Replication of Experience Replay for Continual Learning

Ifrah Idrees, Thao Nguyen, Eric Rosen

Abstract: In this work, we investigate the results reported in the paper “Experience Replay for Continual Learning” [Rolnick et al., 2018] through a replication study of the CLEAR method. More specifically, we review the contributions of this paper and report a detailed specification of our replication process for sequential task learning and implementing the CLEAR method for experience replay. Through our investigation, we found that the authors did not release a codebase of their contributions, and only the baseline, IMPALA [Espeholt et al., 2018], was available. We also found that the CLEAR method required many modifications to the baseline code with vague details not well described in the paper. To this end, we include our solutions to fixing the IMPALA codebase and how we adapted it for the CLEAR method. Lastly, we describe our attempts to implement the CLEAR method with and without behavioral cloning (Figure 2 in the original paper) and report our replicated graph of IMPALA for sequential task learning.

Track: Replicability

NeurIPS Paper Id: https://openreview.net/forum?id=SyeCmVSl8r