(A) study on hippocampal successor representation for transfer learning

Abstract

One of the limitations of reinforcement learning (RL) algorithms is poor task generalizability. On the other hand, humans have the propensity to generalize environmental representations. This study aims to design a human-like generalizable RL algorithm using successor representation (SR), a computational model forming the human predictive map. We propose a novel method to quantify the invariance of the SR and show that it achieves environmental transformation invariance. Second, we implement an SR-Transformer model for task transfer, which best uses the SR's invariance. The proposed model outperforms baseline models on a zero-shot navigation task. We also demonstrate our model's generalizability on an image-based spatial navigation task. Critically, our model can explain various biological phenomena in memory-related brain areas, including the entorhinal grid and hippocampal place cells.