Cross-Class Feature Augmentation for Class Incremental Learning

By leveraging the representations learned in the past, we aim to augment the features Recent deep learning techniques have shown remarkable at each incremental stage to address data deficiency in progress in various computer vision tasks including image the classes belonging to old tasks. To this end, inspired by classification (He et al. 2016; Hu, Shen, and Sun 2018), object adversarial attacks, we adjust the feature representations of detection (Liu et al. 2016; Redmon et al. 2016; Zhu et al. training examples to resemble representations from specific 2021c), semantic segmentation (Chen et al. 2017; Long, target classes that are different from their original classes. Shelhamer, and Darrell 2015; Noh, Hong, and Han 2015), These perturbed features allow a new classifier to maintain and many others. Behind this success is an implicit assumption the decision boundaries for the classes learned up to the that the whole dataset with a predefined set of classes previous stages. Note that this is a novel perspective different should be given in a batch. However, this assumption is from conventional adversarial attack methods (Carlini unlikely to hold in the real-world scenarios which change and Wagner 2017; Goodfellow, Shlens, and Szegedy 2017; dynamically over time. This limits the applicability to realworld Madry et al. 2018; Moosavi-Dezfooli, Fawzi, and Frossard problems because deep neural networks trained under 2016; Zhao, Dua, and Singh 2018), which focus on deceiving changing data distribution often suffer from catastrophic forgetting, models. One may consider generating additional features meaning that the models lose the ability to maintain for each class using the exemplars with the same class labels.