Discriminative Variational Autoencoder for Continual Learning with Generative Replay

Woo-Young Kang, Cheol-Ho Han, Byoung-Tak Zhang

Abstract: Generative replay (GR) is a method to alleviate catastrophic forgetting in continual learning (CL) by generating previous task data and learning them together with the data from new tasks. In this paper, we propose discriminative variational autoencoder (DiVA) to address the GR-based CL problem. DiVA has class-wise discriminative latent embeddings by maximizing the mutual information between classes and latent variables of VAE. Thus, DiVA is directly applicable to classification and class-conditional generation which are efficient and effective properties in the GR-based CL scenario. Furthermore, we use a novel trick based on domain translation to cover natural images which is challenging to GR-based methods. As a result, DiVA achieved the competitive or higher accuracy compared to state-of-the-art algorithms in Permuted MNIST, Split MNIST, and Split CIFAR10 settings.

Keywords: Continual learning, Generative replay, Variational Autoencoder