A paper proposing a regularization method for neural networks that generates synthetic samples according to the training status of a classifier by meta-optimizing the generative model and uses them to train a feature extractor has been accepted to the Asian Conference on Machine Learning (ACML 2023):

Shin'ya Yamaguchi.
Generative Semi-supervised Learning with Meta-Optimized Synthetic Samples.
In Proceedings of the 15th Asian Conference on Machine Learning (ACML), 2023.