Icml2022 | interventional contrastive learning based on meta semantic regularization

Thesis link ：https://arxiv.org/abs/2206.14702

Based on comparative learning (CL) The self supervised learning model of learning visual representations in pairs . Although the current popular CL The model has made great progress , But in this article , We found a phenomenon that has been neglected : When using full image training CL Model time , The performance tested in the full image is better than that tested in the foreground area ; When using foreground area training CL Model time , The performance tested in the complete image is worse than that tested in the foreground area . This observation shows that , The background in the image may interfere with the semantic information of model learning , Its influence has not been completely eliminated . To solve this problem , We build a structural causal model (SCM), Modeling the background as a confounding agent . We propose a regularization method based on backdoor adjustment , That is, interventional contrastive learning based on meta semantic regularization (ICLMSR), Conduct causal intervention on the proposed supply chain management .ICL-MSR Can be integrated into any existing CL In the method , To reduce the background interference of representation learning . It is proved theoretically ICL-MSR It has a smaller error bound . Empirically , Our experiments on several benchmark datasets show that ,ICL-MSR Can improve the most advanced CL Method performance .