Prototypical contrastive learning for image classification

Contrastive learning has become one of the most important representation learning methods because it does not require data to be labeled. However, current contrast learning treats different negative instances equally as being pushed away evenly in the representation space. Intuitively, the impact of different classes of instances in the representation space is different. Therefore, in this paper, we combine contrastive learning and clustering to propose a prototypical contrastive learning (ProCL) for image classification. Specifically, ProCL performs representation learning by clustering semantically similar images into the same group and encouraging clustering consistency between different augmentations of the same image. For a given image, sampling is performed from different clusters in order to ensure semantic differences from negative samples. Moreover, the gaps in semantic information between the prototypes (clustering center) differ, ProCL further weights the negative samples according to the distance between the prototypes, so that those negative samples with appropriate prototype distances have larger weights. This weighting strategy proves to be more effective. The experimental results in several benchmarks demonstrate that ProCL has strong competitive performance.