Improving imbalance classification via ensemble learning based on two-stage learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks require large, high-quality datasets for optimal image classification performance.
• Real-world datasets often exhibit biased distributions, leading to performance degradation due to prior and covariate shifts.
• Existing two-stage learning methods struggle with preserving network representational ability and addressing covariate shift.

Purpose of the Study:

• To develop a robust method for handling imbalanced datasets in deep neural networks.
• To improve model performance by addressing both prior and covariate shifts.
• To enhance the representational ability of networks trained on imbalanced data.

Main Methods:

• Proposed a sample logit-aware reweighting (SLA) method to adjust weights for majority and minority class samples.
• Integrated SLA with logit adjustment for a stable two-stage learning strategy.
• Developed a multi-domain expert specialization model, inspired by ensemble learning, to tackle covariate shift.

Main Results:

• The proposed method effectively repairs weights for hard samples in majority classes and samples in minority classes.
• The combined approach of SLA and logit adjustment forms a stable two-stage learning strategy.
• The multi-domain expert model improved decision-making by averaging expert classifications across different domains.
• Experimental results on CIFAR-LT and ImageNet-LT datasets demonstrated superior performance compared to state-of-the-art methods.

Conclusions:

• The developed two-stage learning method successfully addresses prior and covariate shifts in imbalanced datasets.
• The sample logit-aware reweighting and multi-domain expert specialization offer significant improvements in deep learning for image classification.
• The proposed model achieves excellent performance on challenging imbalanced datasets like CIFAR-LT and ImageNet-LT.