Independent PyTorch reproduction of ICR/ECR attribution-robustness training (Mehrpanah et al., ICML 2026) on ResNet18/CIFAR-10, includes a diagnostic that traces a non-reproducing ICR result to the learning-rate sweep never reaching the edge-of-stability regime the mechanism depends on.
deep-learning pytorch reproducibility cifar10 explainable-ai gradcam resnet18 adversarial-robustness feature-attribution paper-reproduction implicit-regularization vanilla-gradient-descent edge-of-stability
-
Updated
Sep 1, 2026 - Jupyter Notebook