Keynotes | K&Ts | GACs | Talks | Posters | Search

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

Learning to select computations in recurrent circuits

Sixing Chen¹, Frederick Callaway², Sreejan Kumar³, Joni D. Wallis⁴, Erin L. Rich⁵, Marcelo G Mattar¹; ¹New York University, ²Princeton University, ³Columbia University, ⁴University of California, Berkeley, ⁵Mount Sinai School of Medicine

Presenter: Sixing Chen

To cope with their limited computational resources, humans solve problems by adaptively deciding what to compute, a capacity known as meta-reasoning. Yet, determining which computations to perform is itself computationally demanding, raising the question of how such selection can be implemented in the brain. Here, we develop a recurrent neural network model that learns to select computations through meta-reinforcement learning. In a simple choice task, the agent learns to select computations as predicted by optimal symbolic models and reproduces neural dynamics observed in macaque orbitofrontal cortex. In a more complex planning task, the agent learns a planning strategy that replicates key signatures of human strategies. Our work provides a generalizable framework for modeling the adaptive control of thought processes and elucidates how such control can be implemented in neural systems.

Topic Area: Decision-Making, Cognitive Control & Event Cognition