Towards Label-Noise Resistant Learning via Optimal Brain Damage Masking
Researchers have developed a new method called Selective Edge Masking (SEM) to improve the robustness of deep neural networks against noisy labels. By leveraging Optimal Brain Damage theory, SEM adaptively masks less critical connections within the classifier layer, effectively intercepting and suppressing noisy gradients. This plug-and-play mechanism can be integrated with existing noise-robust techniques and has demonstrated superior performance on various benchmarks. AI
IMPACT Introduces a novel technique to enhance model reliability in real-world data scenarios, potentially improving performance in applications with imperfect labeling.