End-to-end novel visual categories learning via auxiliary self-supervision

Semi-supervised learning has largely alleviated the strong demand for large amount of annotations in deep learning. However, most methods have adopted a common assumption that there is always labeled data from same class unlabeled data, which impractical and restricted real-world applications. In this research work, our focus on semi-supervised when categories are disjoint each other. The main challenge how to effectively leverage knowledge they independent other, not belonging categories. Previous state-of-the-art proposed construct pairwise similarity pseudo labels as supervising signals. two issues commonly inherent these methods: (1) All previous comprised multiple training phases, makes it difficult train model an end-to-end fashion. (2) Strong dependence quality limits performance vulnerable noise bias. Therefore, we propose exploit use self-supervision auxiliary task during such will share set surrogate overall signals can regularization. By doing so, all modules algorithm be trained simultaneously, boost capability achieved. Moreover, utilize local structure information feature space label construction, properties more robust noise. Extensive experiments been conducted three frequently used visual datasets, i.e., CIFAR-10, CIFAR-100 SVHN, paper. Experiment results indicated effectiveness achieved new novel datasets.