文章介绍
彭玺团队的一篇发表于PAMI上的文章,附源码github-TPAMI-DCP
文章主要贡献
- We provide a novel insight to the community that the cross-view consistency learning and data recovery are with intrinsic connections in the framework of information theory. Such a theoretical framework is remarkably different from existing MvRL studies which treat the consistency learning and data recovery as two separate problems. (把consistency 和 data recovery 联系到了一块)
- Under our information theoretical framework, we propose DCP which achieves the information consistency and data recoverability through a dual contrastive loss and a dualprediction loss, respectively.(基于信息论框架,实现了infomation consistency 和data recovery)
- To utilize available label information, DCP designs and utilizes the instance- and category-level contrastive loss to enhance the separability of representations. (该框架利用了实现了不同表示之间的区分程度(separability of representations) 利用实例间的对比和类别之间的对比)
- We theoretically and experimentally prove that DCP could learn a sufficient and minimal representation for three tasks, i.e., clustering, classification, and human action recognition. (可以具体实现三个任务,聚类、分类、人类行为的辨别)
文章框架
分三个主要模块:
1、Whin-view Reconstruction : 每个视图的数据分别投影到一个低维的子空间,从而学习每个视图的特定表示。
2、Dual Prediction : 作用是数据恢复作用。
3、Contrastive Learning : 分两个子模块,一个是Instance-Level Contrastive Learning(作用是通过最大化不同视图表示之间的互信息,该损失函数旨在增强跨视图的一致性),一个是Category-level Contrastive Learing(作用是使来自同一类别的样本在表示空间中更接近,而来自不同类别的样本更远)
损失函数
1、总损失函数
2、对比学习损失
contrastive loss
instance-contrastive-learning loss
category-contrastive-learning loss
3、对比预测损失
实验
Table 1 是聚类性能,Table 2 是分类性能