Conceitos essenciais
SAMの一般化性能を向上させるために、F-SAMが提案されました。
Estatísticas
バッチサイズ128でResNet-18 on CIFAR-100でテスト精度80.88%
Citações
"By decomposing the minibatch gradient, we discover that the full gradient component in adversarial perturbation contributes minimally to generalization."
"Friendly perturbation in F-SAM is more 'friendly' to other data points compared with vanilla SAM."