Reproduce “Long-Tailed Recognition via Information-Preservable Two-Stage 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
Fudong Lin, Xu Yuan
Researchers at the University of Delaware introduced "Information-Preservable Two-Stage Learning," a framework for long-tailed classification that achieved state-of-the-art overall accuracies, such as 51.7% on ImageNet-LT and 74.0% on iNaturalist 2018, by combining Balanced Negative Sampling for robust feature representation and Information-Preservable Determinantal Point Process for effective data balancing.
Key insights
- Maximizing mutual information within same-label representation spaces through balanced negative sampling (BNS) is theoretically equivalent to minimizing intra-class distance, leading to superior feature separation.
- A Determinantal Point Process (DPP) can be effectively formulated using prediction probabilities to prioritize sampling of challenging, informative instances for decision boundary rectification.
- Modifying the standard DPP to enforce a fixed cardinality for sampled subsets (IP-DPP) enables controlled balancing of highly imbalanced datasets while preserving crucial information from majority classes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.