Reproduce “Learning without Overwriting: Structural Knowledge Fusion for Continual Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Kanishk Bakshi, Amitansu Pattanaik, Kathiravan Srinivasan

KFN learns sequential image-classification tasks by training a separate specialist for each incoming task, then integrating its representation into a growing shared model with replay-anchored fusion. On ten-task Split CIFAR-100, its worst final task accuracy was 11.94%, versus 2.44% for Replay, and average forgetting was 57.92%, versus 73.44%. These exploratory single-run results do not isolate structural benefits from added capacity.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.