A study of contrastive self-supervised learning generalization based on augmented data

Contrastive learning have been widely applied recently, since only with unlabeled data, it produced results comparable to the state-of-the-art supervised algorithm. Data augmentation plays an essential role in defining effective predictive task, influencing generalization of the model. We find that high-intensity data augmentation is conducive to achieving a wider aggregation of intra-class samples, obtaining more effective features to improve generalization; however, extremely strong augmentation leads to the information loss of different classes, resulting in cross-class deviation of samples, reducing model generalization. We demonstrate our finding through experiment on CIFAR-10 dataset, analyzing the impact of different data augmentation operations on recognition results and the corresponding reasons. Our experimental verification on the relationship between the data augmentation and the generalization of contrastive self-supervised learning agree with our theoretical description.