Augmented Transport Data-Free Class Incremental Learning

Although deep neural networks exhibit high performance on several specific tasks, they suffer from a severe issue of catastrophic forgetting when learning new tasks incrementally. Several incremental learning methods have been recently proposed to address this problem, some of which rely on storing data or constructing generative models to enhance the performance of incremental models. However, storing old data from previous tasks may be limited by memory or privacy issues, and generative models may often be unstable and inefficient during model training. In this study, we present a non-exemplar-based class-incremental learning method to mitigate catastrophic forgetting in incremental learning when saving old class samples is not possible, while distinguishing between old and new classes. We tackle the problem of representation bias and classifier bias that are inherent in class-incremental learning. Specifically, we introduce class enhancement, which enables the model to encounter more classes during training to acquire transferable and diverse representation features, mitigating representation bias. Furthermore, we use backtracking transfer, which employs semantic mapping to limit inter-class relationships by learning class-level semantic relationships across two tasks, to preserve discriminative old knowledge and prevent catastrophic forgetting due to classifier bias. We perform comprehensive experiments on three benchmark datasets that demonstrate the superiority of our proposed method over state-of-the-art methods.