S2OSC: A Holistic Semi-Supervised Approach for Open Set Classification

Open set classification (OSC) tackles the problem of determining whether data are in-class or out-of-class during inference, when only provided with a examples at training time. Traditional OSC methods usually train discriminative generative models owned data, and then utilize pre-trained to classify test directly. However, these always suffer from embedding confusion problem, i.e., partial instances mixed ones similar semantics, making it difficult classify. To solve this we unify semi-supervised learning develop novel algorithm, S2OSC, which incorporates filtering model re-training in transductive manner. In detail, given pool newly coming S2OSC firstly filters mostly distinct using model, annotates super-class for them. Then, trains holistic by combing labeled remaining unlabeled paradigm. Furthermore, considering that streaming form real applications, extend into an incremental update framework (I-S2OSC), adopt knowledge memory regularization mitigate catastrophic forgetting update. Despite simplicity proposed models, experimental results show achieves state-of-the-art performance across variety tasks, including 85.4% F1 on CIFAR-10 300 pseudo-labels. We also demonstrate how can be expanded setting effectively data.