<p>Label refinement methods are designed to improve the quality of training labels by incorporating model predictions into the original training labels. By adjusting the combination coefficient of the noisy label, the impact of noise is reduced, which in turn makes the training process more robust. However, previous label refinement methods are unable to model instance-dependent noise, which is the most realistic type of noise. To address this limitation, we propose a simple approach, probabilistic instance-dependent label refinement (referred to as <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10994_2024_6668_Article_IEq1.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="13" /> </InlineMediaObject> <EquationSource Format="TEX">$\pi$</EquationSource> <EquationSource Format="MATHML"><math> <mi>π</mi> </math></EquationSource> </InlineEquation>-LR). Inspired by the fact that humans are more likely to make mistakes when annotating confusing instances, we propose to estimate the probability of whether a sample is confusing, which can be useful for modeling noise generation. Our approach exploits this concept by assigning a confusing probability <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10994_2024_6668_Article_IEq2.gif" Format="GIF" Height="12" Rendition="HTML" Resolution="72" Type="Linedraw" Width="17" /> </InlineMediaObject> <EquationSource Format="TEX">$\eta _i$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>η</mi> <mi>i</mi> </msub> </math></EquationSource> </InlineEquation> to each instance <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10994_2024_6668_Article_IEq3.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$\varvec{x}_i$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mrow> <mi mathvariant="bold-italic">x</mi> </mrow> <mi>i</mi> </msub> </math></EquationSource> </InlineEquation> from a probabilistic perspective. This provides a clear understanding of how instance-dependent noise affects true labels. Empirical evaluations show that <InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10994_2024_6668_Article_IEq4.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="13" /> </InlineMediaObject> <EquationSource Format="TEX">$\pi$</EquationSource> <EquationSource Format="MATHML"><math> <mi>π</mi> </math></EquationSource> </InlineEquation>-LR improves the robustness of the model in the presence of label noise and outperforms all compared methods on both realistic and synthetic label noise, while maintaining high efficiency in time and space.</p>

Probabilistic instance dependent label refinement for noisy label learning

期刊论文
发表日期：

2025年3月17日

查看全文

Hao-Yuan He,
Yu Liu,
Ren-Biao Liu,
Zheng Xie,
Ming Li

摘要

Label refinement methods are designed to improve the quality of training labels by incorporating model predictions into the original training labels. By adjusting the combination coefficient of the noisy label, the impact of noise is reduced, which in turn makes the training process more robust. However, previous label refinement methods are unable to model instance-dependent noise, which is the most realistic type of noise. To address this limitation, we propose a simple approach, probabilistic instance-dependent label refinement (referred to as $\pi$ $π$ -LR). Inspired by the fact that humans are more likely to make mistakes when annotating confusing instances, we propose to estimate the probability of whether a sample is confusing, which can be useful for modeling noise generation. Our approach exploits this concept by assigning a confusing probability $\eta _i$ $η_{i}$ to each instance $\varvec{x}_i$ $x_{i}$ from a probabilistic perspective. This provides a clear understanding of how instance-dependent noise affects true labels. Empirical evaluations show that $\pi$ $π$ -LR improves the robustness of the model in the presence of label noise and outperforms all compared methods on both realistic and synthetic label noise, while maintaining high efficiency in time and space.