Self-Supervised Learning for OOD Detection

A Simplified Pytorch implementation of Using Self-Supervised Learning Can Improve Model Robustness and Uncertainty(NeurIPS 2019)

The code supports only Multi-class OOD Detection experiment(in-dist: CIFAR-10, Out-of-dist: CIFAR-100/SVHN)

• Command

  • RotNet-OOD

    python test.py --method=rot --ood_dataset=cifar100

  • baseline

    python test.py --method=msp --ood_dataset=svhn

Results (OOD Detection)

• Metric : AUROC

	CIFAR-100	SVHN
Maximum Softmax Probability (baseline)	0.6986	0.7190
RotNet	0.7931	0.9584
RotNet (rot loss only)	0.7132	0.9560
RotNet (KL divergence only)	0.7834	0.8522

Reference

[1] full code(by authors): https://github.com/hendrycks/ss-ood

[2] Using Self-Supervised Learning Can Improve Model Robustness and Uncertainty(NeurIPS 2019): https://arxiv.org/abs/1906.12340

[3] A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks(ICLR 2017): https://arxiv.org/abs/1610.02136