Reproduce “Active Learning Through a Covering Lens” 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
Ofer Yehuda, Avihu Dekel, Guy Hacohen, Daphna Weinshall
ProbCover, an active learning strategy from The Hebrew University of Jerusalem, reformulates sample selection through a "covering lens" to address the "cold start" problem in deep learning. It consistently outperforms existing methods in very low-budget annotation scenarios and demonstrates that active learning effectively enhances semi-supervised models by leveraging self-supervised embeddings.
Key insights
- A theoretical "duality" exists in active learning: optimal strategies depend on budget scale, with maximizing probabilistic coverage for low budgets and minimizing covering radius for high budgets.
- High-quality, semantically rich embeddings from self-supervised learning are crucial for guiding effective active learning strategies, especially when model-dependent uncertainty is unreliable.
- Active learning, specifically ProbCover, is not made redundant by self-supervised learning; instead, it provides a synergistic complement that significantly boosts the performance of semi-supervised models with minimal labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.