Long-Tailed Classification with Fused Latent Features and Contrast Learning
Hongli Yuan, Jian-wei Liu · 2024
In the real world, datasets usually show a long-tailed distribution. This presents a challenge to learning a robust feature extractor. Therefore, many recent studies have introduced contrast learning to learn better representations, but direct transfer does not adapt well to long-tail scenarios, and traditional contrast learning methods do not make full use of the information between features in a batch. To solve this, we modeled class condition distribution using Gaussian Mixture distribution with class Prototypes and Batch data(GM-PB}, which is more suitable for long tail scenario and can boost the model performance. The generalization of GM-PB is proved by deducing the upper bound of the generalization error. Then we propose a feature aggregation method for KNN graphs constructed using logits. In this way, model can capture the relation-level information between features, and get more robust features of tail category. We did intensive experiment on common long-tailed datasets, and the results show that our proposed model is superior to recent methods.