Scholarlink

To develop deep neural networks (DNNs) that maintain strong performance under different natural image corruptions such as noise and blur, namely with strong corruption robustness, is essential for real-world applications. Current methods to enhance corruption robustness focus primarily on data augmentation. But these methods are often empirically designed without a clear theoretical foundation. Here, we show that flatter minima of loss landscapes generally correspond with stronger corruption robustness. Based on this finding, we have developed a method named LASAM-DA that combines layer adaptive sharpness-aware minimization(LASAM) with data augmentation (DA) to boost corruption robustness. LASAM pushes DNNs towards flat minima of loss landscapes, while DA guides DNNs to memorize diverse data distributions. LASAM and DA reinforce each other without interference. Experiments on CIFRA-10-C, CIFAR-100-C and Tiny-ImageNet-C show that LASAM-DA outperforms competing methods and achieves state-of-the-art performance. LASAM-DA is versatile, not restricted to specific data types, network architectures or data augmentation methods. Code is openly accessible at https://github.com/cbmi-group/Flatter-Minima-Helps-for-Robustness .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Flatter Minima of Loss Landscapes Correspond with Strong Corruption Robustness

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Liqun Zhong,
  • Kaijie Zhu,
  • Ge Yang

摘要

To develop deep neural networks (DNNs) that maintain strong performance under different natural image corruptions such as noise and blur, namely with strong corruption robustness, is essential for real-world applications. Current methods to enhance corruption robustness focus primarily on data augmentation. But these methods are often empirically designed without a clear theoretical foundation. Here, we show that flatter minima of loss landscapes generally correspond with stronger corruption robustness. Based on this finding, we have developed a method named LASAM-DA that combines layer adaptive sharpness-aware minimization(LASAM) with data augmentation (DA) to boost corruption robustness. LASAM pushes DNNs towards flat minima of loss landscapes, while DA guides DNNs to memorize diverse data distributions. LASAM and DA reinforce each other without interference. Experiments on CIFRA-10-C, CIFAR-100-C and Tiny-ImageNet-C show that LASAM-DA outperforms competing methods and achieves state-of-the-art performance. LASAM-DA is versatile, not restricted to specific data types, network architectures or data augmentation methods. Code is openly accessible at https://github.com/cbmi-group/Flatter-Minima-Helps-for-Robustness .