CTN: Monday Lab Kim Stachenfeld

Title: Discovering Symbolic Cognitive Models from Human and Animal Behavior with CogFunSearch

Abstract: A key goal of cognitive science is to discover mathematical models that describe how the brain implements cognitive processes. These models often take the form of short computer programs, and constructing them typically requires a great deal of human effort and ingenuity. In this meeting, I’ll share current results from our recent efforts to apply FunSearch [Romera-Paredes et al 2024] to the problem of discovering  programs that reproduce the behavior of humans or other animals performing simple tasks. FunSearch is a recently developed tool that uses Large Language Models (LLMs) in an evolutionary algorithm to discover programs optimized for some objective. For our investigation, we consider datasets from three species performing a classic reward-learning task that has been the focus of a great deal of modeling effort. Our approach reliably discovers models that outperform state-of-the-art cognitive models for each dataset. The discovered programs can readily be interpreted as computational cognitive models, instantiating human-interpretable hypotheses about the learning and decision-making algorithms used by the brain. This is work that we’re wrapping up at DeepMind for ICML and prepping for journal submission, so it’s a great time to get questions, comments, feedback, criticisms, and suggestions for new opportunities! We are also hoping to apply the approach to new tasks/datasets soon, and I’d love to get ideas.