https://arxiv.org/abs/2012.13493
Self-supervised Pre-training with Hard Examples Improves Visual Representations (Chunyuan Li, Xiujun Li, Lei Zhang, Baolin Peng, Mingyuan Zhou, Jianfeng Gao)
self supervision을 pseudo label 예측 문제로 생각해서 label이 필요한 augmentation인 adversarial training과 cutmix를 self supervision에 결합.
#self_supervised #mixup