3 |
The Future of Machine Learning |
UNDERSTANDING THE NEEDS OF DIFFERENT USERS AND USE CASES
As previously discussed, different domains and applications require machine learning systems with different characteristics. For example, the justice system is associated with different pressures than the transportation system. Understanding these different needs and use cases will be important in ensuring the field is able to realize its full potential.
ENGAGEMENT FOR STAKEHOLDERS
As the technical abilities of machine learning systems increase and their societal implications become more pronounced, it will be important for a diverse community of users and developers to help shape both technology development and policy discussions. To push forward the boundaries of machine learning, engineers, data scientists, and computer scientists can benefit from the following perspectives:
- Sociologists could play important roles, including discussing the ethical and societal issues of the technology;
- Psychologists and human factors experts could offer insight on the ways in which humans interact with technology;
- Unions and industrial psychologists could represent the voices of workforce changes that may result from continued development in artificial intelligence;
- Policy makers and regulators could contribute to a dialogue about autonomy and fair information principles; and
- Historians could contribute knowledge about past eras of rapid technological change and inform today’s debate.
This interdisciplinary approach should also be reflected in professional training for artificial intelligence and in policy making for the field.
ALGORITHM DEVELOPMENT AND MODEL CONSTRUCTION
Advances in the technical abilities of machine learning algorithms have contributed to the recent success of the field. Algorithms are “encoded procedures for solving a problem by transforming input data into a desired output, based on specified calculations and procedures.”15 They perform specific tasks, and it is important to consider how they will be used and what will be the explicit trade-offs involved in each use case.
Many applications in public policy highlight these trade-offs. For example, when seeking fairness in a model being used in criminal justice and determining acceptable rates of false positives or false negatives, is it worse to incarcerate a low-risk individual (false positive) than to release from prison a high-risk individual (false negative)? If the algorithms are kept “fair” by reducing false negatives and excluding certain predictive criteria based on legally protected personal characteristics, more high-risk individuals may be released or more low-risk individuals may be incarcerated.
In some cases, such as criminal justice, fairness may become even more important than accuracy. It is important that algorithms are developed and conveyed such that the trade-offs are transparent. This will help increase confidence not just in the algorithm but also in the decision-making process. Context is also important in developing the algorithm, if such data can be collected. It may be useful to encode social norms into the machine learning and optimization processes so as to focus more on fairness.
The broader social consequences of the ubiquitous use of algorithms and increasing personalization are also important to consider. This personalization could lead to “algorithmic bubbles” that classify people based on their usual behaviors, beliefs, and actions and preferentially provide them with information consistent with these. This could insulate currently held beliefs from challenge and could divide societies into segments that cease to communicate.
___________________
15 T. Gillespie, 2014, “The Relevance of Algorithms,” pp. 167-193 in Media Technologies: Essays on Communication, Materiality, and Society, eds. T. Gillespie, P.J. Boczkowski, and K.A. Foot, MIT Press, Cambridge, Mass.
GOVERNANCE AND PUBLIC CONFIDENCE
Given the wide range of approaches to machine learning and possible application areas, public confidence in these systems will be key to their continued success.
The term “governance” in this context refers to a diverse set of instruments and behaviors that shape how machine learning is used: it encompasses codes of practice, institutional norms, and individual behaviors, as well as specific policies and direct regulation by the government.
There are already efforts in place for the government to better understand, offer support, and set standards related to artificial intelligence. For example, algorithms for certain tasks already have to follow existing laws, such as being nondiscriminatory. It is important to note that governance is not synonymous with constraint; instead, it can give the confidence and public acceptance that enables new applications to be developed.
While some have called for further governance of machine learning, both the feasibility and desirability of governance approaches based on targeted regulation of machine learning algorithms have been questioned in different settings. “Artificial Intelligence and Life in 2030” suggests, “Attempts to regulate artificial intelligence in general would be misguided since there is no clear definition of artificial intelligence (it is not any one thing), and the risks and considerations are very different in different domains. Instead, policy makers should recognize that to varying degrees and over time, various industries will need distinct, appropriate, regulations that touch on software built using artificial intelligence or incorporating artificial intelligence in some way.”16 The development of best practices by sector may be more useful in steering the development of the field.
This leads to questions about the governance of the technology itself—and those working on its applications. For example, should data collection be self-regulated for devices such as digital assistants? Because these data are considered unregulated, there are no fair reporting laws that currently apply in the United States. In the European Union, the General Data Protection Regulation, which is expected to take effect in May 2018, will address “automated individual decision making, including profiling” and notes that a person “shall have the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning him or her or similarly significantly affects him or her,” and that, if such processing is necessary, the person will have “at least the right to obtain human intervention [...] to express his or her point of view and to contest the decision.”17 Furthermore, the regulation requires that if profiling is used, the person should have access to “meaningful information about the logic involved.”18
___________________
16 Panel on the One Hundred Year Study on Artificial Intelligence, “Artificial Intelligence and Life in 2030.”
17 DSB-MIT-SYSTEM, “Article 22 EU GDPR: Automated Individual Decision-making, Including Profiling,” http://www.privacy-regulation.eu/en/22.htm, accessed June 16, 2017.
OPEN DISTRIBUTION OF TOOLS AND KNOWLEDGE
It is important to ensure that engineers, designers, business leaders, and inventors have the appropriate knowledge and skills required to make the most out of new technologies. Sharing knowledge—for example, through educational materials and tools—is essential, as is improving communication across industry, academia, and government. Research from different sectors is often complementary, and providing access to data and real examples is crucial to ensure that learning and best practices spread quickly. Continued funding to support academic research and education in the field of machine learning and related areas can also help move the field forward.
Innovation requires access to data, codes, and other resources. Open sourcing of tools and online learning and encouraging researchers to share their findings publicly can spur increased innovation.
MACHINE LEARNING IN THE FUTURE
As machine learning transforms the world in which we live and improves many aspects of our lives, the legal, social, and ethical implications must be considered. The diverse applications of this technology will likely continue to expand and mature. As they do, collaboration, education, awareness of possible adverse implications of this technology, and creativity in addressing problems are crucial to ensure that the pressing challenges of today are mitigated to allow continued growth in the future.
___________________
18 DSB-MIT-SYSTEM, “Article 13 EU GDPR: Information to Be Provided Where Personal Data are Collected from the Data Subject,” https://www.privacy-regulation.eu/en/13.htm, accessed June 16, 2017.