Reproduce “Unlearning with Asymmetric Sources: Improved Unlearning-Utility Trade-off with Public Data” 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

Ahmed Mehdi Inane, Vincent Quirion, Gintare Karolina Dziugaite, Ioannis Mitliagkas

ALU characterizes public data in Langevin unlearning with private and public training data. It retains public examples during noisy projected-gradient fine-tuning on retained private data, bringing the starting model closer in distribution to retraining without forgotten data. Under stated assumptions, the initial Rényi-divergence bound scales with squared forget-set size over squared combined dataset size; utility depends on distribution alignment.

Summary by alphaXiv.

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