Abstract: Traditional contrastive learning algorithms have relied on instance discrimination, which is easy to introduce false negative samples, resulting in algorithm converges to suboptimal solutions and degenerating the performance of downstream tasks. To address this issue, this paper proposes a multi-branch contrastive learning algorithm based on relationship consistency. This algorithm incorporates a nearest neighbor set into the branch network, providing positive samples with their semantic consistency and avoiding false negative samples. Combined with multi-branch training using data augmentation, it minimizes KL divergence to pull positive samples with semantic consistency together and push away negative samples, which enhances the representation ability of network. The temperature parameters in different branches control the smoothness of the sharpen distribution, ensuring the reliability of feature representations. Finally, the method is evaluated on five public datasets and compared with state-of-the-art contrastive learning methods, all achieving satisfactory results.

Key words: contrastive learning, relationship consistency, feature representation, false negative sample, data augmentation