Unsupervised Visual Representation Learning via Mutual Information Regularized Assignment