# Generative Models from the perspective of Continual Learning

## TL;DR

This paper evaluates various generative models and strategies for continual learning across multiple image datasets, finding that GANs with generative replay perform best, though challenges remain with complex datasets like CIFAR10.

## Contribution

It provides a comprehensive comparison of generative models and continual learning strategies on standard benchmarks, highlighting the effectiveness of GANs with replay methods.

## Key findings

- GANs outperform other models in continual learning tasks
- Generative replay strategy yields the best results among tested methods
- Training on CIFAR10 remains unstable and challenging

## Abstract

Which generative model is the most suitable for Continual Learning? This paper aims at evaluating and comparing generative models on disjoint sequential image generation tasks. We investigate how several models learn and forget, considering various strategies: rehearsal, regularization, generative replay and fine-tuning. We used two quantitative metrics to estimate the generation quality and memory ability. We experiment with sequential tasks on three commonly used benchmarks for Continual Learning (MNIST, Fashion MNIST and CIFAR10). We found that among all models, the original GAN performs best and among Continual Learning strategies, generative replay outperforms all other methods. Even if we found satisfactory combinations on MNIST and Fashion MNIST, training generative models sequentially on CIFAR10 is particularly instable, and remains a challenge. Our code is available online \footnote{\url{https://github.com/TLESORT/Generative\_Continual\_Learning}}.