Class-Specific Branch Attention for Mitigating Gradient Interference under Class Imbalance
Researchers have developed a new method called Class-Specific Branch Attention (CSBA) to address performance degradation in deep neural networks caused by class imbalance. This technique identifies and mitigates gradient interference between different classes during training, where majority class gradients can suppress minority class learning. CSBA achieves this by enabling branch-specific channel reweighting, promoting feature decoupling without altering architectural simplicity. The method has shown significant improvements, such as increasing the F1 score for a minority class from 0.261 to 0.522 under severe imbalance. AI
IMPACT Improves model robustness on imbalanced datasets, crucial for real-world applications like medical imaging and fraud detection.