Classification on CUB (test)
94.11AccuracyEPNetSSL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| EPNetSSLBackbone=WRN-28-10, Number of shots=10-shot2020.03 | 94.11 | — | |
| EPNetBackbone=WRN-28-10, Number of shots=10-shot2020.03 | 93.99 | — | |
| EPNetSSLBackbone=RESNET-12, Number of shots=10-shot2020.03 | 92.88 | — | |
| EPNetBackbone=RESNET-12, Number of shots=10-shot2020.03 | 92.68 | — | |
| BaseTransformerBackbone=Res12, Number of shots=5, Number of ways=52022.10 | 90.64 | — | |
| Idefics2Protocol=MTV + 1-shot ICL2024.06 | 89.8 | — | |
| VILA-1.5-8BProtocol=MTV + 1-shot ICL2024.06 | 89.7 | — | |
| MANTIS-LLaMA3-8BProtocol=MTV + 1-shot ICL2024.06 | 89.7 | — | |
| MNPsnoise levels=102023.04 | 88.96 | — | |
| Idefics2Protocol=1-shot ICL2024.06 | 88.7 | — | |
| DeepEMDBackbone=Res12, Number of shots=5, Number of ways=52022.10 | 88.69 | — | |
| VILA-1.5-8BProtocol=1-shot ICL2024.06 | 88.4 | — | |
| MGPnoise levels=102023.04 | 85.48 | — | |
| MELRBackbone=Conv4-64, Number of shots=5, Number of ways=52022.10 | 85.01 | — | |
| IEPTBackbone=Conv4-64, Number of shots=5, Number of ways=52022.10 | 84.33 | — | |
| MANTIS-LLaMA3-8BProtocol=1-shot ICL2024.06 | 84 | — | |
| FEATBackbone=Conv4-64, Number of shots=5, Number of ways=52022.10 | 82.9 | — | |
| BaseTransformerBackbone=Res12, Number of shots=1, Number of ways=52022.10 | 82.27 | — | |
| BaseTransformerBackbone=Conv4-64, Number of shots=5, Number of ways=52022.10 | 82.12 | — | |
| ProtoNetsBackbone=Conv4-64, Number of shots=5, Number of ways=52022.10 | 81.82 | — | |
| Qwen-VL-7BProtocol=MTV + 1-shot ICL2024.06 | 80 | — | |
| CEM2023.09 | 79 | — | |
| Joint CBMBottleneck activation function=Logit2023.09 | 78.16 | — | |
| IntCEMlambda_roll=12023.09 | 78.03 | — | |
| SPAM (Linear, Order 2)Interaction Order=2, Model Architecture=Linear2022.05 | 77.86 | — | |
| IntCEMlambda_roll=0.12023.09 | 77.79 | — | |
| NB2Minteraction_order=pairwise2022.05 | 77.7 | — | |
| SPAM (Neural, Order 2)Interaction Order=2, Model Architecture=Neural2022.05 | 77.62 | — | |
| SPAM (Neural, Order 3)Interaction Order=3, Model Architecture=Neural2022.05 | 77.53 | — | |
| SPAM (Linear, Order 3)Interaction Order=3, Model Architecture=Linear2022.05 | 77.41 | — | |
| NA2Minteraction_order=pairwise2022.05 | 77.13 | — | |
| DE (EF)noise levels=102023.04 | 76.94 | — | |
| DNNModel Architecture=Neural2022.05 | 76.84 | — | |
| MLP2022.05 | 76.84 | — | |
| NBMinteraction_order=first-order2022.05 | 76.83 | — | |
| NAMInteraction Order=1, Model Architecture=Neural2022.05 | 76.32 | — | |
| NAMinteraction_order=first-order2022.05 | 76.32 | — | |
| IntCEMlambda_roll=52023.09 | 76.2 | — | |
| Linear (Order 2)Interaction Order=2, Model Architecture=Linear2022.05 | 76.17 | — | |
| DE (LF)noise levels=102023.04 | 76.16 | — | |
| MCDnoise levels=102023.04 | 76.08 | — | |
| DeepEMDBackbone=Res12, Number of shots=1, Number of ways=52022.10 | 75.65 | — | |
| Joint CBMBottleneck activation function=Sigmoid2023.09 | 74.95 | — | |
| Linear (Order 1)Interaction Order=1, Model Architecture=Linear2022.05 | 74.51 | — | |
| Linear2022.05 | 74.51 | — | |
| TMCnoise levels=102023.04 | 74.19 | — | |
| ETPnoise levels=102023.04 | 73.99 | — | |
| BaseTransformerBackbone=Conv4-64, Number of shots=1, Number of ways=52022.10 | 72.15 | — | |
| XGBoostModel Architecture=Trees2022.05 | 71.86 | — | |
| XGBoost2022.05 | 71.86 | — | |
| MELRBackbone=Conv4-64, Number of shots=1, Number of ways=52022.10 | 70.26 | — | |
| IEPTBackbone=Conv4-64, Number of shots=1, Number of ways=52022.10 | 69.97 | — | |
| FEATBackbone=Conv4-64, Number of shots=1, Number of ways=52022.10 | 68.87 | — | |
| LLaVA-1.6-13BProtocol=1-shot ICL2024.06 | 67.9 | — | |
| BiFTABackbone=ViT-L/14, Zero-shot=true2026.01 | 65.67 | — | |
| WCABackbone=ViT-L/14, Zero-shot=true2026.01 | 65.12 | — | |
| ProtoNetsBackbone=Conv4-64, Number of shots=1, Number of ways=52022.10 | 64.42 | — | |
| DKTKernel=CosSim, Shots=1-shot, Ways=5-ways, Backbone=Conv-4, Number of runs=32019.10 | 63.37 | — | |
| CLIP-DBackbone=ViT-L/14, Zero-shot=true2026.01 | 63.26 | — | |
| DKTKernel=BNCosSim, Shots=1-shot, Ways=5-ways, Backbone=Conv-4, Number of runs=32019.10 | 62.96 | — | |
| CLIP-EBackbone=ViT-L/14, Zero-shot=true2026.01 | 62.53 | — | |
| RelationNetShots=1-shot, Ways=5-ways, Backbone=Conv-4, Number of runs=32019.10 | 62.52 | — | |
| WaffleBackbone=ViT-L/14, Zero-shot=true2026.01 | 62.27 | — | |
| CuPLBackbone=ViT-L/14, Zero-shot=true2026.01 | 62.15 | — | |
| CLIPBackbone=ViT-L/14, Zero-shot=true2026.01 | 62.12 | — | |
| IDEFICS-9BProtocol=1-shot ICL2024.06 | 62 | — | |
| Baseline++Shots=1-shot, Ways=5-ways, Backbone=Conv-4, Number of runs=32019.10 | 61.75 | — | |
| SNGPnoise levels=102023.04 | 61.27 | — | |
| Ind. CBMTraining paradigm=Independent2023.09 | 60.25 | — | |
| DKTKernel=Linear, Shots=1-shot, Ways=5-ways, Backbone=Conv-4, Number of runs=32019.10 | 60.23 | — | |
| MatchingNetShots=1-shot, Ways=5-ways, Backbone=Conv-4, Number of runs=32019.10 | 60.19 | — | |
| LLaVA-1.5-13BProtocol=1-shot ICL2024.06 | 58.24 | — | |
| Qwen-VL-7BProtocol=1-shot ICL2024.06 | 56.5 | — | |
| MAMLShots=1-shot, Ways=5-ways, Backbone=Conv-4, Number of runs=32019.10 | 56.11 | — | |
| Emu 37BProtocol=1-shot ICL2024.06 | 53.56 | — | |
| Seq. CBMTraining paradigm=Sequential2023.09 | 53.41 | — | |
| ProtoNetShots=1-shot, Ways=5-ways, Backbone=Conv-4, Number of runs=32019.10 | 52.52 | — | |
| Flamingo 9BProtocol=1-shot ICL2024.06 | 51.2 | — | |
| Feature TransferShots=1-shot, Ways=5-ways, Backbone=Conv-4, Number of runs=32019.10 | 46.19 | — | |
| IntCEMFraction of concepts intervened=25%, Intervention Policy=Random2023.09 | — | 88.53 | |
| IntCEMFraction of concepts intervened=25%, Intervention Policy=Learnt Policy ψ2023.09 | — | 94.1 | |
| IntCEMFraction of concepts intervened=50%, Intervention Policy=Random2023.09 | — | 96.14 | |
| IntCEMFraction of concepts intervened=50%, Intervention Policy=Learnt Policy ψ2023.09 | — | 99.17 | |
| IntCEMFraction of concepts intervened=75%, Intervention Policy=Random2023.09 | — | 98.98 | |
| IntCEMFraction of concepts intervened=75%, Intervention Policy=Learnt Policy ψ2023.09 | — | 99.82 | |
| IntCEMFraction of concepts intervened=100%, Intervention Policy=Random2023.09 | — | 99.9 | |
| IntCEMFraction of concepts intervened=100%, Intervention Policy=Learnt Policy ψ2023.09 | — | 99.9 |