This paper introduces Selective Edge Masking (SEM), an approach based on Optimal Brain Damage (OBD) theory, to enhance deep neural network robustness against noisy labels by masking redundant connections in the classifier layer, effectively suppressing the backpropagation of noise
Noisy labels are a major challenge in real-world deep learning, causing significant performance degradation due to the backpropagation of incorrect gradients. Existing noise-robust methods have largely focused on loss functions and sample selection. This research shifts focus to dynamic architectural adaptation by rethinking model connectivity. The authors leverage Optimal Brain Damage (OBD) theory, which suggests removing parameters causing negligible loss perturbation, to identify and mask redundant connections within the classifier. This process, termed Selective Edge Masking (SEM), ensures that only critical information paths remain, thereby maintaining the network's fitting capacity while mitigating the influence of label noise during training. Evaluations show that the OBD-driven approach consistently surpasses state-of-the-art methods.