Class incremental learning with task-selective autoencoder

Abstract

With the recent rapid developments of deep learning, deep neural networks have been applied to various computer vision tasks such as object classification, detection, and segmentation, and have shown superior performance. Deep neural networks, however, suffers from the problem called "catastrophic forgetting" that is a tendency of forgetting previously learned tasks when deep neural networks learn new tasks. Catastrophic forgetting is considered as a fundamental limitation of deep neural networks. Knowledge distillation-based methods have been proposed to alleviate catastrophic forgetting problem, but the performance is still not good. In this paper, we propose a Task-selective Autoencoder (TsAE) to improve the performance of class incremental learning of the conventional knowledge distillation-based method. The proposed Task-selective Autoencoder has a simple architecture and a small number of trainable parameters. The experimental results show that the application of Task-selective autoencoder to the existing knowledge distillation-based methods effectively improves classification performance.