Image Classification on CIFAR-10 data-augmented (+) (test)
96.7AccuracyDENSENET
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DENSENETPARAMS.=25.6 M, REACHABLE?=No2017.03 | 96.7 | — | |
| RL (best result)Starting Point=39 LAYERS, 2 POOL LAYERS AT 13 AND 26, 50% SKIPS, Constraints=DISCRETE PARAMS., EXACTLY 39 LAYERS, 2 POOL LAYERS AT 13 AND 26, Post-Processing=ADD MORE FILTERS, SMALL GRID SEARCH, RETRAIN, Params=37.0 M2017.03 | 96.4 | — | |
| WIDE RESNET 40-10+D/OPARAMS.=50.7 M, REACHABLE?=No2017.03 | 96.2 | — | |
| WIDE RESNET 28-10PARAMS.=36.5 M, REACHABLE?=YES2017.03 | 96 | — | |
| EvolutionStarting Point=SINGLE LAYER, ZERO CONVS., Constraints=POWER-OF-2 STRIDES, Post-Processing=NONE, Params=ENSEMB.2017.03 | 95.6 | — | |
| EvolutionStarting Point=SINGLE LAYER, ZERO CONVS., Constraints=POWER-OF-2 STRIDES, Post-Processing=NONE, Params=5.4 M2017.03 | 94.6 | — | |
| EVOLUTIONPARAMS.=5.4 M, REACHABLE?=N/A2017.03 | 94.6 | — | |
| RL (similar search space)Starting Point=20 LAYERS, 50% SKIPS, Constraints=DISCRETE PARAMS., EXACTLY 20 LAYERS, Post-Processing=SMALL GRID SEARCH, RETRAIN, Params=2.5 M2017.03 | 94 | — | |
| RESNETPARAMS.=1.7 M, REACHABLE?=YES2017.03 | 93.4 | — | |
| Q-LearningConstraints=DISCRETE PARAMS., MAX. NUM. LAYERS, NO SKIPS, Post-Processing=TUNE, RETRAIN, Params=11.2 M2017.03 | 93.1 | — | |
| ALL-CNNPARAMS.=1.3 M, REACHABLE?=YES2017.03 | 92.8 | — | |
| HIGHWAYPARAMS.=2.3 M, REACHABLE?=No2017.03 | 92.3 | — | |
| DEEPLY SUPERVISEDREACHABLE?=No2017.03 | 92 | — | |
| NETWORK IN NETWORKREACHABLE?=No2017.03 | 91.2 | — | |
| MAXOUTREACHABLE?=No2017.03 | 90.7 | — | |
| BayesianStarting Point=3 LAYERS, Constraints=FIXED ARCHITECTURE, NO SKIPS, Post-Processing=NONE2017.03 | 90.5 | — | |
| C8-Aug.Backbone=ResNet50, Pre-training=ImageNet-1K2023.10 | — | 94.36 | |
| C8-Aug.Backbone=ViT-Base, Pre-training=ImageNet-1K2023.10 | — | 94.17 | |
| Learned Canonicalization (LC)Backbone=ResNet50, Pre-training=ImageNet-1K2023.10 | — | 92.96 | |
| Learned Canonicalization (LC)Backbone=ViT-Base, Pre-training=ImageNet-1K2023.10 | — | 94.8 | |
| Prior-Regularized LCBackbone=ResNet50, Pre-training=ImageNet-1K2023.10 | — | 95.31 | |
| Prior-Regularized LCBackbone=ViT-Base, Pre-training=ImageNet-1K2023.10 | — | 95.08 | |
| Rotation AugmentationBackbone=ResNet50, Pre-training=ImageNet-1K2023.10 | — | 90.11 | |
| Rotation AugmentationBackbone=ViT-Base, Pre-training=ImageNet-1K2023.10 | — | 93.67 | |
| VanillaBackbone=ResNet50, Pre-training=ImageNet-1K2023.10 | — | 57.77 | |
| VanillaBackbone=ViT-Base, Pre-training=ImageNet-1K2023.10 | — | 63.59 |