Reports of the AAAI 2019 Spring Symposium Series

Symposium Reports

Ioana Baldini, Clark Barrett, Antonio Chella, Carlos Cinelli, David Gamez, Leilani H. Gilpin, Knut Hinkelmann, Dylan Holmes, Takashi Kido, Murat Kocaoglu, William F. Lawless, Alessio Lomuscio, Jamie C. Macbeth, Andreas Martin, Ranjeev Mittu, Evan Patterson, Donald Sofge, Prasad Tadepalli, Keiki Takadama, Shomir Wilson

 The AAAI 2019 Spring Symposi- AI, Autonomous um Series was held Monday through Wednesday, March 25–27, 2019, on Machines, and Human Awareness the campus of Stanford University, Applications of machine learning combined with AI algo- adjacent to Palo Alto, California. The titles of the nine symposia were rithms have propelled unprecedented economic disruptions Artificial Intelligence, Autonomous across diverse fields in industry, military, medicine, finance, Machines, and Human Awareness: User and others. With the forecast for even larger impacts, the Interventions, Intuition and Mutually present economic impact of machine learning is estimated Constructed Context; Beyond Curve in the trillions of dollars. But as autonomous machines Fitting — Causation, Counterfactuals become ubiquitous, recent problems have surfaced. Early on, and Imagination-Based AI; Combin- and again in 2018, Judea Pearl warned AI scientists they ing Machine Learning with Knowledge must “build machines that make sense of what goes on in Engineering; Interpretable AI for Well- their environment,” a warning still unheeded that may Being: Understanding Cognitive Bias impede future development. For example, self-driving vehi- and Social Embeddedness; Privacy- Enhancing Artificial Intelligence and cles often rely on sparse data; self-driving cars have already Language Technologies; Story-Enabled been involved in fatalities, including a pedestrian; and yet Intelligence; Toward Artificial Intelli- machine learning is unable to explain the contexts within gence for Collaborative Open Science; which it operates. Toward Conscious AI Systems; and Verification of Neural Networks.

We propose that these seemingly unrelated prob- to many tasks, including computer vision, natural lems require an interdisciplinary approach to address language processing, and game playing. However, Pearl’s warning. At our symposium, for example, despite all this progress, there is a growing segment papers were presented by AI computer scientists, of the scientific community that questions whether engineers, social scientists, lawyers, physicians, entre- these successes can be extrapolated to create general preneurs, philosophers, and others, who addressed how AI without a major retooling. The goal of this sympo- user interventions may explain the mutual context sium was to bring together researchers across multiple for autonomous machines operating in unfamiliar disciplines (computer science, cognitive science, environments or when experiencing unanticipated economics, medicine, statistics) to discuss the capa- events; how autonomous machines can be taught to bilities of current AI and machine-learning technolo- explain shared contexts by reasoning, inferences or gies and the integration of causal and counterfactual causality and decisions by humans relying on intui- reasoning into the data-driven sciences to help alle- tion; and how human-machine teams may interde- viate their shortcomings. pendently affect human awareness, other teams, and The symposium featured a keynote talk by Judea society and how these teams may be affected in turn. Pearl, professor of computer science and statistics at In short, can context can be mutually constructed and the University of California, Los Angeles, and Turing shared between machines and humans to enhance Award winner for his work on probabilistic and causal performance? For example, in the Uber accident that reasoning. Pearl’s talk focused on the foundations and killed a pedestrian in 2018, the car, which detected types of causal inference in terms of a three-layer the pedestrian 5 seconds before the human driver did, causal hierarchy that sharply distinguishes (1) asso- was a poor team player that did not alert its human- ciations, (2) interventions, and (3) counterfactuals. operator teammate when it easily could have. These theoretical foundations unveil how several By extension, we remain interested in whether shared important problems found throughout society are context follows when machines begin to develop sub- beyond the reach of the current generation of jective states, somewhat like humans’, that allow both machine-learning systems but that can be solved to monitor and report on their joint interpretations of with the tools of causal inference. reality, forcing scientists to rethink the general model The symposium was organized into sessions fol- of human social behavior and thus elevating the value lowing the format “causality + x,” where x is a select of an interdisciplinary approach. If dependence on AI area that included (1) computer vision and imagi- and machine learning continues to grow, we and the nation, (2) machine learning and AI, (3) the social public are also interested in what happens to context sciences and economics, and (4) the health sciences. shared by human-machine teams or society when The speakers were asked to discuss the present and these teams malfunction. As we think through this future of their fields, including the recent advances change in human terms, our ultimate goal is for AI to in causal inference that changed their areas (in terms advance the performance of human-machine teams of both methodology and practice) as well as the for the betterment of society wherever these teams most pressing issues that causal inference tools may interact with other human or machine outsiders. be able to help with in the next few years. After completing most of our invited and regular Specifically, the session on computer vision and presentations over the first two days, on the third day imagination included topics ranging from the construc- of our symposium, we had an extended joint session tion of causal variables using unsupervised learning to with the Privacy-Enhancing Artificial Intelligence and leveraging causal models for enabling unsupervised Language Technologies symposium. In this joint ses- learning techniques (such as GANs) to sample from sion, we discussed how to apply privacy to teams — for interventional distributions. The speakers for the example, to the extent possible, the sharing of context session were Frederick Eberhardt (Caltech), Murat among teammates must remain private to enhance Kocaoglu (MIT-IBM Watson AI Lab), and Mohammed trust and the further sharing of private information Elhoseiny (KAUST). within a team. That is, what teammates share should The machine learning and AI session featured a not be disclosed outside of the team context unless discussion on how causal and counterfactual reason- rules or laws have been violated. ing guides human cognition and decision making, William F. Lawless, Ranjeev Mittu, and Donald Sofge its relationship with reinforcement learning, and served as cochairs of this symposium and wrote initial explorations of how it can be related to deep this report. learning. The speakers for the session were Tobias Gerstenberg (Stanford University), Thomas Dietterich (Oregon State Un