Integrating complex augmentations like Mixup into generative classifiers

Develop methods to incorporate complex data augmentations, specifically Mixup, into class-conditional generative classifiers while preserving or improving their performance and robustness characteristics.

Background

The authors highlight practical challenges for deep generative classifiers, including inference cost and training workflows. While discriminative classifiers commonly leverage augmentations such as Mixup to improve generalization, the paper notes that it is unclear how to adapt such augmentations to the generative classification paradigm.

Addressing this gap requires designing augmentation strategies compatible with learning class-conditional likelihoods p(x|y), ensuring they integrate coherently with diffusion or autoregressive training objectives and maintaining the generative classifier’s strengths under distribution shift.