Dual Compensation Residual Networks for Class Imbalanced Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models struggle with class-imbalanced datasets, leading to poor generalization.
• Existing re-balancing methods often cause overfitting on minority (tail) classes and underfitting on majority (head) classes.

Purpose of the Study:

• To propose Dual Compensation Residual Networks (DCRN) for improved performance on class-imbalanced data.
• To address the challenges of fitting both tail and head classes effectively.

Main Methods:

• Introduced dual Feature Compensation Module (FCM) and Logit Compensation Module (LCM) to combat overfitting by correcting feature drift in tail classes.
• Developed a Residual Balanced Multi-Proxies Classifier (RBMC) to alleviate underfitting by enhancing head class learning through uniform learning with a residual path.

Main Results:

• The proposed FCM and LCM modules effectively estimate and compensate for feature drift in tail classes.
• The RBMC classifier improves learning of head classes by utilizing a residual learning path.
• Experiments on Long-tailed and Class-Incremental benchmarks demonstrate the method's efficacy.

Conclusions:

• Dual Compensation Residual Networks offer a robust solution for learning generalizable representations and classifiers on imbalanced datasets.
• The DCRN approach successfully balances performance across both head and tail classes, outperforming existing methods.