董浩華

Make two other auxiliary tasks and both rely on self-generated labels. -Treat all images in each case as one class -Channel shuffle Pre-training using the feature extractor after self-supervised learning is 4% more accurate than the ImageNet-based pre-trained model.