Logit-driven data generation as a data augmentation procedure

Dr. Venkatesan wagers Dr. Zhou a lunch meal of the winner's choosing that a peer-reviewed article will be published by the end of the year 2024, that demonstrates the use of data, generated from a teacher network conditioned on the logits created in a fashion similar to mix-up for augmenting the dataset during the training of a student network.

The supposition here is that if in mix-up the authors were able to mix labels and interpolate data and treat this new combination of label-data as supervision, then it must also be possible to use a discriminator to create a logit and generate an image, use this logit-data pair as supervision to augment a supervised dataset.