A Novel Adaptive Data Transformation for Contrastive Learning

Although contrastive learning has been playing a critical role in pattern recognition, how to optimize positive pairs through data transformation is still not well developed up to now. In this paper, we propose a novel Adaptive Data Transformation, named ADTrans, to identify an optimal sequence of data transformations, which enables generating high-quality positive pairs adaptively during contrastive training. Extensive experiments on benchmark datasets have shown that ADTrans can improve the performance of representation learning on downstream tasks significantly, including image classification, instance segmentation, and object detection. It can achieve a classification accuracy of 12% and 9% higher than the existing MOCO v2, SimSiam, and BYOL on the STL 10 and TinyImageNet datasets, respectively, with the ResNet-18 backbone. Moreover, it outperforms MOCO v2 on COCO instance segmentation, object detection, and Pascal VOC instance segmentation.