Machine learning is changing everyday activities through improvements in functions such as voice recognition, object detection, and image perception. The spheres of medicine, transportation, finance, government, and education could be transformed by such innovative technology, as could the way individuals and communities live and work. For example, some tasks typically performed by humans are already being performed by computer systems, and the range and complexity of tasks these systems can perform is growing. This creates novel tensions arising from the legal, social, and ethical implications of machine learning for humans now and in future generations.
To explore these issues in greater depth, the Raymond and Beverly Sackler U.S.-U.K. Scientific Forum “The Frontiers of Machine Learning” was held on January 31 and February 1, 2017, at the Washington, D.C., headquarters of the National Academies of Sciences, Engineering, and Medicine. A planning committee of distinguished scholars from the United States and the United Kingdom organized the forum.
The forum drew over 60 in-person attendees and over 500 webcast participants from academia, government, industry, and philanthropy. Participants included industry leaders, machine learning researchers, and experts in privacy and the law. They represented strong proponents of widespread adoption of machine learning as well as those with concern for the societal tensions that arise with expanding the use of machine learning. This 2-day meeting included presentations and discussions on the current state of the art in machine learning, its relationship
to other fields, challenges in the field, and key considerations for its future development.
This publication provides a summary of the observations shared by forum participants. It does not reflect a consensus of the participants or the views of the sponsoring organizations. Instead, it is intended to provide perspectives and suggestions from a selection of individuals working at the frontier of machine learning and studying its implications. It examines the evolution of the field and its applications, some of the social and ethical tensions to which advances in machine learning have contributed, and considerations for the future development of the field.
The terms “artificial intelligence,” “machine learning,” and “robotics” are often used interchangeably, but important distinctions exist.
While specific definitions vary, artificial intelligence is, generally speaking, any method for programming computers to enable them to carry out tasks or behaviors that would require intelligence if performed by humans. Early work in this field focused on automated reasoning; using these approaches, programs were written as sets of logical statements against which queries were processed by theorem proving and search. For example, computers could be programmed with the rules for board games and then tasked with finding sequences of moves that could defeat an opponent. To obtain the logical rules (or inferences) for such systems, researchers in the field first turned to interviewing experts in the relevant domains (e.g., medical diagnosis, fault diagnosis). When the need to reason under uncertainty became increasingly evident, especially in medicine and in engineering applications, the field embraced probabilistic models as a replacement for systems of logical rules. This in turn required developing methods for probabilistic reasoning. However, in many applications, it was very difficult to build probabilistic models manually and determine their parameters. The field then turned to machine learning and related methods in statistics and pattern recognition to help address this challenge. Instead of interviewing human experts and seeking to codify the steps they take to solve a problem, these systems examine data—that is, examples of how to solve the problem at hand—to find patterns and rules that can drive decision making, such as clustering and classification.
Machine learning draws from a variety of fields, including computer science, statistics, engineering, cognitive science, and neuroscience. Researchers in machine learning develop both the mathematical foundations and the practical applications of systems that learn from data.
Broadly speaking, machine learning can be divided into three
branches: supervised, unsupervised, and reinforcement learning. “Supervised learning” is based on a program that learns from previous examples: a model with specified parameters receives input and makes a prediction based on function estimation. This prediction is then compared to the actual outcome, and the similarity or difference between the two then informs the updates to the model parameters. For example, a supervised learning model can use photos with information about what is shown in those photos to then recognize and identify features in new photos. “Unsupervised learning” can be used where there is a dearth of labeled data on which to train algorithms. An unsupervised machine learning system processes a large amount of data, finding patterns in the data and learning the characteristics that make data points more or less similar to each other. “Reinforcement learning” emphasizes learning by trial and error, using rewards or punishments for success or failure at a task as a means to discover successful ways of behaving in complex environments.
The field of machine learning continues to advance at a rapid pace as a result of improved computational resources, better algorithms and tools, increased availability of data to train systems, and the ingenuity of those developing machine learning systems. There are countless applications of machine learning in society that help people become more organized and make processes more efficient—for example, in organizing digital photo collections, managing spam in email platforms, and supporting navigation devices. Machine learning can solve problems related to classification, regression, clustering, dimensionality reduction, semi-supervised learning, and reinforcement learning. These are often referred to as the “canonical” problems in machine learning and are described in Box 1.
The term “artificial intelligence” can often evoke ideas of human-like intelligence in machines, known as “artificial general intelligence.” The goal of artificial general intelligence research is to create artificial intelligence systems that have the ability to adapt to a range of tasks and could develop a breadth of capabilities that would match or exceed those of humans. Such systems are different from current artificial intelligence systems that are designed to solve specific target tasks but lack any knowledge or capability beyond those tasks. Some scholars have argued that once systems achieve human-level breadth and capability, they could start improving themselves and therefore rapidly exceed human capabilities. Most researchers maintain,
however, that there is no technical basis for such concerns and that matching human capabilities is not a critical threshold in the development of artificial intelligence. These researchers worry that a focus on developing systems with human-level intelligence detracts from the need to address near-term questions about how to create artificial intelligence systems that behave correctly, safely, and reliably.
Machine learning is a quickly evolving field with numerous application areas. One of the earliest applications of the technology, “information retrieval” involves “finding material (usually documents) of an unstructured nature (usually text) that satisfies an information need from within large collections (usually stored on computers).”1 Machine learning provides algorithms and models useful for modern information retrieval. Users benefit from this technology each time they search their e-mail or browse the Web.
“Speech recognition” is the process by which a computer system recognizes spoken words and phrases and translates them into machine-readable format. Though the first speech recognition system was developed in the 1950s, these computer systems could translate only a small number of speech signals into words and phrases. Newer software in modern speech recognition has advanced to the point of understanding and using natural speech patterns (e.g., speed, tone, word choice) that are characteristic of normal conversation. Using layers of neural networks,2 known as deep learning3 approaches, the sound waves from speech are converted into accurate plain text. However, it still remains difficult to distinguish one voice from another. Developing computer systems that can understand the meaning of the plain text and make appropriate responses is still challenging.
“Targeted advertising,” which became popular in the 1990s, seeks to present advertisements to consumers that increase the likelihood that the consumer will make a certain purchase or perform a certain action. To do this, the companies collect data on their consumers’ behaviors or purchase histories and apply machine learning methods to develop predictive models of future customer behavior.
1 C.D. Manning, P. Raghavan, and H. Schütze, 2008, Introduction to Information Retrieval, Cambridge University Press, Cambridge, U.K.
2 Neural networks are “computing system[s] made up of a number of simple, highly interconnected processing elements, which process information by their dynamic state response to external inputs.” (M. Caudill, 1989, Neural networks primer: Part I, AI Expert 2(12): 46-52.) The outputs are equivalent to the sum of the inputs multiplied by their weights. More specifically, the inputs are applied, the actual versus desired output is calculated, the error is propagated backward into the network, and weights are adjusted to minimize the error. Neural networks are trained repeatedly, with inputs adjusted each time, to get closer to a desired output. It remains difficult to predict what will happen with these networks, especially with unknown or untrained input, so they must continually be trained with new samples—simply changing the code is insufficient. Over time the development of more complex, large-scale, multilayered neural networks has allowed researchers to work on problems at scale.
In finance, machine learning is used for assessing risk and managing assets. Machine learning has also found applications in bioinformatics where complex biological data (e.g., genetic code) are analyzed using either supervised or unsupervised learning to better understand the ways in which cells function.
Automated vehicles can independently navigate some public roadways by sensing their environment and responding appropriately. Such vehicles learn where to steer and how to locate other vehicles, people, and objects based on a range of machine learning techniques, including using deep neural networks to analyze image and video data. Advances in (1) data handling and communications bandwidth, (2) computing power, (3) learning algorithms, and (4) efficient inference algorithms help make automated vehicles more reliable, and advances in machine learning theory could improve the technology even further. Important challenges remain, such as identifying better metrics to evaluate the severity of driving errors (e.g., causing a head-on collision versus running a red light), improving three-dimensional perception, and functioning under poor weather or lighting conditions.
Automated vehicles have the potential to reduce vehicle fatalities, increase mobility for the elderly and those with disabilities, decrease environmental pollution, and support new models of public transportation. However, questions remain about the likelihood of these
3 Deep learning is a powerful class of machine learning methods that explore data representations using supervised, semi-supervised, or unsupervised learning. An essential characteristic of deep learning is that these methods are not just one function; rather, they are made up of a set of composable subfunctions for model building. Deep learning involves neural network models with algorithmic innovation, larger data sets, larger compute resources, and better software tools, and it is compatible with many variants of machine learning. Deep learning has a number of similarities to the human brain: artificial neural networks act like neurons in the brain to connect various input concepts with potential outputs. E-mail spam filters, Internet search engines, speech recognition devices, translators, photo recognition programs, automatic email repliers, object detection in automated vehicles, and navigation apps all rely on deep learning to simplify tasks for the user. Much modern deep learning demands building differential architectures that can do everything. Although deep learning allows more intricate processing than more basic machine learning approaches, the method is data hungry, compute intensive, deficient in representing uncertainty, easily deceived by oppositional examples, difficult to optimize, and opaque. In order to preserve interpretability and practicality when using deep learning processes, it is important to stress test these models in new environments.
benefits being fully realized. For example, could this technology result in more cars on the road, lengthen commutes, or lead to urban sprawl? Alternatively, might fewer people choose to own cars, leading to increased carpooling and a reduced number of cars on the road? If this is the case, fewer parking lots would be needed across communities, and cities could find new ways to use the space. Any significant change in car ownership also has potential economic impacts; the automotive sector is a large component of many developed economies, and the effects of such disruption are difficult to predict.
Automated vehicle deployment could also affect pedestrians and cyclists, as it changes the commonly accepted—if unwritten—rules of the road. For example, how can a pedestrian make eye contact with a driver before crossing a street when there is no driver in an automated vehicle? Much of the communication among road users and many of the predictions humans make about the behavior of other road users when navigating a vehicle-populated environment is not explicit. Rather, it relies on reading subtle observational cues combined with past experience and knowledge of the environmental norms (e.g., driving behavior in the United States is different from that in India). These questions open interesting avenues of investigation, and researchers are developing models to better understand how humans interact with intelligent devices in potentially dangerous situations. This raises one facet of a broader question about vehicle safety standards: should new standards be developed, or do current road safety standards translate for automated vehicles?
Machine learning is also finding new applications in data science itself. There are not yet enough data scientists, statisticians, and machine learning experts to dedicate adequate time to understanding the data, building models, and making predictions. One possible solution to this shortage of experts is the Automatic Statistician project, which is creating systems that are able to automate data analysis.4 As interest in data science grows, and the sophistication of these tools increases, such projects could help support the use of machine learning in a broad range of research fields or novel applications.