Reproduce “Fine-Grained Analysis of Catastrophic Forgetting in Continual Learning via Algorithmic Stability” 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
Guodong Zheng, Yan Sun, Peng Wang, Enneng Yang, Jiafei Lyu, Xiaochun Cao, Dacheng Tao, Li Shen
The authors analyze how SGD schedules affect forgetting in continual learning by tracking how changes to training examples propagate across updates and how gradients from earlier and later tasks align. Their bounds compare constant, task-resetting decay, and continuous decay. In Split CIFAR-100 experiments, continuous decay yields near-zero forgetting in tested settings, but also the lowest average accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.