Reproduce “Hard Negative Mining for Metric Learning Based Zero-Shot Classification” 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
Maxime Bucher, Stéphane Herbin, Frédéric Jurie
Researchers from CentraleSupélec and Université Paris-Saclay developed hard negative mining strategies to improve metric learning for Zero-Shot Classification. Their Uncertainty/Correlation-based sampling method achieves 81.2% accuracy on the AwA dataset, outperforming random negative sampling and existing state-of-the-art methods while also exhibiting faster convergence.
Key insights
- Carefully selecting informative "hard" negative examples significantly improves Zero-Shot Classification performance over random sampling.
- Considering both model uncertainty and intra-class attribute correlation when selecting negatives leads to more effective learning and faster convergence.
- Increasing the ratio of intelligently selected negative to positive pairs during training generally enhances accuracy in metric learning for ZSC.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.