Image Classification on CIFAR-10 (Top-1 Accuracy vs. Probability Thresholds)
93.23Top-1 Acc (p=0.0)Fine-tuning
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Fine-tuningBackbone=ResNet-32, Noise Type=Flip2022.11 | 93.23 | 82.47 | 74.07 | — | — | |
| Focal LossBackbone=ResNet-32, Noise Type=Flip2022.11 | 93.03 | 86.45 | 80.45 | — | — | |
| CrossEntropyBackbone=ResNet-32, Noise Type=Flip, Implementation=Reported from original paper2022.11 | 92.89 | 76.83 | 70.77 | — | — | |
| CrossEntropyBackbone=ResNet-32, Noise Type=Flip, Implementation=Our implementation2022.11 | 92.33 | 90.56 | 86.25 | 26.67 | 13.58 | |
| Reed-HardBackbone=ResNet-32, Noise Type=Flip2022.11 | 92.31 | 88.28 | 81.06 | — | — | |
| MW-netBackbone=ResNet-32, Noise Type=Flip, Implementation=Our implementation2022.11 | 92.19 | 90.74 | 87.63 | 42.41 | 27.19 | |
| MentorNetBackbone=ResNet-32, Noise Type=Flip2022.11 | 92.13 | 86.3 | 81.76 | — | — | |
| MW-netBackbone=ResNet-32, Noise Type=Flip, Implementation=Reported from original paper2022.11 | 92.04 | 90.33 | 87.54 | — | — | |
| D2LBackbone=ResNet-32, Noise Type=Flip2022.11 | 92.02 | 87.66 | 83.89 | — | — | |
| Advisor Network (Ours)Backbone=ResNet-32, Noise Type=Flip, Implementation=Ours2022.11 | 91.87 | 91.09 | 90.26 | 89.34 | 82.47 | |
| GLCBackbone=ResNet-32, Noise Type=Flip2022.11 | 91.02 | 89.68 | 88.92 | — | — | |
| Co-teachingBackbone=ResNet-32, Noise Type=Flip2022.11 | 89.87 | 82.83 | 75.41 | — | — | |
| L2RWBackbone=ResNet-32, Noise Type=Flip2022.11 | 89.25 | 87.86 | 85.66 | — | — | |
| Self-pacedBackbone=ResNet-32, Noise Type=Flip2022.11 | 88.52 | 87.03 | 81.63 | — | — | |
| S-ModelBackbone=ResNet-32, Noise Type=Flip2022.11 | 83.61 | 79.25 | 75.73 | — | — |