Contrast to Divide: Self-Supervised Pre-Training for Learning with Noisy Labels

Benchmark Model Rank Results
image-classification-on-cifar-10-with-noisyC2D (ELR+ with SimCLR, ResNet-34)#1Accuracy (under 20% Sym. label noise): 96.74 ± 0.12
image-classification-on-cifar-10-with-noisyC2D (DivideMix with SimCLR, ResNet-18)#4Accuracy (under 20% Sym. label noise): 96.23 ± 0.09
image-classification-on-clothing1mELR+ with C2D (ResNet-50)#16Accuracy: 74.58 ± 0.15%
image-classification-on-mini-webvision-1-0DivideMix with C2D (ResNet-50)#11Top-1 Accuracy: 79.42 ± 0.34Top-5 Accuracy: 92.32 ± 0.33