Schedule of Events | Symposia

Sketchpad Series

Latent structure inference supports behavioral flexibility

Poster Session A - Saturday, March 29, 2025, 3:00 – 5:00 pm EDT, Back Bay Ballroom/Republic Ballroom

Tiffany A. Morton1, Daniel Kimmel1, Daphna Shohamy1,2,3; 1Zuckerman Mind Brain Behavior Institute, Columbia University, 2Kavli Institute for Brain Science, Columbia University, 3Columbia University, NY, NY

For flexible behavior, agents must adapt their responses to stimuli based on features of the environment, which are often collectively referred to as “context”. Yet, it remains unclear how one forms a mental representation of context that is abstracted away from its observable features. We hypothesized that, early in learning, agents represent conjunctions of observable features to support context-dependent behavior, but, with experience, gradually transition to an abstract representation of context. To investigate this, we developed a serial reversal learning task where the set of action and outcome contingencies for several stimuli changed in successive blocks of trials. Each set of contingencies corresponded to a latent context that subjects could exploit to infer optimal actions after a block change. Critically, throughout the course of learning, we systematically varied the number of features required to define the latent context, such that the benefit of a context-based representation increased with conjunction size. Preliminary results show that subjects utilize the latent structure to infer the optimal response. Early in learning, performance was greater on trials where the latent context was resolvable with fewer features (i.e., a simple conjunction). However, with experience, performance on trials requiring complex conjunctions improved. These results are consistent with our hypothesis that the representation of latent structure transitioned from the conjunction of observable features to an abstracted concept of context. These findings contribute to a deeper understanding of how abstracted representations of context are constructed and may inform predictions for the underlying neural representations.

Topic Area: EXECUTIVE PROCESSES: Goal maintenance & switching

CNS Account Login

CNS2025-Logo_FNL_HZ-150_REV

March 29–April 1  |  2025

Latest from Twitter