Integrate Countersign Justification with airSlate SignNow

Integrate countersign justification

data augmentation is usually used to form new examples that share the same class label as the original instance however we also know of data augmentations that either corrupt the class label or particularly push the image to be out of the distribution of say natural images an example of this is the jigsaw augmentation where you patch up an image and then scramble the patches so you destroy the global structure of the image this paper explores these kinds of augmentations termed as negative data augmentations and sees how we can integrate them into our training loops to make use of these kinds of priors as well the authors show very interesting and exciting results with training gans and contrastive learning frameworks such as contrastive predictive coding and moco i think this is a really exciting strategy to extend the use of data augmentation and further improve the ability to learn with less labeled data [Music] this video will explain negative data augmentation the high level motivation behind negative data augmentation is that most of these papers on data augmentation and discussing different ways to transform our existing data examples are focused on this concept of label preserving or semantically preserving transformations we have images of things like flowers dogs cats or we have text sequences or we have any kind of data domain and we're trying to find transformations that give us more data but preserve the class label of our original data instance so a lot of the problems that are related to data augmentation are with this difficulty of trying to find these transformations for data domains outside of images that are semantically preserving or label preserving so this idea of negative data augmentation is saying well what about the the transformations that aren't label preserving how can we still use those because those still have an interesting prior about what is out of the support of the data distribution so the motivating question behind this research paper is is there a use for intentionally out of distribution data augmentations and shown here in image augmentation this is a great example of this we have our original image of a dog and then we have the jigsaw augmentation so the jigsaw augmentation randomly crops an image and then shuffles the ordering of the patches from the crops so that you have this new potential image pixel grid for learning from and this is completely out of distribution for natural images there's no dog image that would be organized like this so this is an intentionally out of distribution kind of data augmentation and this paper negative data augmentation is going to integrate these kind of augmentations into contrastive learning and generative adversarial networks to see how we can leverage these data augmentations which i think is a really exciting direction because it's much easier to find uh or i don't know about much easier because i haven't really thought about this too much but it does...