Reproduce “Parameter-Level Soft-Masking 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
Tatsuya Konishi, Mori Kurokawa, Chihiro Ono, Zixuan Ke, Gyuhak Kim, Bing Liu
SPG addresses task-incremental learning, where old training data are discarded and task IDs remain available, by scaling shared-network updates according to parameter importance for earlier tasks. On ImageNet-100 split into 100 tasks, it reaches 58.1±0.5% final average accuracy, compared with WSN’s 51.9±0.4%. This does not eliminate forgetting: backward transfer is −1.2%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.