enabled navigation devices, to wireless mobile telecommunications networks that evolved from analog to increasingly broadband digital technologies (2G to 3G to 4G to 5G), and most recently, navigation systems that supplement mapping and simple navigation with real-time, crowd-sourced traffic conditions (Mostafa et al., 2018). In research contexts, ad hoc networks enable motor vehicles to communicate directly with each other about emergent situations and driving conditions (Zongjian et al., 2016).
The achievement of successful self-driving cars has been and continues to be evolutionary. In terms of supporting the act of driving itself, automatic transmissions and anti-lock braking systems were early driver-assistance technologies. More recently, we have seen the development of driver-assistance AI applications that rely on sensing mechanisms such as radar, sonar, lidar, and cameras with signal processing techniques, which enable lane departure warnings, blind spot assistance, following distance alerts, and emergency braking (Sujit, 2015).
This recent use of cameras begins to distinguish AI techniques from the prior signal processing techniques. AI processes video data from the cameras at a level of abstraction comparable to that at which a human comprehends. That is, the machine extracts objects such as humans, cyclists, road signs, other vehicles, lanes, and other relevant factors from the video data and has been programmed to identify and interpret the images in a way that is understandable to a human. This combination of computer vision with reasoning comprises a specialized AI for driving.
However, for all the laudable goals, including improving driving safety, errors remain and sometimes those errors are fatal. This is possibly the single most important lesson that the health care domain can learn from the increasing use of AI in other industries. As reported in the press, the woman killed in spring 2018
as she walked her bicycle across the street was sensed by the onboard devices, but the software incorrectly classified her as an object for which braking was unnecessary. It was also reported that the “backup” driver of this autonomous vehicle was distracted, watching video on a cell phone (Laris, 2018).
The point of the above example is not that AI (in this case a self-driving car) is evil. Rather, we need to understand AI not in isolation but as part of a human–AI “team.” Certainly, humans without any AI assistance do far worse; on average in 2017, 16 pedestrians were killed each day in traffic crashes (NHTSA, 2018). Reaching back to the 1960 quote from Licklider, it is important to note that the title of his article was “Man–Computer Symbiosis.” In this example, the driver–computer symbiosis failed. Even the conceptualization that the human was considered a backup was wrong. The human is not just an alternative to AI; the human is an integral part of the complete system. For clinicians to effectively manage this symbiosis, they must (1) understand their own weaknesses (e.g., fatigue, biases), (2) understand the limits of the sensors and analytics, and (3) be able to assist or assume complete manual control of the controlled process in time to avoid an unfortunate outcome. AI must be viewed as a team member, not an “add-on” (Johnson and Vera, 2019).
Other examples of AI outside of medicine are outlined below. The theme of symbiosis is ubiquitous.
Professional Services
Although AI is often associated with physical devices and activities, it is actually very well suited to professional activities that rely on reasoning and language. For example, x.ai offers the seemingly mundane but intricate service of coordinating professionals’ calendars. This is offered through a chat bot, which exercises not only natural language interpretation and generation but also logical reasoning to perform the scheduling.
In another domain, LawGeex and other vendors offer an AI that performs legal contract review and both discovers and appropriately escalates issues found in contract language. In addition, the AI can propose redline edits in coordination with a lawyer. Such an AI streamlines the review of standard contracts, such as nondisclosure agreements, that consume significant, expensive time with little probable value. As with many AI products, the AI enhances the human’s capabilities and effectiveness, rather than operating as an autonomous agent.
Accounting and auditing are beginning to utilize AI for repetitive task automation such as accounts receivable coding and anomaly detection in audits (Amani and Fadlalla, 2017). Once again, the reason for this application of AI is speed and accuracy, when paired with a human.
Engineers and architects have long applied technology to enhance their design, and AI is set to accelerate that trend (Noor, 2017). Unusual AI-generated structural designs are in application today. A well-known example is the partition in the Airbus A320, in which AI algorithms utilized biomimetics to design a material almost half as light and equally as strong as the previous design (Micallef, 2016).
Finance has also been an early adopter of machine learning and AI techniques. The field of quantitative analytics was born in response to the computerization of the major trading exchanges. This has not been a painless process. One of the early “automated” trading strategies, portfolio insurance, is widely believed to have either caused or exacerbated the 1987 stock market crash (Bookstaber, 2007). The failure of Long-Term Capital Management offers another cautionary example (Lowenstein, 2011). This fund pursued highly leveraged arbitrage trades, where the pricing and leverage were algorithmically determined. Unexpected events caused the fund to fail spectacularly, requiring an almost $5 billion bailout from various financial institutions. Despite these setbacks, today all major banks and a tremendous number of hedge funds pursue trading strategies that rely on systematic machine learning or AI techniques. Most visible are the high-frequency trading desks, which rely on AI technologies to place, cancel, and execute orders at a speed as minute as one-hundredth of a microsecond, far faster than a human can think, react, and act (Seth, 2019).
Media
Content recommendation, most often based on an individual’s previous choices, is the most widely visible application of AI in media. Large distribution channels such as Netflix and Amazon leverage machine learning algorithms for content recommendation to drive sales and engagement (Yu, 2019). These systems initially relied on algorithms such as collaborative filters to identify customers similar to others in terms of what media they consumed and enjoyed. More recently, deep learning techniques have been found superior for this task (Plummer, 2017). Health care examples are limited (Chen and Altman, 2015).
An interesting development has been that early techniques relied on metadata (descriptive features of media) in order to generate recommendations. More recent techniques utilize AI to generate metadata from the media itself, to personalize the presentations of recommendations, and then create recommendations. For instance, computer vision is used now to index film to identify faces, brands, and locations, which are coupled with human tags to create rich metadata (Yu, 2019).
In the music industry, startups such as Hitwizard and Hyperlive generalize these two ideas to attempt to predict which songs will be popular (Interiano et al., 2018). First, the AI structures the music into features such as loudness, beats per
minute, and key, among others. Then, it compares this structured representation of the song to others that have been successful in order to identify similarity, and hence the new song’s likelihood of also becoming a hit. The general complaint that all the music on the radio “sounds the same” may be based in part on the need to conform to the styles “approved” by the algorithms.
An emerging AI capability is generative art. Google initially released software called Deep Dream, which was able to create art in the style of famous artists, such as Vincent van Gogh (Mordvintsev et al., 2015). This technique is now used in many cell phone apps, such as Prisma Photo Editor,3 as “filters” to enhance personal photography.
Another more disturbing use of AI surfaces in the trend known as “deepfakes,” technology that enables face and voice swapping in both audio and video recordings (Chesney and Citron, 2018). The deepfake technique can be used to create videos of people saying and doing things that they never did, by swapping their faces, bodies, and other features onto videos of people who did say or do what is portrayed in the video. This initially emerged as fake celebrity pornography, but academics have demonstrated that the technique can also be used to create fake political videos (BBC News, 2017). The potential effect of such technology, when coupled with the virality of social networks, for the dissemination of false content is terrifying. Substantial funding is focused on battling deepfakes (Villasenor, 2019). An ethical, societal, and legal response to this technology has yet to emerge.
Compliance and Security
Security is well suited to the application of AI, because the domain exists to detect the rare exception, and vigilance in this regard is a key strength of all computerized algorithms. One current application of AI in security is automated license plate reading, which relies on basic computer vision (Li et al., 2018). Because license plates conform to strict standards of size, color, shape, and location, the problem is well constrained and thus suitable for early AI.
Predictive policing has captured the public imagination, potentially due to popular representations in science fiction films such as Minority Report (Perry et al., 2013). State-of-the-art predictive policing technology identifies areas and times of increased risk of crime rather than identifying the victim or perpetrator (Kim et al., 2018). Police departments typically utilize such tools as part of larger strategies. However, implementations of these technologies can propagate racial, gender, and other kinds of profiling when based on historically biased datasets (Caliskan et al., 2017; Garcia, 2016) (see Chapter 1).
___________________
3 See https://prisma-ai.com.
In the commercial sector, AI is finding increasing application in compliance. AI technologies can read e-mails, chat logs, and AI-transcribed phone calls in order to identify insider trading, theft, or other abuses. Research efforts are under way to identify theft, diversion, and abuse from all manner of dispensing devices.
Note that many of these applications are also controversial for privacy concerns and surveillance capacity and scope, in addition to their potential to propagate racial, gender, and sexual orientation biases (see Chapter 1). Large gaps remain in the goal of aligning population and cultural expectations and preferences with the regulation and legislation of privacy, which is a subject covered in more detail in Chapter 7.
Space Exploration
Space exploration is another area—an unusual and interesting one, at that—in which AI has been employed. One might provocatively claim that there is a planet in our solar system (probably) populated exclusively by robots, and that one of those robots is artificially intelligent. On Mars, NASA has sent robot rovers to explore the surface. In Curiosity, the most recent rover, NASA included a navigation and target acquisition AI called AEGIS (Autonomous Exploration for Gathering Increased Science System) (Chien and Wagstaff, 2017; Francis et al., 2017). This AI allows the rover to autonomously select rocks likely to yield successful observational studies. The necessity for AI derives from the communication latency between an Earth-based controller and the distant rover that can cause inefficiency or danger, such as in response to unexpected volcanic eruptions. The NASA Jet Propulsion Laboratory is currently designing an autonomous AI that will enable a self-sufficient probe to explore the methane subsurface oceans of Titan and Europa (Troesch et al., 2018).
AI BY INDIVIDUAL FUNCTIONS
While the concept of an “intelligent machine” has interested philosophers at least as far back as Descartes, the most popular conception was first proposed by Alan Turing (1950). Turing proposed that the question of a machine’s ability to think is not constructive, and instead one ought to test whether or not a machine can perform as well as a human in conversation. This has come to be known as the “Turing test,” which still serves as the prototype litmus test for many AI tasks. Indeed, in all AI areas—from specialized reasoning, to speech, to vision—researchers attempt to outperform humans using AI.
Where the preceding section considers the many current applications of AI, in this section we consider the components, or faculties, that comprise an AI
system or tool. The core of AI is reasoning, whether that is achieved symbolically, probabilistically, or through mathematical optimization. But before an AI system or tool can reason, a representation of the domain of reasoning must first be made. For instance, to play chess, the computer must somehow hold the current state of the board, the rules of the game, and the desired outcome in its memory. Effectively structuring or representing reality is often the key to AI. A key observation is that these representations are often layered; stated differently, an effective representation comprises a hierarchy of abstractions.
Consider chess again: the base representation is the field and players, the next layer may be particular formations of pieces, the next evolving set plays, and so on. By reasoning at higher and higher levels of abstraction, AIs can achieve effectiveness without requiring true human intelligence. George V. Neville-Neil writes:
We have had nearly 50 years of human/computer competition in the game of chess but does that mean any of those computers are intelligent? No, it does not—for two reasons. The first is that chess is not a test of intelligence; it is the test of a particular skill—of playing chess. The second reason is that thinking chess was a test of intelligence was based on a false cultural premise that brilliant chess players were brilliant minds, more gifted than those around them. Yes, many intelligent people excel at chess, but chess, or any other single skill, does not denote intelligence. (Neville-Neil, 2017)
Thus, an AI system typically receives input from sensors (afferent systems) and operates in the environment through displays/effectors (efferent systems). These capture reality, represent and reason over it, and then affect reality, respectively. The standard representation of this is a keyboard and mouse as inputs, a central processing unit (CPU) and data storage units for processing, and a monitor for output to a human user. A robot with AI may contain sonar for inputs, a CPU, and motorized wheels for its outputs. More sophisticated AIs, such as a personal assistant, may have to interpret and synthesize data from multiple sensors such as a microphone, a camera, and other data inputs in order to interact with a user through a speaker or screen. As each of the effectors, representations, and AI reasoning systems improves, it becomes more seamless, or human-like, in its capabilities.
AI Technologies
AI in the guise of “the next impossible thing computers will do,” almost by definition, occupies the forefront of technology at whatever time it is considered. As a result, AI is often conflated with its enabling technology. For instance, in the
1980s there were hardware machines, called Lisp machines, specifically created to execute the AI algorithms of the day (Phillips, 1999). Today, technologies such as graphics processing units (GPUs), the Internet of Things (IoT), and cloud computing are closely associated with AI, while not being AI in and of themselves. In this section, we briefly review and clarify.
GPUs and, recently, tensor processing units (TPUs) are computer hardware elements specialized to perform mathematical calculations rapidly. They are much like the widely understood CPUs, but rather than being generalized so that they are able to perform any operation, GPUs and TPUs are specialized to perform calculations more useful to machine learning algorithms and hence AI systems. The operations in question are linear algebra operations such as matrix multiplication. The GPUs and TPUs enable AI operations.
IoT is the movement to collect sensor data from all manner of physical devices and make them available on the Internet. Examples abound including lightbulbs, doors, cameras, and cars; theoretically anything that can be manufactured might be included. IoT is associated with AI because the data that flow from these devices comprise the afferent arm of AI systems. As IoT devices proliferate, the range of domains to which AI can be applied expands. In the emergent Internet of Medical Things, patient-generated physiological measurements (e.g., pulse oximeters and sphygmomanometers) are added to the data collected from these “environmental” devices. It will be crucial that we “understand the limitations of these technologies to avoid inappropriate reliance on them for diagnostic purposes” (Deep Blue, 2019; Freedson et al., 2012) and appreciate the social risks potentially created by “intervention-generated inequalities” (Veinot et al., 2018).
Cloud computing abstracts computation by separating the computer services from the proximate need for a physical computer. Large technology companies such as Amazon, Google, Microsoft, and others have assembled vast warehouses filled with computers. These companies sell access to their computers and, more specifically, the services they perform over the Internet, such as databases, queues, and translation. In this new landscape, a user requiring a computer service can obtain that service without owning the computer. The major advantage of this is that it relieves the user of the need to obtain and manage costly and complex infrastructure. Thus, small and relatively technologically unsophisticated users, including individuals and companies, may benefit from advanced technology. Most data now are collected in clouds, which means that the complex computing hardware (such as GPU machines) and services (such as natural language processing [NLP]) needed for AI are easily obtainable. Cloud computing also creates challenges in multinational data storage and other international law complexities, some of which are briefly discussed in Chapter 7.
Reasoning and Learning
Computers are recognized as superior machines for their rigorous logic and expansive calculation capabilities, but the point at which formal logic becomes “thinking” or “intelligence” has proven difficult to pinpoint (Turing, 1950). Expert systems that have been successful in military and industrial settings have captured the imagination of the public with the Deep Blue versus Kasparov chess matches. More recently, Google DeepMind’s AlphaGo defeated Lee Sedol at the game Go, using deep learning methods (Wikipedia, 2019a).
Adaptive Learning
A defining feature of a machine learning system is that the programmer does not instruct the computer to perform a specific task but, rather, instructs the computer how to learn a desired task from a provided dataset. Programs such as deep learning, reinforcement learning, gradient boosting, and many others comprise the set of machine learning algorithms. The programmer also provides a set of data and describes a task, such as images of cats and dogs and the task to distinguish the two. The computer then executes the machine learning algorithm upon the provided data, creating a new, derivative program specific to the task at hand. This is called training. That program, usually called the model, is then applied to a real-world problem. In a sense, then, we can say that the computer has created the model.
The machine learning process described above comprises two phases, training and application. Once learning is complete, it is assumed that the model is unchanging. However, an unchanging model is not strictly necessary. A machine learning algorithm can alternatively continue to supplement the original training data with data and performance encountered in application and then retrain itself with the augmented set. Such algorithms are called adaptive because the model adapts over time.
All static models in health care degrade in performance over time as characteristics of the environment and targets change, and this is one of the fundamental distinctions between industrial and health care processes (addressed in more detail in Chapter 6). However, adaptive learning algorithms are one of the family of methods that can adapt to this constantly changing environment, but they create special challenges for regulation, because there is no fixed artifact to certify or approve. To draw a health care analogy, the challenge would be like the U.S. Food and Drug Administration (FDA) attempting to approve a molecule that evolved over time. Although it is possible to certify that an adaptive algorithm performs to specifications at any given moment and that the algorithm by which it learns is sound, it is an open question as to whether the future states of an adaptive algorithm
can be known to perform at the same or better specification—that is, whether it can be declared safe. FDA has issued guidance that it will approve adapting devices, but significant challenges remain in this domain. For further discussion of these challenges in the regulatory and legislative context, see Chapter 7.
Reinforcement Learning
Understood best in the setting of video games, where the goal is to finish with the most points, reinforcement learning examines each step and rewards positive choices that the player makes based on the resulting proximity to a target end state. Each additional move performed affects the subsequent behavior of the automated player, known as the agent in reinforcement learning semantics. The agent may learn to avoid certain locations to prevent falls or crashes, touch tokens, or dodge arrows to maximize its score. Reinforcement learning with positive rewards and negative repercussions is how robot vacuum cleaners learn about walls, stairs, and even furniture that moves from time to time (Jonsson, 2019).
Computer Vision
Computer vision is a domain of AI that attempts to replicate the human visual apparatus (see Chapter 1). The machine should be able to segment, identify, and track objects in still and moving images. For example, some automobile camera systems continuously monitor for speed limit signs, extract that information, and display it on the dashboard. More advanced systems can identify other vehicles, pedestrians, and local geographic features. As noted above, combining similar computer vision systems with reasoning systems is necessary for the general problem of autonomous driving.
Language and Conversation
The language and conversation domain of AI can be segmented into the interpretation and generation of spoken and written words (see Chapter 1). Textual chatbots that assist humans in tasks such as purchases and queries is one active frontier. Today, spoken words are mainly encountered in the consumer realm in virtual assistants such as Alexa, Siri, Cortana, and others, such as those embedded in cars. These AI systems typically convert audio data into textual data for processing, apply NLP or natural language understanding for the task at hand, generate a textual response, and then convert that into audio. While full conversations are currently beyond the state of the art, simple intent or question- and-answer tasks are now commercially available.
Touch and Movement
Early applications of the robotics domain of AI appeared in industrial manufacturing, autopilots and autonomous vehicles, and home robots such as the Roomba, with additional research having gone into humanoid and canine-like robots. Making four-legged robots walk, run, and recover from falls, in particular, has been vexing. Building on the adaptive learning discussion above, the use of simulated data to speed robot training, which augments but does not fully replace the engineered control mechanisms, is a recent advance in robotics (Hwangbo, 2019).
Smell and Taste
Electronic noses are still marginal but increasingly useful technology (Athamneh et al., 2008). They couple chemosensors with classification systems in order to detect simple and complex smells. There is not yet significant research into an electronic tongue, although early research similar to that concerning electronic noses exists. Additionally, there is early research on computer generation of taste or digital gustation, similar to the computer generation of speech; however, no applications of this technology are apparent today.
KEY STAKEHOLDERS
It is not too strong to assert that AI is the fruit of U.S. government–funded research, carried out initially by programs such as the Defense Advanced Research Projects Agency (DARPA), which funded academic pioneers at the Massachusetts Institute of Technology (MIT), Stanford University, and Carnegie Mellon University in the 1960s. However, as the utility and impact of these technologies has accelerated, a number of other countries have made significant investments in AI. Furthermore, by providing a constant and deep market for AI technologies through energy, space exploration, and national defense, the governments of the United States, China, the European Union, and other countries have enabled AI to take root and thrive.
United States
The United States has been coordinating strategic research and development (R&D) investment in AI technologies for a number of years. In reaction to the Soviet Union’s launch of Sputnik in 1957, the U.S. government founded DARPA, which poured significant funding into computing, resulting in the early
advancement of AI as well as the Internet. Most foundational AI technology was supported through DARPA funding, beginning in the 1960s.
In 2016, DARPA’s AI R&D plan established a number of strategic categories and aims for federal investment, which included recommendations for developing an implementation framework and workforce for AI R&D. The National Institutes of Health has also articulated strategic goals for AI within its Strategic Plan for Data Science, and the U.S. Department of Health and Human Services in conjunction with the Robert Wood Johnson Foundation commissioned a report on how AI will shape the future of public health, community health, and health care delivery (JASON, 2017; NIH, 2018).
In 2018, DARPA announced “AI Next,” a $2 billion program to support the further development of AI. Additionally, under the Trump administration, the National Science and Technology Council established the Select Committee on Artificial Intelligence, which is tasked with publicizing and coordinating federal R&D efforts related to AI (White House Office of Science and Technology Policy, 2018). In February 2019, President Donald Trump issued an Executive Order 13859, “Maintaining American Leadership in Artificial Intelligence,” which charged the Select Committee on Artificial Intelligence with the generation of a report and a plan (Trump, 2019). This plan, released in June 2019, outlines the national governmental strategy for AI R&D, which includes a broad scope of seven focus areas to guide interagency collaboration, education and training programs, and directed funding programs (NSTC, 2019).
China
China is a recent but energetic participant in the AI community. In 2017, Chinese inventors filed more AI patents than any other country (World Intellectual Property Organization, 2018). In addition, although the specific amount of government funding that China has allocated for AI research is difficult to know, CBInsights (2018) estimated that China represented 48 percent of global AI research funding in 2018, dwarfing other countries’ contributions to the field. The Boston Consulting Group reached similar conclusions, noting that up to 85 percent of Chinese companies have either adopted AI processes or are running pilot initiatives to do so. The report could say the same for just approximately 50 percent of companies in the United States, France, Germany, and Switzerland and 40 percent of companies in Austria and Japan (Duranton et al., 2018).
China has a number of advantages that make AI progress more feasible. Foremost, the government actively supports companies’ pushes into AI, pursuing its stated goal of becoming the world leader in the field by 2030 (Mozur, 2017). Additionally, data are more available in China, as there are at least 700 million
Internet-connected smartphone users (Gerbert et al., 2018). Finally, in the health care market in particular, Chinese privacy laws are more lax than in the United States (Simonite, 2019).
European Union and the United Kingdom
In the past decade, Europe has developed a robust AI ecosystem, placing it well within the ranks of the United States and China. With an equal balance of corporate non-R&D and R&D entities, the region is second to the United States for most players in the space (EU Commission Joint Research Centre, 2018). However, unlike the United States and China where government funding has propelled the industry, AI in Europe stems from accelerating investment from private equity and venture capital firms (Ernst and Young, 2019).
To increase global competitiveness there has been a recent uptick of national attention and investment in AI (Ernst and Young, 2019). Several countries, namely Germany and France, have written national AI strategies, although each differs in motivation and approach. Dedicating €3 billion to AI research and development, Germany aims to expand the integration of AI in business processes. Comparatively, the French plan focuses on the potential of AI for defense and security, transportation, and health (Franke and Sartori, 2019).
Ultimately, the European Union sees the strength in advancing its AI agenda through coordination among its member states. In April 2018, 25 EU countries pledged to work together to boost “Europe’s technology and industrial capacity in AI” while “addressing socio-economic challenges and ensuring an adequate legal and ethical framework” (European Union Member States Representatives, 2018). Adhering to its commitment to ethics, the European Union released guidelines in April 2019 for the development of trustworthy AI solutions that could foreseeably shape the regulation of AI in the European Union and overseas (EU Commission, 2019).
United Kingdom
It is unsurprising that the United Kingdom leads Europe in AI. As previously discussed, Turing’s efforts in the 1950s planted the seeds for the country’s leadership in this area, which, for the most part, has been cultivated by the country’s thriving startup community (UK House of Commons, Science and Technology Committee, 2016; UK House of Lords, 2018). More recently, the government has engaged in improving the country’s AI standing through financial support and the launch of several AI initiatives aimed at exploring and preparing for the sustainable procurement of AI technology. Building on the findings of these initiatives, the United Kingdom
in 2018 unveiled its AI Sector Deal, a broad industry plan to stimulate innovation, build digital infrastructure, and develop workforce competency in data science, engineering, and mathematics (UK Government, 2019).
Academic
In the United States, MIT, Stanford, and Carnegie Mellon pioneered AI research in the 1960s, and these, and many others, continue to do so today. Cambridge University in the United Kingdom and Tsinghua University in China also produce leading AI research. Also important was the role of commercially supported research institutes, such as Nokia Bell Labs (formerly Bell Laboratories), which supported much of Claude E. Shannon’s pioneering work in digital communications and cryptography, and Xerox PARC, which continues with laboratories such as Microsoft Research and Facebook’s Building X (Shannon, 1940; Xerox, 2019). Indeed, there is significant tension between commercial facilities and academic institutions regarding talent (Reuters, 2016). Uber, for instance, at one point recruited almost the entire computer vision faculty of Carnegie Mellon, to the university’s consternation.
AI-specific conferences, such as the Conference on Neural Information Processing Systems4 (NeurIPS), the preeminent academic AI conference, attract thousands of abstract submissions annually. Furthermore, the number of AI submissions to journals and conferences that are not AI specific is increasing annually.
Commercial Sector
Programmers have always developed AI systems in order to achieve specific goals, and this inherent usefulness of the technology has frequently spawned or spun out into commercial activity. This activity has been focused in the technology sector, but a significant development during 2017 and 2018 was the focus on the digital transformation of other industries seeking to lay a foundation so that they might capitalize on the advantages of AI. Health care is one such sector.
In addition to the academic participants in the founding Dartmouth Summer Research Project on Artificial Intelligence, the sole commercial participant was Trenchard More of IBM. Since the 1956 meeting, IBM has maintained an active role in the AI community. In 2011, IBM captured the public imagination by winning Jeopardy! with its AI system Watson (Wikipedia, 2019b). Watson has evolved into a commercially available family of products and has also been deployed with variable success in the clinical setting (Freeman, 2017; Herper, 2017).
___________________
4 See https://nips.cc.
Many of the most successful technology companies, including Amazon, Facebook, Google, Microsoft, Tesla, and Uber, are deeply reliant on AI within their products and have contributed significantly to its expansion.
In addition to these more established companies, the past 5 years have witnessed something akin to a Cambrian explosion of the number of startups in the AI space. A good, if instantly outdated, reference is CBInsight’s AI 100, which lists the 100 most promising startups in the field (CBInsights, 2019).
Professional Societies
In addition, a number of civil and professional societies exist that provide leadership and policy in the AI space. These include, but are not limited to, those listed below.
IEEE (Institute of Electrical and Electronics Engineers) is the professional society for engineers, scientists, and allied professionals, including computer scientists, software developers, information technology professionals, physicists, and medical doctors. IEEE has formed several societies, such as the Signal Processing Society and Computational Intelligence Society, both of which produce publications relating to AI.
The Association for Computing Machinery (ACM) is a professional society for computing educators, researchers, and professionals. A special interest group on AI, known as SIGAI, exists within ACM.
The Electronic Frontier Foundation is a nonprofit agency concerned with digital liberties and rights. It considers AI ethics and laws and aims to protect the rights of users and creators of AI technology.
Notably in health care, the American Medical Association (AMA) passed its first policy recommendation on augmented intelligence in June 2018. The policy states that AMA will “promote [the] development of thoughtfully designed, high-quality, clinically validated health care AI” (AMA, 2018). Furthermore, the AMA Journal of Ethics dedicated its February 2019 issue to AI in health care.
Nonprofits
As a reaction to an increasing concentration of AI within the commercial sector, OpenAI was founded as a nonprofit research entity. Along the same vein as the open source movement, which maintains the general availability of software, OpenAI’s mission is to promote the accessibility of AI intellectual property to all for the purposes of developing technology that services the public good (OpenAI, 2019).
Public–Private Partnerships
There are other notable public–private partnerships that have recently formed to work in the AI sector to bridge collaboration in these spaces. For example, Partnership for AI is a large consortium of more than 90 for-profit and nonprofit institutions in multiple countries to share best practices and further research in AI, to advance public understanding of AI, and to promote socially benevolent applications in AI (Partnership on AI, 2018). Another example is AINow, which is hosted by New York University but receives funding from a variety of large for-profit and nonprofit institutions interested in AI development.
KEY CONSIDERATIONS
In summary, we would like to highlight key considerations for future AI applications and endeavors, which can be learned by examining AI’s history and evaluating AI up to the current era.
- As described in more detail in Chapters 3 and 4, history has shown that AI has gone through multiple cycles of emphasis and disillusionment in use. It is critical that all stakeholders be aware of and actively seek to educate and address public expectations and understanding of AI (and associated technologies) in order to manage hype and establish reasonable expectations, which will enable AI to be applied in effective ways that have reasonable opportunities for sustained success.
- Integration of reinforcement learning into various elements of the health care system will be critical in order to develop a robust, continuously improving health care industry and to show value for the large efforts invested in data collection.
- Support and emphasis for open source and free tools and technologies for use and application of AI will be important to reduce cost and maintain wide use of AI technologies as the domain transitions from exponential growth to a future plateau stage of use.
- The domain needs strong patient and consumer engagement and empowerment to ensure that preferences, concerns, and expectations are transmitted and ethically, morally, and appropriately addressed by AI stakeholder users.
- Large-scale development of AI technologies in industries outside of health care should be carefully examined for opportunities for incorporation of those advances within health care. Evaluation of these technologies should include consideration for whether they could effectively translate to the processes and workflows in health care.
REFERENCES
AMA (American Medical Association). 2018. AMA passes first policy recommendations on augmented intelligence. Press release. https://www.ama-assn.org/presscenter/press-releases/ama-passes-first-policy-recommendations-augmented-intelligence.
Amani, F. A., and A. M. Fadlalla. 2017. Data mining applications in accounting: A review of the literature and organizing framework. International Journal of Accounting Information Systems 24:32–58. https://doi.org/10.1016/j.accinf.2016.12.004.
Athamneh, A. I., B. W. Zoecklein, and K. Mallikarjunan. 2008. Electronic nose evaluation of cabernet sauvignon fruit maturity. Journal of Wine Research 19: 69–80.
BBC News. 2017. Fake Obama created using AI video tool. https://www.youtube.com/watch?v=AmUC4m6w1wo.
Bookstaber, R. 2007. A demon of our own design: Markets, hedge funds, and the perils of financial innovation. New York: John Wiley & Sons.
Buchanan, B. G. 2005. A (very) brief history of artificial intelligence. AI Magazine 26(4):53–60.
Caliskan, A., J. J. Bryson, and A. Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science 356:183–186.
Case, N. 2018. How to become a centaur. MIT Press. https://doi.org/10.21428/61b2215c.
CBInsights. 2018. Artificial intelligence trends to watch in 2018. https://www.cbinsights.com/research/report/artificial-intelligence-trends-2018 (accessed November 5, 2019).
CBInsights. 2019. AI100: The artificial intelligence startups redefining industries. https://www.cbinsights.com/research/artificial-intelligence-top-startups.
Chen, J. H., and R. B. Altman. 2015. Data-mining electronic medical records for clinical order recommendations: Wisdom of the crowd or tyranny of the mob? AMIA Joint Summits on Translational Science Proceedings. Pp. 435–439. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4525236.
Chesney, R., and D. K. Citron. 2018. Deep fakes: A looming challenge for privacy, democracy, and national security. California Law Review 107. https://dx.doi.org/10.2139/ssrn.3213954.
Chien, S., and K. L. Wagstaff. 2017. Robotic space exploration agents. Science Robotics 2(7). https://doi.org/10.1126/scirobotics.aan4831.
Deep Blue. 2019. IBM’s 100 icons of progress. http://www-03.ibm.com/ibm/history/ibm100/us/en/icons/deepblue.
Duranton, S., J. Erlebach, and M. Pauly. 2018. Mind the (AI) gap: Leadership makes the difference. Boston, MA: Boston Consulting Group. https://media-publications.bcg.com/france/Mind-the-(AI)-Gap-Press-deckVF.pdf.
Ernst &Young. 2019. Artificial intelligence in Europe–Germany: Outlook for 2019 and beyond. Prepared for Microsoft. https://www.ey.com/Publication/vwLUAssets/ey-artificial-intelligence-in-europe/$FILE/ey-artificial-intelligence-in-europe-germany.pdf.
EU Commission. 2019. Ethics Guidelines for Trustworthy Artificial Intelligence. https://ec.europa.eu/digital-single-market/en/news/ethics-guidelines-trustworthy-ai.
European Commission Joint Research Centre. 2018. Artificial intelligence: A European perspective. http://publications.jrc.ec.europa.eu/repository/bitstream/JRC113826/ai-flagship-report-online.pdf.
European Union Member States Representatives. 2018. EU declaration on cooperation on artificial intelligence. https://ec.europa.eu/jrc/communities/en/node/1286/document/eu-declaration-cooperation-artificial-intelligence.
Francis, R., T. Estlin, G. Doran, S. Johnstone, D. Gaines, V. Verma, M. Burl, J. Frydenvang, S. Montano, R. C. Wiens, S. Schaffer, O. Gasnault, L. DeFlores, D. Blaney, and B. Bornstein. 2017. AEGIS autonomous targeting for ChemCam on Mars Science Laboratory: Deployment and results of initial science team use. Science Robotics 2(7). https://doi.org/10.1126/scirobotics.aan4582.
Franke, U., and P. Sartori. 2019. Machine politics: Europe and the AI revolution. Policy brief. https://www.ecfr.eu/publications/summary/machine_politics_europe_and_the_ai_revolution.
Freedson, P., H. R. Bowles, R. Troiano, and W. Haskell. 2012. Assessment of physical activity using wearable monitors: Recommendations for monitor calibration and use in the field. Medicine and Science in Sports and Exercise 44:S1–S4. https://doi.org/10.1249/MSS.0b013e3182399b7e.
Freeman, D. H. 2017. A reality check for IBM’s AI ambitions. MIT Technology Review. https://www.technologyreview.com/s/607965/a-reality-check-for-ibms-ai-ambitions.
Frohlich, H., R. Balling, N. Beerenwinkel, O. Kohlbacher, S. Kumar, T. Lengauer, M. H. Maathuis, Y. Moreau, S. A. Murphy, T. M. Przytycka, M. Rebhan, H. Rost, A. Schuppert, M. Schwab, R. Spang, D. Stekhoven, J. Sun, A. Weber, D. Ziemek, and B. Zupan. 2018. From hype to reality: Data science enabling personalized medicine. BMC Medicine 16:150.
Garcia, M. 2016. Racist in the machine: The disturbing implications of algorithmic bias. World Policy Journal 33:111–117.
Gartner, Inc. 2018. 5 trends emerge in the Gartner Hype Cycle for Emerging Technologies. SmarterWithGartner. https://www.gartner.com/smarterwithgartner/5-trends-emerge-in-gartner-hype-cycle-for-emerging-technologies-2018.
Gerbert, P., M. Reeves, S. Ransbotham, D. Kiron, and M. Spira. 2018. Global competition with AI in business: How China differs. MIT Sloan Management Review. https://sloanreview.mit.edu/article/global-competition-of-ai-in-business-how-china-differs (accessed November 8, 2019).
Halevy, A., P. Norvig, and F. Pereira. 2009. The unreasonable effectiveness of data. IEEE Intelligent Systems 24:8–12. https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/35179.pdf.
Haraway, D. 2000. A cyborg manifesto. In The Cybercultures Reader, edited by D. Bell and B. Kennedy. London, UK, and New York: Routledge.
Hastie, T., R. Tibshirani, and J. Friedman. 2001. The elements of statistical learning: Data mining, inference, and prediction. New York: Springer-Verlag.
Herper, M. 2017. MD Anderson benches IBM Watson in setback for artificial intelligence in medicine. Forbes. https://www.forbes.com/sites/matthewherper/2017/02/19/md-anderson-benches-ibm-watson-in-setback-for-artificial-intelligence-in-medicine/#54ada2e83774.
Hutson, M. 2017. A matter of trust. Science 358:1375–1377. https://doi.org/10.1126/science.358.6369.1375.
Hwangbo, J., J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsounis, V. Koltun, and M. Hutter. 2019. Learning agile and dynamic motor skills for legged robots. Science Robotics 4:13.
Interiano, M., K. Kazemi, L. Wang, J. Yang, Z. Yu, and N. L. Komarova. 2018. Musical trends and predictability of success in contemporary songs in and out of the top charts. Royal Society Open Science 5:171274.
JASON. 2017. Artificial intelligence for health and health care. https://www.healthit.gov/sites/default/files/jsr-17-task-002_aiforhealthandhealthcare12122017.pdf.
Johnson, M., and A. H. Vera. 2019. No AI is an island: The case for teaming intelligence. AI Magazine 40:17–27.
Jonsson, A. 2019. Deep reinforcement learning in medicine. Kidney Disease (Basel) 5(1):18–22.
Kim, S., P. Joshi, P. S. Kalsi, and P. Taheri. 2018. Crime analysis through machine learning. 9th Annual Information Technology, Electronics and Mobile Communication Conference, Vancouver.
Kline, R. R. 2011. Cybernetics, automata studies, and the Dartmouth Conference on Artificial Intelligence. IEEE Annals of the History of Computing 33(4):5–16.
Kolko, J. 2018. 5 questions we should be asking about automation and jobs. Harvard Business Review. https://hbr.org/2018/12/5-questions-we-should-be-asking-about-automation-and-jobs (accessed November 5, 2019).
Krizhevsky, A., I. Sutskever, and G. E. Hinton. 2012. ImageNet classification with deep convolutional neural networks. Presented at 25th International Conference on Neural Information Processing Systems. https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf (accessed November 8, 2019).
Laris, M. 2018. Raising the bar: Uber details shortcomings in self-driving car that killed pedestrian. The Washington Post. https://www.washingtonpost.com/local/trafficandcommuting/raising-the-bar-uber-details-shortcomings-in-self-driving-car-that-killed-pedestrian/2018/11/01/5152bc54-dd42-11e8-b3f0-62607289efee_story.html.
Li, H., P. Wang, M. You, and C. Shenb. 2018. Reading car license plates using deep neural networks. Image and Vision Computing 27:14–23.
Licklider, J. C. R. 1960. Man-computer symbiosis. IRE Transactions on Human Factors in Electronics 1:4–11.
Lighthill, J. 1973. Artificial intelligence: A general survey. In Artificial Intelligence: A Paper Symposium. Science Research Council, UK.
Lowenstein, R. 2011. When genius failed. New York: Random House.
McCarthy, J. 1958. Programs with common sense. Paper presented at Symposium on Mechanization of Thought Processes at National Physical Laboratory, Teddington, UK.
McCarthy, J. C. 1959. Programs with common sense. http://www-formal.stanford.edu/jmc/mcc59/mcc59.html.
McCarthy, J. 1974. Review of “Artificial Intelligence: A General Survey,” by Professor Sir James Lighthill. Artificial Intelligence 5:317–322.
McCarthy, J., M. L. Minsky, N. Rochester, and C. E. Shannon. 2006. A proposal for the Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955. AI Magazine 27(4):12–14.
Mervis, J. 2017. Not so fast. Science 358:1370–1374.
Metz, C. 2019. Turing award won by 3 pioneers in artificial intelligence. The New York Times. https://www.nytimes.com/2019/03/27/technology/turing-award-ai.html.
Micallef, K. 2016. Airbus generates bionic design for flights of the future. Redshift. https://www-p7.autodesk.com/redshift/bionic-design.
Miller, R. A., H. E. Pople, and J. D. Myers. 1982. Internist-1, an experimental computer-based diagnostic consultant for general internal medicine. New England Journal of Medicine 307:468–476.
Moor, J. 2006. The Dartmouth College Artificial Intelligence Conference: The next fifty years. AI Magazine 27(4):87–91.
Mordvintsev, A., M. Tyka, and C. Olah. 2015. Deepdream. GitHub. https://github.com/google/deepdream.
Mostafa, A. N., C. Pranamesh, S. Anuj, B. G. Stephen, and H. Mingyi. 2018. Evaluating the reliability, coverage, and added value of crowdsourced traffic incident reports from Waze. Transportation Research Record: Journal of the Transportation Research Board 2672:34–43.
Mozur, P. 2017. Beijing wants A. I. to be made in China by 2030. The New York Times. https://www.nytimes.com/2017/07/20/business/china-artificial-intelligence.html (accessed November 8, 2019).
Murphy, K. P. 2013. Machine learning: A probabilistic perspective. Cambridge, MA: MIT Press.
Neville-Neil, G. V. 2017. The chess player who couldn’t pass the salt. Communications of the ACM 60(4):24–25. https://doi.org/10.1145/3055277.
Newquist, H. P. 1994. The brain makers: The history of artificial intelligence. Upper Saddle River, NJ: Prentice Hall.
NHTSA (National Highway Traffic Safety Administration). 2018. 2017 fatal motor vehicle crashes: Overview. Traffic Safety Facts Research Note DOT HS 812 603. Washington, DC: U.S. Department of Transportation. https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/812603 (accessed November 8, 2019).
NIH (National Institutes of Health). 2018. NIH strategic plan for data science. https://datascience.nih.gov/strategicplan (accessed December 7, 2019).
Noor, A. K. 2017. AI and the future of machine design. Mechanical Engineering 139(10):38–43.
NRC (National Research Council). 1966. Language and machines: Computers in translation and linguistics: A report. Washington, DC: National Academy Press. https://doi.org/10.17226/9547.
NSTC (National Science and Technology Council). 2019. The national artificial intelligence research and development strategic plan: 2019 update. https://www.nitrd.gov/pubs/National-AI-RD-Strategy-2019.pdf (accessed December 13, 2019).
OpenAI. 2019. OpenAI. https://openai.com.
Partnership on AI. 2018. Meet the partners. https://www.partnershiponai.org/partners.
Perry, W. L., B. McInnis, C. Price, S. Smith, and J. S. Hollywood. 2013. Predictive policing: The role of crime forecasting in law enforcement operations. Santa Monica, CA: RAND Corporation. https://www.rand.org/content/dam/rand/pubs/research_reports/RR200/RR233/RAND_RR233.pdf.
Phillips, E. M. 1999. If it works, it’s not AI: A commercial look at artificial intelligence startups. M. S. thesis, Massachusetts Institute of Technology.
Plummer, L. 2017. This is how Netflix’s top-secret recommendation system works. Wired. https://www.wired.co.uk/article/how-do-netflixs-algorithms-work-machine-learning-helps-to-predict-what-viewers-will-like.
Reuters. 2016. One year after announcing pact, the Uber-Carnegie Mellon Partnership is stalled. Fortune. http://fortune.com/2016/03/21/uber-carnegie-mellon-partnership.
Rosenblatt, F. 1958. The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review 65(6):386–408.
Seth, S. 2019. The world of high-frequency algorithmic trading. Investopedia. https://www.investopedia.com/articles/investing/091615/world-high-frequency-algorithmic-trading.asp.
Shannon, C. E. 1940. A symbolic analysis of relay and switching circuits. Cambridge, MA: Massachusetts Institute of Technology.
Shortliffe, E. H. 1974. MYCIN: A rule-based computer program for advising physicians regarding antimicrobial therapy selection. Palo Alto, CA: Stanford Artificial Intelligence Laboratory, Stanford University.
Simonite, T. 2019. How health care data and lax rules help China prosper in AI. Wired. https://www.wired.com/story/health-care-data-lax-rules-help-china-prosper-ai.
Sujit, S. B. 2015. Effective test strategy for testing automotive software. International Conference on Industrial Instrumentation and Control, Prune, India. https://www.semanticscholar.org/paper/Effective-test-strategy-for-testing-automotive-Barhate/aeff9109e6214cd4586af032e968c5daabf945d2.
Sumner, W. I. 1993. Review of Iliad and Quick Medical Reference for primary care providers. Archives of Family Medicine 2:87–95.
Topol, E. 2019. Deep medicine: How artificial intelligence can make healthcare human again. New York: Basic Books.
Troesch, M., S. Chien, Y. Chao, J. Farrara, J. Girton, and J. Dunlap. 2018. Autonomous control of marine floats in the presence of dynamic, uncertain ocean currents. Robotics and Autonomous Systems 108:100–114.
Trump, D. 2019. Executive Order 13859: Maintaining American leadership in artificial intelligence. Federal Register 84(3967):3967–3972. https://www.federalregister.gov/documents/2019/02/14/2019-02544/maintaining-american-leadership-in-artificial-intelligence.
Turing, A. M. 1950. Computing machinery and intelligence. Mind 49:433–460.
UK Government. 2019. AI sector deal. Policy paper. https://www.gov.uk/government/publications/artificial-intelligence-sector-deal/ai-sector-deal.
UK House of Commons, Science and Technology Committee. 2016. Robotics and artificial intelligence. Fifth Report of Session 2016–17. https://publications.parliament.uk/pa/cm201617/cmselect/cmsctech/145/145.pdf.
UK House of Lords, Select Committee on Artificial Intelligence. 2018. AI in the UK: Ready, willing and able? Report of Session 2017–19. https://publications.parliament.uk/pa/ld201719/ldselect/ldai/100/100.pdf.
Veinot, T. C., H. Mitchell, and J. S. Ancker. 2018. Good intentions are not enough: How informatics interventions can worsen inequality. Journal of the American Medical Informatics Association 25:1080–1088. https://doi.org/10.1093/jamia/ocy052.
Villasenor, J. 2019. Artificial intelligence, deepfakes, and the uncertain future of truth. TechTank blog, Brookings Institution. https://www.brookings.edu/blog/techtank/2019/02/14/artificial-intelligence-deepfakes-and-the-uncertain-future-of-truth.
Weiner, N. 1948. Cybernetics: Or control and communication in the animal and the machine. Cambridge, MA: MIT Press.
White House Office of Science and Technology Policy. 2018. Summary of the 2018 White House summit on artificial intelligence for American industry. https://www.whitehouse.gov/wp-content/uploads/2018/05/Summary-Report-of-White-House-AI-Summit.pdf.
Wikipedia. 2019a. AlphaGo versus Lee Sedol. https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol.
Wikipedia. 2019b. Watson (computer). https://en.wikipedia.org/wiki/Watson(computer).
World Intellectual Property Organization. 2018. World intellectual property indicators 2018. https://www.wipo.int/edocs/pubdocs/en/wipo_pub_941_2018.pdf.
Xerox. 2019. Xerox and PARC define the era of artificial intelligence. https://www.xerox.com/en-kn/insights/artificial-intelligence-today.
Xu, N., and K. Wang. 2019. Adopting robot lawyer? The extending artificial intelligence robot lawyer technology acceptance model for legal industry by an exploratory study. Journal of Management & Organization 1–19. https://doi.org/10.1017/jmo.2018.81.
Yu, A. 2019. How Netflix uses AI, data science, and machine learning—From a product perspective. Medium. https://becominghuman.ai/how-netflix-uses-ai-and-machine-learning-a087614630fe.
Zhang, D. D., G. Peng, and Y. O. Yao. 2019. Artificial intelligence or intelligence augmentation? Unravelling the debate through an industry-level analysis. SSRN. https://dx.doi.org/10.2139/ssrn.3315946.
Zongjian, H., C. Jiannong, and L. Xuefeng. 2016. SDVN: Enabling rapid network innovation for heterogeneous vehicular communication. IEEE Network 30:10–15.
Suggested citation for Chapter 2: Fackler, J., and E. Jackson. 2020. Overview of current artificial intelligence. In Artificial intelligence in health care: The hope, the hype, the promise, the peril. Washington, DC: National Academy of Medicine.
This page intentionally left blank.
