Reproduce “Class-Balanced Loss Based on Effective Number of Samples” 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
Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, Serge Belongie
The paper introduces class-balanced loss weighting for long-tailed image classification, scaling each example’s loss by the inverse effective sample count of its class. This count follows a simplified model of diminishing coverage and is controlled by β; the weighting can be added to standard losses. On iNaturalist 2018, ResNet-50 with class-balanced focal loss has lower validation top-1 error than softmax cross-entropy at 224×224 input, limited to that comparison.
Key insights
- The value of additional training samples exhibits diminishing returns due to data overlap and redundancy, rather than a linear increase in information.
- Modeling data overlap through a simplified random covering problem provides a principled, theoretically grounded basis for re-weighting imbalanced data.
- The hyperparameter "beta" controls the degree of diminishing returns, reflecting the inherent diversity or redundancy within a class, and can provide insights into dataset characteristics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.