Image Classification on CIFAR-10 (test) (Accuracy vs Sample Size)
88.5Accuracy (N=100)DE ELBo
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DE ELBoModel Method=DE ELBo (ours), Backbone=ViT-B/16, Regularization=L2-SP2025.02 | 88.5 | 94.1 | 97.4 | 98.2 | |
| L2-SPModel Method=MAP + GS, Backbone=ViT-B/162025.02 | 86.5 | 93.8 | 97.2 | 98.2 | |
| L2-SPModel Method=MAP + GS, Backbone=ConNeXt-Tiny2025.02 | 85.4 | 94.1 | 96.9 | 97.9 | |
| DE ELBoModel Method=DE ELBo (ours), Backbone=ConNeXt-Tiny, Regularization=L2-SP2025.02 | 83.9 | 94.5 | 97.1 | 97.8 | |
| DE ELBoModel Method=DE ELBo (ours), Backbone=ResNet-50, Pretraining=PTYL (SSL)2025.02 | 69.9 | 88.5 | 95.1 | 96.8 | |
| PTYL (SSL)Model Method=MAP + GS, Backbone=ResNet-502025.02 | 69 | 88.7 | 95.4 | 97.3 | |
| L2-SPModel Method=MAP + GS, Backbone=ResNet-502025.02 | 67.7 | 87.7 | 94.6 | 97 | |
| L2-SPModel Method=MAP + Laplace, Backbone=ResNet-502025.02 | 67.5 | 88.2 | 95.1 | 96.9 | |
| DE ELBoModel Method=DE ELBo (ours), Backbone=ResNet-50, Regularization=L2-SP2025.02 | 62.3 | 87.1 | 91.3 | 92.9 | |
| DE ELBoModel Method=DE ELBo (ours), Backbone=ResNet-50, Pretraining=PTYL2025.02 | 60.2 | 78.1 | 90.6 | 96.7 | |
| Linear probingModel Method=MAP + GS, Backbone=ResNet-502025.02 | 60.1 | 74.5 | 81.4 | 83.2 | |
| L2-zeroModel Method=MAP + GS, Backbone=ResNet-502025.02 | 58.4 | 73.6 | 81.6 | 83.2 | |
| PTYLModel Method=MAP + GS, Backbone=ResNet-502025.02 | 57.5 | 78.4 | 90.6 | 96.6 | |
| PTYL (SSL)Model Method=MAP + Laplace, Backbone=ResNet-502025.02 | 35 | 72 | 76.4 | 94.1 |