arXiv 2019

Unsupervised Data Augmentation

arXiv 2019 google-research/uda

Unlike previous methods that use random noise such as Gaussian noise or dropout noise, UDA has a small twist in that it makes use of harder and more realistic noise generated by state-of-the-art data augmentation methods.

DATA AUGMENTATION SENTIMENT ANALYSIS TEXT CLASSIFICATION

Unsupervised Data Augmentation for Consistency Training

arXiv 2019 google-research/uda

In this work, we present a new perspective on how to effectively noise unlabeled examples and argue that the quality of noising, specifically those produced by advanced data augmentation methods, plays a crucial role in semi-supervised learning.

DATA AUGMENTATION SEMI-SUPERVISED IMAGE CLASSIFICATION TEXT CLASSIFICATION TRANSFER LEARNING