Robust Generalization against Photon-Limited Corruptions via Worst-Case Sharpness Minimization

Code

• github.com/zhuohuangai/sharpdro
  OfficialIn paperpytorch★ 13
• github.com/2024-MindSpore-1/Code10/tree/main/SharpDRO-ms-main
  mindspore★ 0
• github.com/yangyucheng000/papercode-2/tree/main/SharpDRO-ms-main
  mindspore★ 0
• github.com/feedbackward/addro
  pytorch★ 0

Abstract

Robust generalization aims to tackle the most challenging data distributions which are rare in the training set and contain severe noises, i.e., photon-limited corruptions. Common solutions such as distributionally robust optimization (DRO) focus on the worst-case empirical risk to ensure low training error on the uncommon noisy distributions. However, due to the over-parameterized model being optimized on scarce worst-case data, DRO fails to produce a smooth loss landscape, thus struggling on generalizing well to the test set. Therefore, instead of focusing on the worst-case risk minimization, we propose SharpDRO by penalizing the sharpness of the worst-case distribution, which measures the loss changes around the neighbor of learning parameters. Through worst-case sharpness minimization, the proposed method successfully produces a flat loss curve on the corrupted distributions, thus achieving robust generalization. Moreover, by considering whether the distribution annotation is available, we apply SharpDRO to two problem settings and design a worst-case selection process for robust generalization. Theoretically, we show that SharpDRO has a great convergence guarantee. Experimentally, we simulate photon-limited corruptions using CIFAR10/100 and ImageNet30 datasets and show that SharpDRO exhibits a strong generalization ability against severe corruptions and exceeds well-known baseline methods with large performance gains.

Reproductions