CS Special Seminar: Stratis Tsirtsis "Behavioral machine learning"

Behavioral machine learning

Stratis Tsirtsis
Max Planck Institute for Software Systems
Google Scholar

Abstract: Artificial intelligence systems based on machine learning have achieved remarkable success in a variety of tasks, from image recognition to text generation. However, as these systems are increasingly used to assist humans in high-stakes decisions, expectations increase—we expect them to consider our social norms, account for how their decisions affect and are affected by humans, and provide explanations for their reasoning. In other words, we expect them to behave like humans. In this talk, I will share my work and perspective on behavioral machine learning, a field that bridges machine learning with insights from the social and behavioral sciences. Along the way, I will dive deeper into two projects. First, I will introduce a game-theoretic model in which a decision maker (e.g., a bank) uses a predictive model to make decisions under transparency, while a population of individuals (e.g., loan applicants) strategically adapt their features in response. In this context, I will present a method to find a decision policy that incentivizes individuals to invest effort in actions that lead to self-improvement. Second, I will introduce a method based on causal modeling that enables large language models to perform counterfactual reasoning about their own outputs—assessing how a specific output sequence would have changed if, in retrospect, certain parts of the output had been different. I will conclude with a discussion of exciting directions for future work in behavioral machine learning with applications in online platforms, healthcare, and cognitive science.

Bio: Stratis is a Ph.D. candidate at the Max Planck Institute for Software Systems (MPI-SWS). His research bridges machine learning with economics and cognitive science, focusing on ways artificial intelligence can enhance and assist high-stakes decision making under uncertainty. His interests center on developing novel methodologies that leverage causality, game theory, and optimization to create machine learning systems capable of accounting for and emulating human behavior. Beyond his work at MPI-SWS, Stratis has conducted research at Stanford University's Department of Psychology and Meta AI.