Semantic-Aware Logit Adjustment for Long-Tailed Recognition

Training deep neural networks on imbalanced data is difficult as conventional training methods exhibit bias towards the head classes. Existing methods address long-tailed bias issue through frequency correction or representation learning, but they do not quantify class-level feature overlap and optimize the loss accordingly. In this work, we introduce Semantic-Aware Logit Adjustment (SALA), a training loss which integrates the concept of frequency priors together with semantic crowdedness penalty based on the similarities of the classifiers’ prototypes. Our method not only considers the rarity of each class, but also the degree of overlap of the representations between the classes to optimize the most structurally crowded classes. Extensive experiments performed on standard benchmarks like ImageNet-LT, CIFAR-10-LT, CIFAR-100-LT and iNaturalist 2018 show highly competitive performance, particularly excelling in rare-class accuracy.