关键词: 关键词: 深度聚类, 对比学习, 语义特征, 过拟合, 正则化

Abstract: Abstract: Contrastive learning between data views has achieved remarkable success in the field of deep clustering. However, since all supervised information of one view originates from the other view, contrastive learning yields only a minimal adequate representation of the information shared by both and weakens the non-shared information between views. Considering the richness of feature representations required for image clustering, it cannot be guaranteed that all information relevant to the clustering task is included in the shared information between views, i.e., the minimum adequate representation obtained by contrastive learning is insufficient for the clustering task. Therefore, a method for optimizing semantic features is proposed, which uses reconstruction loss as a regularization term in the pre-training phase to increase the mutual information between the feature representation and the input, thus approximating the introduction of more information relevant to the clustering task and reducing the risk of overfitting shared information by contrastive learning. In the fine-tuning stage, the traditional clustering algorithm and clustering network are abandoned to be updated simultaneously, and the similarity differences between the picture nearest neighbors are used as losses to update the clustering network in order to maximize the use of the semantic information of the nearest neighbors between pictures. By experimenting the method on CIFAR10,CIFAR100,STL10 datasets, the experimental results show that the accuracy of the method is higher than all the compared methods, and it improves 2.7% over the suboptimal SCAN method on STL10 dataset, and the other evaluation indexes achieve the leading position, which verifies the effectiveness of the method.

Key words: Keywords: deep clustering, contrastive learning, semantic features, overfitting, regularization