Analysis of Weight Balancing on Long-Tailed Recognition Problem

Abstract

Long-tailed recognition, where the per-class sample size is highly skewed, has recently gained in importance and is challenging because the accuracy of data belonging to classes with a few samples deteriorates in naive training. Two-stage training with weight decay and weight clipping has been proposed to improve the accuracy of long-tailed data. However, this method requires tuning many hyperparameters in the second stage of training, and why it is effective for long-tailed data is unknown. We analyzed the algorithm and found that it can be decomposed into the increase in the FDR of the feature extractor by weight decay and logit adjustment by weight decay and weight clipping. On the basis of this analysis, we propose a training algorithm without the second stage that results in both improved accuracy and simplification.