Image Classification on CIFAR-10 (Standard and Robust Accuracy)
87.24Standard AccuracyRCT-AF (β = 1)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| RCT-AF (β = 1)Activation Function=RCT-AF, β=1, max |σ''|=7.0, Backbone=ResNet-18, Training Method=DAJAT, Evaluation Protocol=AutoAttack2026.03 | 87.24 | 54.79 | |
| RCT-AF (β = 0)Activation Function=RCT-AF, β=0, max |σ''|=7.0, Backbone=ResNet-18, Training Method=DAJAT, Evaluation Protocol=AutoAttack2026.03 | 86.65 | 54.57 | |
| RCT-AF (β = 2)Activation Function=RCT-AF, β=2, max |σ''|=7.0, Backbone=ResNet-18, Training Method=DAJAT, Evaluation Protocol=AutoAttack2026.03 | 86.08 | 54.49 | |
| ReLUActivation Function=ReLU, max |σ''|=∞, Backbone=ResNet-18, Training Method=DAJAT, Evaluation Protocol=AutoAttack2026.03 | 85.23 | 51.37 | |
| GELUActivation Function=GELU, max |σ''|=0.798, Backbone=ResNet-18, Training Method=DAJAT, Evaluation Protocol=AutoAttack2026.03 | 84.88 | 51.69 | |
| Leaky ReLUActivation Function=Leaky ReLU, max |σ''|=∞, negative slope=0.01, Backbone=ResNet-18, Training Method=DAJAT, Evaluation Protocol=AutoAttack2026.03 | 84.52 | 51.24 | |
| MishActivation Function=Mish, max |σ''|=0.362, Backbone=ResNet-18, Training Method=DAJAT, Evaluation Protocol=AutoAttack2026.03 | 82.63 | 48.75 | |
| SwishActivation Function=Swish, max |σ''|=0.5, Backbone=ResNet-18, Training Method=DAJAT, Evaluation Protocol=AutoAttack2026.03 | 82.53 | 48.08 | |
| ELUActivation Function=ELU, max |σ''|=1.0, Backbone=ResNet-18, Training Method=DAJAT, Evaluation Protocol=AutoAttack2026.03 | 78.39 | 43.11 |