Reproduce “Continual Unsupervised Representation 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
Dushyant Rao, Francesco Visin, Andrei A. Rusu, Yee Whye Teh, Razvan Pascanu, Raia Hadsell
The authors introduce CURL, an unsupervised continual-learning system that infers data groupings, adds mixture components when existing ones model examples poorly, and replays generated examples from earlier model snapshots. On sequential MNIST, expansion at threshold −200 yielded 77% cluster accuracy, versus 51% without expansion at the same 25-component capacity. Its best 77.74% result required snapshots aligned with class duration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.