Cross-Class Feature Augmentation for Class Incremental Learning

We propose a novel class incremental learning approach by incorporating afeature augmentation technique motivated by adversarial attacks. We employ aclassifier learned in the past to complement training examples rather thansimply play a role as a teacher for knowledge distillation towards subsequentmodels. The proposed approach has a unique perspective to utilize the previousknowledge in class incremental learning since it augments features of arbitrarytarget classes using examples in other classes via adversarial attacks on apreviously learned classifier. By allowing the cross-class featureaugmentations, each class in the old tasks conveniently populates samples inthe feature space, which alleviates the collapse of the decision boundariescaused by sample deficiency for the previous tasks, especially when the numberof stored exemplars is small. This idea can be easily incorporated intoexisting class incremental learning algorithms without any architecturemodification. Extensive experiments on the standard benchmarks show that ourmethod consistently outperforms existing class incremental learning methods bysignificant margins in various scenarios, especially under an environment withan extremely limited memory budget.