Contrastive learning with auxiliary model

Abstract

Contrastive Learning (CL) achieves excellent performance to reduce the requirements of annotation in many domains recently. Most CL methods treat all samples other than an anchor as negative samples and move them far away in training. However, these samples may be semantically similar and should belong to the same class. Recent false negative sample identification methods only rely on a unique embedding space, and the decisions may be biased. This study proposes a false negative sample identification method with an auxiliary model shared with the backbone of the main CL model in order to provide an additional point of view on the decisions. The auxiliary model is trained by minimizing the standard contrastive loss while maximizing diversity with the main model. False negative samples are identified by the fusion of the similarity measures in both feature spaces learned by the two models. The standard datasets are used to evaluate the effectiveness of our method. Experimental results confirm that our method achieves more reliable and accurate performance compared with the state-of-the-art CL methods. Using auxiliary models improves the false negative sample identification significantly. Our proposed false negative sample identification method improves the accuracy and stability of contrastive learning. Code is available at https://github.com/czzerone/Dual-Contrastive-Learning.

Lab Members

Patrick Chan

Associate Professor, Vice Dean

Patrick Chan works on machine learning, deep learning, image processing, adversarial learning, and secure machine learning.

Xuanming Liang 梁轩铭

My research interests include large language model, Model Safety, Aaget applications.