Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Dynamics of neural representations that support generalization under continual learning

Daniel L. Kimmel¹, Kim Stachenfeld¹, Nikolaus Kriegeskorte¹, Stefano Fusi¹, C. Daniel Salzman, Daphna Shohamy¹; ¹Columbia University

Presenter: Daniel L. Kimmel

Abstraction and generalization are essential for flexible decision-making in novel situations. Recent work in humans and monkeys has shown how abstract variables are encoded by the representational geometry of single-neuron population activity. However, these observations are typically made after learning has converged, leaving open the question of how these representations form. To address this question, we developed a factorized model of temporal abstraction that builds on the successor representation. The model disentangles the contributions of different levels of abstract learning—from stimulus-response associations to generalizable task schema—in the form of a *factorized prediction error* that relates the change in relational knowledge to a predicted change in representational geometry on each trial. We fit the model to the behavior of human participants performing a context-dependent decision task during fMRI. The model captured the learning dynamics at multiple timescales, including the increasing contribution of generalization as participants transferred abstracted relational knowledge between novel task instances. In fMRI, BOLD activity in orbitofrontal cortex, hippocampus, and amygdala correlated more with learning attributed to generalization than to the other levels of abstraction. Moreover, the relative dominance of generalization over the other levels increased across task instances in entorhinal cortex, as well as orbitofrontal cortex and hippocampus. Finally, individual variation in the generalization neural signal correlated with behavioral performance on key trials that required relational knowledge. Our findings align with recent proposals for how the brain generalizes abstracted knowledge to current task-relevant states. Our approach offers a computational framework for probing the dynamics of representational geometry under continual abstract learning.

Topic Area: Predictive Processing & Cognitive Control

Extended Abstract: Full Text PDF