Table 2. Summary of test accuracies (%) on CIFAR datasets.
| Dataset | Model | Top-1 | Top-5 | FP Top-1a | FP Top-5a | Top-1 gap | Top-5 gap |
|---|---|---|---|---|---|---|---|
| CIFAR-10a | AresB-10 | 90.74 | – | – | – | – | – |
| AresB-18 | 91.90 | – | 93.02 | – | 1.12 | – | |
| AresB-34 | 92.71 | – | 93.62 | – | 0.91 | – | |
| CIFAR-100a | AresB-10 | 69.45 | 91.70 | – | – | – | – |
| AresB-18 | 73.01 | 92.57 | 75.61 | 93.05 | 2.60 | 0.48 | |
| AresB-34 | 74.73 | 93.25 | 76.76 | 93.37 | 2.03 | 0.12 |
Note:
a
Full-precision (denoted as FP) counterparts of the AresB-18 and AresB-34 models are based on the evaluation results of the ResNet-18 and ResNet-34 models.