NCMatch: Semi-supervised Learning with Noisy Labels via Noisy Sample Filter and Contrastive Learning

Abstract

Semi-supervised learning (SSL) has been widely studied in recent years, which aims to improve the performance of supervised learning by utilizing unlabeled data. However, the presence of noisy labels on labeled data is an inevitable consequence of either limited expertise in the labeling process or inadvertent labeling errors. While these noisy labels can have a detrimental impact on the performance of the model, leading to decreased accuracy and reliability. In this paper, we introduce the paradigm of Semi-Supervised Learning with Noisy Labels (SSLNL), which aims to address the challenges of semi-supervised classification when confronted with limited labeled data and the presence of label noise. To address the challenges of SSLNL, we propose the Noisy Samples Filter (NSF) module, which effectively filters out noisy samples by leveraging class agreement. Additionally, we propose the Semi-Supervised Contrastive Learning (SSCL) module, which harnesses the power of high-confidence pseudo-labels generated by a semi-supervised model to enhance the extraction of robust features through contrastive learning. Extensive experiments on CIFAR-10, CIFAR-100, and SVHN datasets demonstrate that our proposed method, outperforms state-of-the-art methods in addressing the problems of SSL and SSLNL, thus validating the effectiveness of our solution.