Spring Distinguished Speaker Series Fosters Conversations on the Future of AI with Leaders in the Field
This past quarter, the Distinguished Speaker Series brought in data science Professors Margo Seltzer, Tom Griffiths, and Moshe Vardi to campus for insights from the frontiers of the field. Sponsored jointly by the University of Chicago’s Data Science Institute, the Department of Statistics, the Department of Computer Science, and the Committee on Computational and Applied Mathematics, the series hosts leading researchers to share theoretical advances and practical challenges in the space and spur dialogues on campus. Read on for highlights from this spring’s talks, on the metascience of data science, AI’s limitations, and how to think about tradeoffs between efficiency vs. resilience in AI.
Margo Seltzer
For the opening session of Spring Quarter, Professor Margo Seltzer, the Canada 150 Research Chair in Computer Systems and the Cheriton Family Chair in Computer Science at the University of British Columbia, examined how traditional software engineering approaches fall short when it comes to addressing the needs of data science because it was designed for entirely different purposes.
Professor Seltzer, who co-founded and served as CTO of Sleepycat Software and has developed several widely-used software packages, detailed the generations of approaches to handling data provenance and reproducibility since 1997. Among them were efforts she and her team have worked on, such as Reproducibility as a Service (RaaS). But while there has ostensibly been a commitment to improved transparency and reproducibility, Professor Seltzer’s research found most publicly-posted scripts are not, in fact, reproducible. Professor Seltzer emphasized that future strategies must go beyond signaling to outcomes, stressing that “the artifact,” or proof of compliance “is not the goal; the goal is insights and results.”
Tom Griffiths
Next, Tom Griffiths, the Henry R. Luce Professor of Information Technology, Consciousness and Culture in the Departments of Psychology and Computer Science at Princeton University, and a co-author of “Algorithms to Live By” challenged conventional predictions about the capabilities of artificial intelligence.
In AI, he said, we’ve expected something superhuman when in fact, AI will be “heterogeneously superintelligent, outperforming humans in a wide range of settings while systematically underperforming in others.” How, then, do we understand this “jagged frontier?” Professor Griffiths discussed how cognitive science can provide a framework for understanding these performance patterns, including cases of “predictable failures” by large language models. For example: AI is better at counting when the number is 30 as opposed to 29. Why? Because 30 appears more frequently on the internet.
Professor Griffiths also discussed how model design might differ depending on whether a given task demands more or less reasoning, such as facial recognition which relies on more holistic impressions. Professor Griffiths concluded by emphasizing that specific kinds of AI models are not necessarily better than humans but rather different: “We evolved under certain constraints to solve certain computational problems. They evolved under others and are consequently different.”
Moshe Vardi
The final lecture of the quarter brought George Distinguished Service Professor in Computational Engineering at Rice University Moshe Vardi to campus. Drawing on events from recent history—including the supply chain disruptions of the early pandemic—Professor Vardi examined how both industry and society have prioritized efficiency above all else—leaving us hugely vulnerable to unforeseen disruptions.
In the ML space, “greedy algorithms” tend to get stuck in local optima (the best case for a specific scenario) and require a systemic intervention to spur them to a broader, more global perspective. Professor Vardi argued for a paradigm shift from prioritizing efficiency to emphasizing resilience, and made a case for societal intervention to promote this. As precedent, he pointed to automobile design elements like bumpers which became universal not due to aesthetic preferences, but because they were found to improve safety and therefore required by regulation. Possible tech security approaches included incorporating accountability by requiring certification as in civil engineering, and incentivizing some degree of redundance to allow more resilience than leaner “just in time” approaches.
Thank you to our speakers and to everyone who attended these talks throughout the year. These sessions created opportunities to exchange ideas and opened up discussion about the future of AI to the broader UChicago community.
The lineup for next year’s Distinguished speaker series will be announced in the fall. The Data Science Institute will be moving into its new home at 5460 South University Avenue this summer and will host next year’s series there. For more information about the Distinguished Speaker Series and future events, visit the Data Science Institute website.
