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

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Neural implementation of temporal difference error via meta reinforcement learning

Christopher M Kim¹, Carson C Chow²; ¹Howard University, ²NIH

Presenter: Christopher M Kim

Dopamine response has been thought to encode the temporal difference error in reinforcement learning. Recent experimental studies captured the key aspects of temporal difference error in the dopamine response, leading to an evidence for a hard-wired circuit computing the temporal difference error in the brain. However, a neural network model for computing the temporal difference error within the framework of reinforcement learning has not been developed yet. Here, we develop such model by learning to produce the temporal difference error and using the approximate error for reinforcement learning. Our model performance was comparable to the standard models. Analysis of trained networks predicts that the correlation between reward and value increases with layer depth. In sum, our work provides a reinforcement learning framework to generate dopamine response in a network, opening the door to develop realistic models to understand dopamine activity.

Topic Area: Methods, Tools, Theory & Neural Coding