NeurIPS 2020

GAN Memory with No Forgetting

Meta Review

The paper presents an GAN-based method to learn from multiple datasets sequentially. To adapt to multiple datasets, the authors propose modulation tricks from style transfer, e.g., FiLM and AdaFM, which allows for generating different types of data without forgetting. The reviewer all agree on acceptance for the following reasons: interesting idea for incremental learning, addressing the catastrophic forgetting issue with a reasonable approach, strong experimental results, etc. This AC agrees to accept the paper.