Twin Contrastive Learning with Noisy Labels

TL;DR

This paper introduces TCL, a twin contrastive learning approach that effectively handles noisy labels by modeling data with GMMs, detecting out-of-distribution examples, and using cross-supervision to improve representation learning, achieving significant performance gains.

Contribution

TCL is a novel contrastive learning framework that explicitly models label noise and out-of-distribution detection using GMMs and entropy regularization, improving robustness in noisy label scenarios.

Findings

TCL achieves 7.5% accuracy improvement on CIFAR-10 with 90% noisy labels.

TCL outperforms existing methods on multiple benchmarks.

TCL effectively detects noisy and out-of-distribution samples.

Abstract

Learning from noisy data is a challenging task that significantly degenerates the model performance. In this paper, we present TCL, a novel twin contrastive learning model to learn robust representations and handle noisy labels for classification. Specifically, we construct a Gaussian mixture model (GMM) over the representations by injecting the supervised model predictions into GMM to link label-free latent variables in GMM with label-noisy annotations. Then, TCL detects the examples with wrong labels as the out-of-distribution examples by another two-component GMM, taking into account the data distribution. We further propose a cross-supervision with an entropy regularization loss that bootstraps the true targets from model predictions to handle the noisy labels. As a result, TCL can learn discriminative representations aligned with estimated labels through mixup and contrastive learning. Extensive experimental results on several standard benchmarks and real-world datasets demonstrate the superior performance of TCL. In particular, TCL achieves 7.5\% improvements on CIFAR-10 with 90\% noisy label -- an extremely noisy scenario. The source code is available at \url{https://github.com/Hzzone/TCL}.