Ultimate Negative Sampling for Contrastive Learning

Unsupervised learning has received more attention due to the superior performance of contrastive learning methods. Most contrastive methods use data augmentation techniques to construct positive and negative pairs. The augmented view of the same sample is regarded as a positive sample, while the rest are negative samples. This negative sampling strategy has strong randomness and ignores samples that are semantically similar to anchors, namely sampling bias. This problem has been addressed by weighting the similarity of negative samples. In this paper, we propose a novel ultimate negative sampling for contrastive learning. Unlike random sampling, we set a more extreme negative sample selection mechanism based on the ideal representation of the sample. Furthermore, we constrain the consistency between samples across the space. Experiment results demonstrate the proposed method’s superiority on multiple benchmark datasets.