Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning

TL;DR

This work addresses the challenge of accurately estimating noisy class posteriors in noisy label learning, which is critical for learning classifier-consistent models. It introduces Part-Level Multi-labeling (PLM), which augments supervision by generating part-level labels through instance cropping and couples them with a novel single-to-multiple transition matrix within a label joint training framework, yielding the joint networks $g^u$ and $g^e$ and the relation $g^p(\boldsymbol{x})=g^u(\boldsymbol{x}) g^e(\boldsymbol{x})$. The training objective combines a standard classification loss on the noisy posterior with a multi-label loss on the part-level labels, driving the model to integrate diverse parts and mitigate overemphasis on misleading features. Extensive experiments on synthetic and real-world noisy datasets show that PLM improves estimation of $P(\tilde{Y}|\boldsymbol{x})$ and overall NLL performance, and can enhance existing transition-matrix based methods when integrated as an auxiliary component.

Abstract

In noisy label learning, estimating noisy class posteriors plays a fundamental role for developing consistent classifiers, as it forms the basis for estimating clean class posteriors and the transition matrix. Existing methods typically learn noisy class posteriors by training a classification model with noisy labels. However, when labels are incorrect, these models may be misled to overemphasize the feature parts that do not reflect the instance characteristics, resulting in significant errors in estimating noisy class posteriors. To address this issue, this paper proposes to augment the supervised information with part-level labels, encouraging the model to focus on and integrate richer information from various parts. Specifically, our method first partitions features into distinct parts by cropping instances, yielding part-level labels associated with these various parts. Subsequently, we introduce a novel single-to-multiple transition matrix to model the relationship between the noisy and part-level labels, which incorporates part-level labels into a classifier-consistent framework. Utilizing this framework with part-level labels, we can learn the noisy class posteriors more precisely by guiding the model to integrate information from various parts, ultimately improving the classification performance. Our method is theoretically sound, while experiments show that it is empirically effective in synthetic and real-world noisy benchmarks.

Figures (2)

• Figure 1: Illustration of overemphasis that arises when learning noisy class posterior with classification loss (Figure \ref{['fig:before']}), as well as the framework we proposed to alleviate this overemphasis (Figure \ref{['fig:method']}). Class activation maps are used to visualize feature importance for estimation, where the highlighted areas (with stronger red intensity) represent the focus regions of the model.
• Figure 2: Mean estimation errors of noise class posterior for 5 trials on Mnist, CIFAR-10 and CIFAR-100. The error bars of standard deviation are shaded in each plot. The lower the better.