Less Labeled Data? Here's the Solution: The SimCLRv2

The complication of learning information from only a few labeled data has troubled machine learning researchers for a long time, especially in applications of computer vision. To tackle this problem, a new research shows promising solution. What's new: A Google Brain Team led by Ting Chen and other fellow colleagues have formulated a simple framework for semi-supervised learning, which utilises very few labeled data and a large amount of unlabeled data to perform classification on the ImageNet database with an accuracy that outperforms the standard supervised training. Key insight: Semi-supervised learning which involves unsupervised pretraining followed by supervised fine-tuning has been copiously used for natural language processing, however, their application in computer vision has shown propitious results only very recently. The researchers carried forward this idea for use in computer vision by developing an improved variant of a previously proposed contrastive learning framework, SimCLR.