Mixing augmentations like mixup, cutmix, augmix and so on shows a promising improvement in all machine learning aspects.

In the semi supervised regime, SEMI-VIT proposed “probabilistic pseudo mixup”

The goal of this project is to try to understand if mixing labeled and unlabeled data helps generalization.

Since SSL requires scedualing for thresholding in most cases, it could lead to better understanding about a curriculum augmentation strategy, similar to CUDA.

<img src=’/images/mixing.png’,width=”600”/>
image from https://arxiv.org/pdf/2212.10888