Ablation study on the proposed self-supervised adversarial training, which is integrated by cross entropy reformulation in equation (10), adversarial training in equation (15), augmentation alignment in equation (15) and adversarial invariant regularization (AIR) in Xu et al. (2023b) where CIFAR-10 was used
| Methods | CE reform. | AT | Aug align. | AIR | RA | SA |
|---|---|---|---|---|---|---|
| SimCLR baseline | 0.03 | 91.77 | ||||
| + AT (RoCL) | ✓ | 39.93 | 78.37 | |||
| + CE reformulation | ✓ | 40.98 | 77.42 | |||
| + CE + AT (ours (15)) | ✓ | ✓ | 48.35 | 76.34 | ||
| + CE + AT + aug alignment | ✓ | ✓ | ✓ | 51.15 | 76.90 | |
| + CE + AT + aug alignment + AIR | ✓ | ✓ | ✓ | ✓ | 53.05 | 77.21 |
| Methods | Aug align. | |||||
|---|---|---|---|---|---|---|
| SimCLR baseline | 0.03 | |||||
| + | ✓ | 39.93 | 78.37 | |||
| + | ✓ | 40.98 | 77.42 | |||
| + | ✓ | ✓ | 48.35 | 76.34 | ||
| + | ✓ | ✓ | ✓ | 51.15 | 76.90 | |
| + | ✓ | ✓ | ✓ | ✓ | 77.21 |
Sharing content requires targeting cookies to be enabled. Please update your cookie preferences to use this feature.