Towards robust long-tailed recognition: A class-balanced loss based on example forgetting
Hewei Yu, Jiayi Chen · Mathematical Foundations of Computing · 2025
Although deep learning has achieved remarkable success on balanced datasets, its performance often degrades significantly when applied to real-world imbalanced datasets. Among the various approaches to address long-tailed recognition, cost-sensitive reweighting methods have emerged as a widely adopted solution. In this work, we demonstrate that there exists substantial information overlap among samples, and samples that are 'forgotten' more frequently during training tend to carry more unique information, thereby contributing more significantly to the overall information volume. Building on this insight, we propose a novel Class-Balanced Loss Based on Example Forgetting (FCB loss), which integrates hard example mining and example forgetting into the computation of the total information volume. This mechanism is then utilized to make a weight factor within the loss function. The proposed FCB loss not only effectively adjusts the contribution of classes with varying attributes, but also exhibits robustness to hyperparameter variations. Extensive experiments demonstrate that FCB loss significantly enhances the accuracy of long-tailed recognition tasks without the need for extensive hyperparameter tuning. And, the more prominent the long-tailed phenomenon is, the better the results are. Additionally, we introduce a novel resampling strategy that, when combined with FCB loss, ensures balanced class contributions while enhancing training efficiency and stability.