In real-world classification tasks, it is difficult to collect trainingsamples from all possible categories of the environment. Therefore, when aninstance of an unseen class appears in the prediction stage, a robustclassifier should be able to tell that it is from an unseen class, instead ofclassifying it to be any known category. In this paper, adopting the idea ofadversarial learning, we propose the ASG framework for open-categoryclassification. ASG generates positive and negative samples of seen categoriesin the unsupervised manner via an adversarial learning strategy. With thegenerated samples, ASG then learns to tell seen from unseen in the supervisedmanner. Experiments performed on several datasets show the effectiveness ofASG.
展开▼
