Attribution Robustness on Imagenette (val)
84AccuracyExplicit Curv. Reg. (ECR)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Explicit Curv. Reg. (ECR)Backbone=ResNet50, Comp. Cost=F + Bx + Bθx, hparams.=lr + λ2026.05 | 84 | 11 | 2 | |
| Implicit Curv. Reg. (ICR)Backbone=ResNet50, Comp. Cost=F + Bθ, hparams.=lr2026.05 | 83 | 10 | 1 | |
| Sharp. Aware Min. (SAM)Backbone=ResNet50, Comp. Cost=2F + 2Bθ, hparams.=lr + ρ2026.05 | 79 | 30 | 5 | |
| Adv. Training Reg. (ATR)Backbone=ResNet50, Comp. Cost=2F + Bx + Bθ, hparams.=lr + ϵ2026.05 | 78 | 21 | 3 | |
| No Regularization (Base)Backbone=ResNet50, Comp. Cost=F + Bθ, hparams.=-2026.05 | 77 | 61 | 5 | |
| Ante-hoc Act. Reg. (AAR)Backbone=ResNet50, Comp. Cost=F + Bθ, hparams.=lr + β2026.05 | 75 | 10 | 4 | |
| Explicit Curv. Reg. (ECR)Backbone=ViT-B/16, Comp. Cost=F + Bx + Bθx, hparams.=lr + λ2026.05 | 74 | 8 | 2 | |
| Implicit Curv. Reg. (ICR†)Backbone=ViT-B/16, Comp. Cost=F + Bθ, hparams.=lr2026.05 | 73 | 7 | 1 | |
| Adv. Training Reg. (ATR)Backbone=ViT-B/16, Comp. Cost=2F + Bx + Bθ, hparams.=lr + ϵ2026.05 | 68 | 8 | 3 | |
| No Regularization (Base)Backbone=ViT-B/16, Comp. Cost=F + Bθ, hparams.=-2026.05 | 67 | 11 | 5 | |
| Sharp. Aware Min. (SAM)Backbone=ViT-B/16, Comp. Cost=2F + 2Bθ, hparams.=lr + ρ2026.05 | 66 | 10 | 5 | |
| Ante-hoc Act. Reg. (AAR)Backbone=ViT-B/16, Comp. Cost=F + Bθ, hparams.=lr + β2026.05 | 63 | 7 | 4 | |
| Post-hoc Act. Reg. (PAR)Backbone=ViT-B/16, Comp. Cost=F + Bθ, hparams.=lr + β2026.05 | 59 | 6 | 4 | |
| Post-hoc Act. Reg. (PAR)Backbone=ResNet50, Comp. Cost=F + Bθ, hparams.=lr + β2026.05 | 10 | 0 | 4 |