Reproduce “The Forgetting-Retention Dilemma: Certified Unlearning Theory in 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
Yiting Hu, Lingjie Duan, Qian Zhang
The authors develop certified task-level unlearning for sequential, ℓ2-regularized continual learning. Their methods use Hessian corrections and calibrated Gaussian noise; a hybrid combines curvature corrections with natural forgetting to reduce storage. On asynchronous CIFAR-100, the hybrid uses 2,232.40 MB peak GPU memory and has 0.0463 approximation error, versus 5,678.90 MB and 0.0087 for Gauss–Newton Hessian; this evidence is limited to the studied setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.