Class-wise unlearning on CIFAR-100
0.988ToWENGRAM
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ENGRAMBackbone=ResNet50, alpha=12026.06 | 0.988 | 0.757 | 0.573 | |
| RETRAINBackbone=ResNet502026.06 | 0.985 | 0.734 | 0.525 | |
| ENGRAMBackbone=ResNet50, alpha=best2026.06 | 0.983 | 0.767 | 0.496 | |
| SALUNBackbone=ResNet502026.06 | 0.98 | 0.75 | 0.634 | |
| RANDOM LABELBackbone=ResNet502026.06 | 0.964 | 0.799 | 0.778 | |
| NEGGRAD+Backbone=ResNet502026.06 | 0.894 | 0.632 | 0.12 | |
| FINE-TUNEBackbone=ResNet502026.06 | 0.86 | 0.768 | 0.585 | |
| l1-SPARSEBackbone=ResNet502026.06 | 0.837 | 0.77 | 0.546 | |
| NEGGRADBackbone=ResNet502026.06 | 0.593 | 0.677 | 0.149 |