Image Classification on ImageNet Robustness Generalization Suite (Sketch, A, R, V2)
67.77Top-1 Acc (V2)MMA + NeRP
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| MMA + NeRPNeRP=true2026.05 | 67.77 | 50.07 | 51.77 | 78.2 | 61.95 | — | |
| PromptSRC + NeRPNeRP=true2026.05 | 66.87 | 49.77 | 52.77 | 78.13 | 61.89 | — | |
| CoOpTPT=true2023.11 | 66.83 | 49.29 | 57.95 | 77.27 | 62.84 | — | |
| CoCoOp + NeRPNeRP=true2026.05 | 66.73 | 50.2 | 52.47 | 77 | 61.6 | — | |
| MMRL + NeRPNeRP=true2026.05 | 66.63 | 49.27 | 53.33 | 77.93 | 61.79 | — | |
| GraphAdapterBackbone=ViT-B/16, Shot count=16-shot optimization, Variant=Generalization optimized (alpha=0.8)2023.09 | 65.6 | 49.23 | 50.57 | 77.73 | 60.78 | — | |
| GraphAdapterBackbone=ViT-B/16, Shot count=16-shot optimization, Variant=Few-shot optimized2023.09 | 65.57 | 48.57 | 49.23 | 77.2 | 60.14 | — | |
| TaskResBackbone=ViT-B/16, Shot count=16-shot optimization2023.09 | 65.3 | 49.13 | 50.37 | 77.7 | 60.63 | — | |
| PromptAlignTPT=false2023.11 | 65.29 | 50.23 | 59.37 | 79.33 | 63.55 | — | |
| Co-CoOpTPT=true2023.11 | 64.85 | 48.27 | 58.47 | 78.65 | 62.61 | — | |
| MMRLNeRP=false2026.05 | 64.47 | 49.17 | 51.12 | 77.32 | 60.52 | — | |
| CLIPTPT=true2023.11 | 64.35 | 47.94 | 54.77 | 77.06 | 60.81 | — | |
| PromptSRCNeRP=false2026.05 | 64.35 | 49.55 | 50.9 | 77.8 | 60.65 | — | |
| MMANeRP=false2026.05 | 64.33 | 49.13 | 51.12 | 77.32 | 60.48 | — | |
| CoOpTPT=false2023.11 | 64.2 | 47.99 | 49.71 | 75.21 | 59.28 | — | |
| CoOpBackbone=ViT-B/16, Shot count=16-shot optimization2023.09 | 64.18 | 46.71 | 48.41 | 74.32 | 58.41 | — | |
| Co-CoOpTPT=false2023.11 | 64.07 | 48.75 | 50.63 | 76.18 | 59.91 | — | |
| CoCoOpNeRP=false2026.05 | 64.07 | 48.75 | 50.63 | 76.18 | 59.91 | — | |
| CLIPTPT=false2023.11 | 60.86 | 46.09 | 47.87 | 73.98 | 57.2 | — | |
| Zero-shot CLIPBackbone=ViT-B/16, Evaluation protocol=Zero-shot2023.09 | 60.83 | 46.15 | 47.77 | 73.96 | 57.18 | — | |
| GraphAdapterBackbone=ResNet-101, Shot count=16-shot optimization, Variant=Few-shot optimized2023.09 | 59.6 | 40.83 | 28.77 | 67.13 | 49.08 | — | |
| TaskResBackbone=ResNet-101, Shot count=16-shot optimization2023.09 | 59.5 | 41.7 | 29.87 | 68.07 | 49.79 | — | |
| GraphAdapterBackbone=ResNet-101, Shot count=16-shot optimization, Variant=Generalization optimized (alpha=0.8)2023.09 | 59.5 | 41.6 | 30 | 68.1 | 49.8 | — | |
| TaskResBackbone=ViT-B/32, Shot count=16-shot optimization2023.09 | 59.2 | 42.5 | 31.43 | 69.33 | 50.62 | — | |
| GraphAdapterBackbone=ViT-B/32, Shot count=16-shot optimization, Variant=Generalization optimized (alpha=0.8)2023.09 | 59.1 | 42.7 | 31.73 | 69.43 | 50.74 | — | |
| GraphAdapterBackbone=ViT-B/32, Shot count=16-shot optimization, Variant=Few-shot optimized2023.09 | 59 | 41.7 | 29.57 | 68.67 | 49.74 | — | |
| CoOpBackbone=ResNet-101, Shot count=16-shot optimization2023.09 | 58.66 | 39.08 | 28.89 | 63 | 47.41 | — | |
| CoOpBackbone=ViT-B/32, Shot count=16-shot optimization2023.09 | 58.08 | 40.44 | 30.62 | 64.45 | 48.4 | — | |
| GraphAdapterBackbone=ResNet-50, Shot count=16-shot optimization, Variant=Generalization optimized (alpha=0.8)2023.09 | 56.58 | 35.89 | 23.07 | 60.86 | 44.1 | — | |
| TaskResBackbone=ResNet-50, Shot count=16-shot optimization2023.09 | 56.47 | 35.83 | 22.8 | 60.7 | 43.95 | — | |
| GraphAdapterBackbone=ResNet-50, Shot count=16-shot optimization, Variant=Few-shot optimized2023.09 | 56.4 | 34.5 | 21.88 | 58.94 | 42.93 | — | |
| Linear Probe CLIPBackbone=ViT-B/16, Evaluation protocol=Linear Probe2023.09 | 56.26 | 34.77 | 35.68 | 58.43 | 46.29 | — | |
| CoOpBackbone=ResNet-50, Shot count=16-shot optimization2023.09 | 55.11 | 32.74 | 22.12 | 54.96 | 41.23 | — | |
| Zero-shot CLIPBackbone=ResNet-101, Evaluation protocol=Zero-shot2023.09 | 54.81 | 38.71 | 28.05 | 64.38 | 46.49 | — | |
| Zero-shot CLIPBackbone=ViT-B/32, Evaluation protocol=Zero-shot2023.09 | 54.79 | 40.82 | 29.57 | 65.99 | 47.79 | — | |
| Zero-shot CLIPBackbone=ResNet-50, Evaluation protocol=Zero-shot2023.09 | 51.34 | 33.32 | 21.65 | 56 | 40.58 | — | |
| Linear Probe CLIPBackbone=ResNet-101, Evaluation protocol=Linear Probe2023.09 | 50.05 | 26.8 | 19.44 | 47.19 | 35.87 | — | |
| Linear Probe CLIPBackbone=ViT-B/32, Evaluation protocol=Linear Probe2023.09 | 49.73 | 28.06 | 19.67 | 47.2 | 36.17 | — | |
| Linear Probe CLIPBackbone=ResNet-50, Evaluation protocol=Linear Probe2023.09 | 45.97 | 19.07 | 12.74 | 28.16 | 28.16 | — | |
| Lin. P. w/ PCA and FastICAEncoder=ViT16, Shots=12024.02 | 30.26 | 19.17 | 26.41 | 38.87 | 28.68 | 35.2 | |
| Lin. P. w/ FastICAEncoder=ViT16, Shots=12024.02 | 29.31 | 13.91 | 23.24 | 28.61 | 23.77 | 34.35 | |
| Lin. P.Encoder=ViT16, Shots=12024.02 | 27.64 | 16.34 | 21.84 | 34.28 | 25.02 | 32.42 | |
| Lin. P. w/ FastICAEncoder=ViT32, Shots=12024.02 | 24.8 | 9.97 | 12.23 | 21.23 | 17.06 | 29.21 | |
| Lin. P. w/ PCA and FastICAEncoder=ViT32, Shots=12024.02 | 24.45 | 12.39 | 12.56 | 27.61 | 19.25 | 29.05 | |
| Lin. P. w/ PCA and FastICAEncoder=RN101, Shots=12024.02 | 24.17 | 11.63 | 12.28 | 26.38 | 18.62 | 28.5 | |
| Lin. P. w/ FastICAEncoder=RN101, Shots=12024.02 | 23.33 | 8.35 | 10.71 | 17.87 | 15.07 | 27.5 | |
| Lin. P.Encoder=ViT32, Shots=12024.02 | 22.99 | 11.93 | 11.56 | 25.25 | 17.93 | 26.99 | |
| Lin. P.Encoder=RN101, Shots=12024.02 | 21.48 | 9.9 | 10.17 | 23.85 | 16.35 | 26.05 | |
| Lin. P. w/ PCA and FastICAEncoder=RN50, Shots=12024.02 | 20.26 | 6.85 | 8.05 | 17.54 | 13.18 | 24.06 | |
| Lin. P. w/ FastICAEncoder=RN50, Shots=12024.02 | 19.68 | 6.37 | 7.21 | 13.84 | 11.77 | 23.22 | |
| Lin. P.Encoder=RN50, Shots=12024.02 | 18.24 | 5.68 | 6.55 | 15.41 | 11.47 | 21.74 |