Memory Replay GANs: learning to generate images from new categories without forgetting

TL;DR

This paper introduces Memory Replay GANs, a novel framework that enables generative adversarial networks to learn new categories sequentially while effectively mitigating forgetting of previously learned categories.

Contribution

The paper proposes Memory Replay GANs with a memory replay generator and two methods to prevent forgetting, advancing sequential learning in generative models.

Findings

Memory Replay GANs generate high-quality images for new categories.

The approach significantly reduces forgetting of previous categories.

Experimental results outperform baseline methods on multiple datasets.

Abstract

Previous works on sequential learning address the problem of forgetting in discriminative models. In this paper we consider the case of generative models. In particular, we investigate generative adversarial networks (GANs) in the task of learning new categories in a sequential fashion. We first show that sequential fine tuning renders the network unable to properly generate images from previous categories (i.e. forgetting). Addressing this problem, we propose Memory Replay GANs (MeRGANs), a conditional GAN framework that integrates a memory replay generator. We study two methods to prevent forgetting by leveraging these replays, namely joint training with replay and replay alignment. Qualitative and quantitative experimental results in MNIST, SVHN and LSUN datasets show that our memory replay approach can generate competitive images while significantly mitigating the forgetting of previous categories.