This paper shows findings about combine self-supervised learning with supervised learning can improve model's performance on robustness and uncertainty. It serves like a summary with experiments.
In this paper, they found self-supervision can:
improve model's robustness to adversarial examples, label corruption and common input corruptions.
benefits out-of-distribution (OOD) detection on difficult, near-distribution outliers
This paper has essentially no theory, but proposes lots of interesting directions for future work.
http://arxiv.org/abs/1906.12340
This paper shows findings about combine self-supervised learning with supervised learning can improve model's performance on robustness and uncertainty. It serves like a summary with experiments.
In this paper, they found self-supervision can:
This paper has essentially no theory, but proposes lots of interesting directions for future work.