New SA-SAM method improves deep neural network robustness at high sparsity

Researchers have developed Sparsity-Adaptive Sharpness-Aware Minimization (SA-SAM), a new method to improve the robustness of deep neural networks against common corruptions, especially at high sparsity levels. SA-SAM adjusts the perturbation radius based on sparsity to maintain robustness, and also introduces Magnitude-Weighted Hessian (MWH) for model pruning. Experiments on CIFAR and ImageNet corruption datasets showed SA-SAM achieved better robustness than existing methods at 80-90% sparsity while maintaining clean accuracy. AI

IMPACT Enhances model robustness and efficiency for deployment in real-world, potentially noisy, conditions.

RANK_REASON The cluster contains a research paper detailing a new method for deep neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• CIFAR-100-C
• CIFAR-10-C
• CLMP
• ImageNet-100-C
• Magnitude-Weighted Hessian
• Sam
• Sa Samuha
• Sparsity-Adaptive Sharpness-Aware Minimization

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →