Mode-Aware GAN: Continual Adaptation for Conditional Image Generation

TMLR Paper4110 Authors

Abstract: Continuously learning new modes in generative models while preserving previously learned ones is a significant challenge, particularly with limited training samples. Here, we propose a Mode Affinity Score tailored for continual learning within conditional generative adversarial networks. This score, derived from the discriminators, measures the similarity between generative tasks. By leveraging this score, new modes can be seamlessly integrated into the model through an interpolation process among the closest learned modes, guided by the computed affinity scores. This approach enhances generation performance and mitigates the risk of catastrophic forgetting. Extensive experiments demonstrate the efficacy of our method compared to existing techniques, even when using significantly fewer training samples.

Submission Length: Regular submission (no more than 12 pages of main content)

Previous TMLR Submission Url: https://openreview.net/forum?id=X1ZAZQAHmp&referrer=%5Bthe%20profile%20of%20Cat%20P.%20Le%5D(%2Fprofile%3Fid%3D~Cat_P._Le1)

Changes Since Last Submission: The font has been fixed.

Assigned Action Editor: ~Piyush_Rai1

Submission Number: 4110