Self-Supervised Classification Network

We present Self-Classifier – a novel self-supervised end-to-end classification learning approach. learns labels and representations simultaneously in single-stage manner by optimizing for same-class prediction of two augmented views the same sample. To guarantee non-degenerate solutions (i.e., where all are assigned to class) we propose mathematically motivated variant cross-entropy loss that has uniform prior asserted on predicted labels. In our theoretical analysis, prove degenerate not set optimal is simple implement scalable. Unlike other popular unsupervised contrastive representation approaches, it does require any form pre-training, expectation-maximization, pseudo-labeling, external clustering, second network, stop-gradient operation, or negative pairs. Despite its simplicity, approach sets new state art ImageNet; even achieves comparable state-of-the-art results learning. Code available at https://github.com/elad-amrani/self-classifier .