论文题目：Interventional Contrastive Learning with Meta Semantic Regularizer
论文来源：ICML2022
一、Motivation
When the CL model is trained with full images, the performance tested in full images is better than that in foreground areas; when the CL model is trained with foreground areas, the performance tested in full images is worse than that in foreground areas. This observation reveals that back grounds in images may interfere with the model learning semantic information and their influence has not been fully eliminated. 当对比学习模型使用全部的图像进行训练的时候，其在全部图片上的测试效果是好的；当只是使用预训练的图片训练模型时，其在全部图片上的测试效果比只在预训练的数据上的测试效果差。这个发现主要由两个原因造成：其一是预训练数据集和微调数据集存在gap；其二是对比学习任务中，每一个正样本只有一个正例，这会导致存在假的负样本。
”1” represents training and testing on full images;
”2” represents training on full images and testing on foreground images;
”3” represents training and testing on foreground images;
”4” represents training on foreground images and testing on full images.
”1” represents training and testing on full images;
”2” represents training on full images and testing on foreground images.
对比上面两个图可以发现，该论文提出的模型不仅使得效果变好了，也使得gap变小了。
二、Model
其模型结构如上图所示，与传统的对比学习方法相比，其多了一个板块msr，其使用msr对原始图像进行编码，然后将其获得的权重用于修正两个随机增强的表示，最后使用对比学习计算损失。
其目标函数如下所示：
三、Data$Experiments
其数据集采用的是CIFAR和ImageNet
四、Performance
五、Conclusion
实验结果表明了该方法的优越性，其不仅在一定程度上缓解了预训练数据集与微调数据集存在gap的情况，还在多个数据集上都提升了backbone的实验效果。
六、References
[1] Qiang W, Li J, Zheng C, et al. Interventional contrastive learning with meta semantic regularizer.In ICML. PMLR, 2022: 18018-18030.
[2] Krizhevsky, A., Hinton, G., et al. Learning multiple layers of features from tiny images. 2009.
[3] Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee, 2009.
