Research talk: Towards efficient generalization in continual RL using episodic memory

Speaker: Mandana Samiei, PhD Student, McGill University and Mila (Quebec AI Institute) Reinforcement learning (RL) is a powerful, brain-inspired framework to train agents for making sequential decisions in artificial intelligence. In this talk, the researchers consider two scenarios wherein RL can be challenging. The first is when non-stationarity plays an important role in the environment, and the second is when data and compute available to the agent are limited. We then discuss mitigation principles inspired by the brain’s capacity for episodic memory, that is, the subjective memory of specific previous events. However, the classical implementation of episodic memory in RL is computationally inefficient for storing and retrieving information. Besides that, simple episodic memories do not show good generalization to novel tasks. Despite the recent progress made by episodic memory in RL on the speed of learning, efficient generalization remains an open area for future explorations. The researchers pro