Pay Attention to Attention Distribution: A New Local Lipschitz Bound for Transformers

Benchmark Model Rank Results
image-classification-on-cifar-10JaSMin (ViT-B, k=0, lambda=1e-2, mean)–Percentage correct: 98.55
image-classification-on-cifar-100JaSMin (ViT-B, k=10, lambda=1e-3, max)–Percentage correct: 90.94