Fisher-Shaped EWC (FS-EWC) improves parameter protection in continual learning by minimizing the worst-case empirical Fisher Information Matrix trace within a neighborhood of future updates. Unlike standard EWC, it uses trajectory-based model merging to favor solutions less sensitive to parameter displacement.
HOW THIS AFFECTS YOU
●
researcherYou can utilize this method to mitigate catastrophic forgetting in models without requiring expensive data replay buffers.