Reproduce “Why Forget-Only Unlearning Needs Memorization” 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
Luka Radić, Vikrant Singhal, Amartya Sanyal
The authors show forget-only unlearning cannot always reproduce retraining when training discards information needed to identify the retraining target. For canonical threshold learning, the fitted model carries log q bits, while supporting m selected deletions requires at least m[log q − βq log(q−1) − h(βq)] bits. This bound concerns a constructed distribution and specified guarantees; a learner retaining more information can avoid the obstruction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.