Poster Presentation

Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms

Flexible inference of history-dependent learning rules in de novo learning

Yuhan Helena Liu¹, Victor Geadah¹, Jonathan W. Pillow¹; ¹Princeton University

Presenter: Yuhan Helena Liu

Understanding how animals learn new tasks from scratch is a central challenge in cognitive neuroscience. Existing models typically impose fixed parametric forms on learning rules, which may miss richer dynamics used by animals. More flexible approaches exist, but are largely limited to bandit settings rather than learning new input-output mappings (i.e., *de novo* learning). Here, we infer learning rules directly from behavioral data during *de novo* task learning using a hybrid modeling framework that separates the decision policy from the learning rule. We model the decision policy with a generalized linear model (GLM), providing an interpretable parameterization of choice behavior, while parameterizing trial-by-trial updates of the GLM weights with a neural network. We introduce TinyRNNGLM, a recurrent extension with only two hidden units, whose low dimensionality enables direct interpretation of the latent state. It outperforms a feedforward, memoryless variant by capturing cross-trial dependencies, and also outperforms a low-dimensional RNN policy model that predicts choice directly, highlighting the benefit of separating the decision policy from the learning rule. In mice, the latent state encodes recent reward history, with larger updates following rewarded sequences. Overall, these results indicate that hybrid models support both flexible inference and interpretability in *de novo* learning, while a low-dimensional recurrent extension reveals reward-history-dependent structure in the learning process.

Topic Area: Methods, Tools, Theory & Neural Coding