ResearchBenchmarksImage Classification on ImageNet (Corruption Robustness Metrics)For AI assistantsMarkdownContent indexImage Classification on ImageNet (Corruption Robustness Metrics)Follow0Accuracy (Clean)Normal-0.001-0.000500.0005Feb 9, 2022Evaluation resultsMethodMethodLinksAccuracy (Clean)Accuracy (Hue Corruption)Accuracy (Saturation Corruption)Accuracy (Rotation)Accuracy (Brightness Corruption)Accuracy (Contrast Corruption)Accuracy (L-inf, 20-step)Accuracy (Auto L-inf)NormalTraining=Normal†Training=Normal†2022.02034.34.912.29.55.9100100MadryTraining=Madry†∞Training=Madry†∞2022.02039.96.719.914.610.846.353.6Fast-ATTraining=Fast-AT†∞Training=Fast-AT†∞2022.02049.311.129.822.81948.954.1GAT-fTraining=GAT-f†Training=GAT-f†2022.02017.73.36.83.93.357.965.1