On the Importance of Gradients for Detecting Distributional Shifts in the Wild

Benchmark Model Rank Results
out-of-distribution-detection-on-imagenet-1k-10GradNorm (ResNetv2-101)#29FPR95: 61.42
out-of-distribution-detection-on-imagenet-1k-3GradNorm#24FPR95: 50.03
out-of-distribution-detection-on-imagenet-1k-8GradNorm (ResNetv2-101)#13FPR95: 46.48
out-of-distribution-detection-on-imagenet-1k-9GradNorm (ResNetv2-101)#17FPR95: 60.86