View in #questions-forum on Slack
@Kaustubh_Olpadkar: Is EWC memory-based method like GEM?!
Context: I was reading GEM paper. The following lines from paper indicate that.
On the other hand, memory-based methods such as EWC and GEM lead to higher ACC as the number of passes through the data increases. However, GEM suffers less negative BWT than EWC, leading to a higher ACC.
@Arthur_Douillard: I recall that only GEM (for its gradient constraint loss) used rehearsal data. Not EWC
@andcos: I agree with @Arthur_Douillard. EWC only needs to store additional values for each adaptive parameter, not previous patterns.
@Kaustubh_Olpadkar: Any idea if this official implementation from GradientEpisodicMemory team is correct? They are using the memory for ewc.