I have just come across the idea of self-supervised learning. It seems that it is possible to get higher accuracies on downstream tasks when the network is trained on pretext tasks. Suppose that I want to do image classification on my own set of images. I have limited data on these images and may...