Active diversification of head-class features in bilateral-expert models for enhanced tail-class optimization in long-tailed classification

Training deep learning models on long-tailed datasets is a challenging task since the classification performance of tail classes with fewer samples is always unsatisfactory. Currently, many long-tailed methods have achieved success. However, some methods always improve tail-class performance at the expense of head class performance due to limited model capability. To address this issue, we propose a novel algorithm-level method inspired by information theory to balance the information space of each class and boost tail-class performance while minimizing head-class sacrifice. Our method involves actively eliminating the redundant feature information of head classes to save space for tail classes during training. Specifically, we use a bilateral expert model and design a duplicate information disentanglement (DID) module that can extract duplicate and redundant information from bilateral-expert features. This allows us to develop a head diversity loss to decrease the extracted duplicate and redundant information of head classes and a tail distillation loss to increase the label information of tail classes. The joint result of these two losses allows our model to fully leverage the information space for improved tail-class performance without compromising head-class performance. The effectiveness and practicability of our method are verified by five datasets with long-tailed distributions for visual recognition or fault diagnosis tasks. Experimental results demonstrate that our method outperforms currently available mainstream methods, which we attribute to the effectiveness of our proposed DID module and the incorporation of two long-tailed losses.