Classification Attribution on ImageNet (misclassified samples)
66.71Ins. ScoreGreedy
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GreedyBackbone=ResNet-101, Attribution Target=model's wrong prediction2026.06 | 66.71 | 5.59 | 86.72 | 83.99 | 1,738.14 | 3.84 | |
| PhaseWinBackbone=ResNet-101, Attribution Target=model's wrong prediction2026.06 | 63.11 | 5.75 | 84.76 | 80.04 | 987.42 | 6.39 | |
| GreedyModel Architecture=CLIP ViT-L/14, Attribution Target=ground-truth class2026.06 | 48.27 | 2.28 | 74.25 | 71.61 | 1,755.77 | 4.23 | |
| GreedyBackbone=ResNet-101, Attribution target=ground-truth class2026.06 | 43.58 | 1.82 | 66.06 | 64.14 | 1,738.14 | 3.8 | |
| PhaseWinModel Architecture=CLIP ViT-L/14, Attribution Target=ground-truth class2026.06 | 42.91 | 2.43 | 69.89 | 64.58 | 1,046.46 | 6.68 | |
| PhaseWinBackbone=ResNet-101, Attribution target=ground-truth class2026.06 | 38.69 | 1.85 | 61.01 | 57.51 | 955.35 | 6.39 | |
| RISEBackbone=ResNet-101, Attribution Target=model's wrong prediction2026.06 | 37.33 | 10.65 | 63.15 | 52.52 | 5,000 | 0.74 | |
| HSICBackbone=ResNet-101, Attribution Target=model's wrong prediction2026.06 | 35.43 | 10.67 | 60.87 | 52.03 | 1,536 | 2.31 | |
| Gradient IntegralBackbone=ResNet-101, Attribution Target=model's wrong prediction2026.06 | 29.28 | 16.22 | 57.61 | 40.47 | — | — | |
| GreedyBackbone=CLIP RN101, Attribution Target=ground-truth class, MEC=1748.26, A-C=2.752026.06 | 27.47 | 1.05 | 48.02 | 41.84 | — | — | |
| GradientBackbone=ResNet-101, Attribution Target=model's wrong prediction2026.06 | 27.47 | 17.37 | 57.32 | 37.7 | — | — | |
| IGOS++Backbone=ResNet-101, Attribution Target=model's wrong prediction2026.06 | 26.8 | 16.95 | 59.84 | 35.25 | — | — | |
| PhaseWinBackbone=CLIP RN101, Attribution Target=ground-truth class, MEC=835.51, A-C=5.112026.06 | 21.93 | 1.07 | 42.67 | 31.33 | — | — | |
| RISEModel Architecture=CLIP ViT-L/14, Attribution Target=ground-truth class2026.06 | 19.9 | 5.16 | 45.41 | 37.74 | 5,000 | 0.91 | |
| Grad-ECLIPModel Architecture=CLIP ViT-L/14, Attribution Target=ground-truth class2026.06 | 18.32 | 5.08 | 44.75 | 38.98 | — | — | |
| RISEBackbone=ResNet-101, Attribution target=ground-truth class2026.06 | 14.56 | 3.18 | 33.07 | 29.43 | 5,000 | 0.66 | |
| HSICModel Architecture=CLIP ViT-L/14, Attribution Target=ground-truth class2026.06 | 14.47 | 6.42 | 36.2 | 32.27 | 1,536 | 2.36 | |
| IGOS++Model Architecture=CLIP ViT-L/14, Attribution Target=ground-truth class2026.06 | 11.81 | 8.75 | 33.61 | 22.26 | — | — | |
| HSICBackbone=ResNet-101, Attribution target=ground-truth class2026.06 | 11.02 | 3.73 | 27.51 | 25.14 | 1,536 | 1.79 | |
| IGOS++Backbone=ResNet-101, Attribution target=ground-truth class2026.06 | 10.2 | 4.87 | 27.25 | 19.51 | — | — | |
| GradientModel Architecture=CLIP ViT-L/14, Attribution Target=ground-truth class2026.06 | 9.96 | 9.68 | 32.51 | 18.52 | — | — | |
| RISEBackbone=CLIP RN101, Attribution Target=ground-truth class, MEC=5000.00, A-C=0.482026.06 | 9.75 | 1.77 | 24.33 | 17.93 | — | — | |
| Gradient IntegralModel Architecture=CLIP ViT-L/14, Attribution Target=ground-truth class2026.06 | 9.31 | 10.42 | 31.69 | 16.53 | — | — | |
| Gradient IntegralBackbone=ResNet-101, Attribution target=ground-truth class2026.06 | 9.17 | 4.88 | 24.71 | 20.77 | — | — | |
| GradientBackbone=ResNet-101, Attribution target=ground-truth class2026.06 | 8.98 | 5.08 | 24.71 | 20.07 | — | — | |
| HSICBackbone=CLIP RN101, Attribution Target=ground-truth class, MEC=1536.00, A-C=1.352026.06 | 7.95 | 2.06 | 20.74 | 15.81 | — | — | |
| Gradient IntegralBackbone=CLIP RN101, Attribution Target=ground-truth class2026.06 | 6.15 | 3.61 | 18.32 | 10.48 | — | — | |
| GradientBackbone=CLIP RN101, Attribution Target=ground-truth class2026.06 | 6.05 | 3.49 | 18.49 | 10.43 | — | — | |
| IGOS++Backbone=CLIP RN101, Attribution Target=ground-truth class2026.06 | 5.28 | 3.43 | 18.39 | 7.29 | — | — |