Explainable AI Evaluation on ImageNet-S
6.785SelectionSplitLRP (sc, wta)
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| SplitLRP (sc, wta)Layer=(1.0.2, pos), Abs=y, Backbone=ResNet502026.01 | 6.785 | 0.483 | 56.7 | 5 | 0.686 | 0.696 | |
| SplitCAM (sc, α 0.5, wta)Layer=(4.1.1, +, g), Abs=y, Backbone=ResNet502026.01 | 6.526 | 0.611 | 95.6 | 6.6 | 0.737 | 0.187 | |
| SplitLRP (sc, wta)Layer=(conv1, pos), Abs=y, Backbone=ResNet502026.01 | 6.313 | 0.506 | 58.7 | 5.6 | 0.697 | 0.755 | |
| GradCAM++Layer=4.1.2, Abs=n, Backbone=ResNet502026.01 | 6.054 | 0.556 | 94.5 | 7.7 | 0.838 | 0.14 | |
| SplitGrad (α 0.5, wta)Layer=(4.1.2, g), Abs=y, Backbone=ResNet502026.01 | 5.69 | 0.654 | 97.2 | 8.3 | 0.886 | 0.23 | |
| SplitGrad (α 0.3, wta)Layer=(4.1.2, g), Abs=y, Backbone=ResNet502026.01 | 5.663 | 0.649 | 97.2 | 8.2 | 0.882 | 0.219 | |
| SplitCAM (sc, α 0.3, wta)Layer=(4.1.2, g), Abs=n, Backbone=ResNet502026.01 | 5.654 | 0.673 | 97.2 | 8.1 | 0.88 | 0.233 | |
| DeconvolutionLayer=conv1, Abs=y, Backbone=ResNet502026.01 | 5.474 | 0.478 | 66.6 | 15.1 | 1.177 | 0.618 | |
| OcclusionLayer=conv1, Abs=n, Backbone=ResNet502026.01 | 5.3 | 0.179 | 72.6 | 13 | 1.03 | 0.785 | |
| Gradient SHAPLayer=conv1, Abs=y, Backbone=ResNet502026.01 | 5.269 | 0.527 | 77.9 | 16.6 | 1.227 | 1.074 | |
| Integrated GradientsLayer=conv1, Abs=y, Backbone=ResNet502026.01 | 5.265 | 0.528 | 77.7 | 16.7 | 1.226 | 0.994 | |
| Feature AblationLayer=conv1, Abs=n, Backbone=ResNet502026.01 | 4.655 | 0.136 | 68.4 | 18.8 | 1.388 | 1.077 | |
| SplitGrad (α 0.3, wta)Layer=(1.0.2, g), Abs=y, Backbone=ResNet502026.01 | 4.449 | 0.513 | 77.7 | 14.9 | 1.266 | 0.813 | |
| DeconvolutionLayer=conv1, Abs=n, Backbone=ResNet502026.01 | 3.934 | 0.484 | 66.6 | 14.7 | 1.203 | 0.751 | |
| SplitCAM (sc, α 0.5, wta)Layer=(2.0.1, g), Abs=n, Backbone=ResNet502026.01 | 3.682 | -1.37 | 88.7 | 15.8 | 1.447 | 1.078 | |
| SplitGrad (α 0.5, wta)Layer=(3.1.1, g), Abs=y, Backbone=ResNet502026.01 | 3.588 | 0.559 | 88 | 13 | 1.366 | 0.859 | |
| SplitCAM (sc, α 0.3, wta)Layer=(1.0.2, g), Abs=n, Backbone=ResNet502026.01 | 3.503 | 0.487 | 74.7 | 16.8 | 1.395 | 1.277 |