Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

IRNet: Iterative Refinement Network for Noisy Partial Label Learning

Zheng Lian, Mingyu Xu, Lan Chen, Licai Sun, Bin Liu, Lei Feng, Jianhua Tao

TL;DR

This work extends partial label learning (PLL) to the more challenging noisy PLL setting where the ground-truth label may lie outside the candidate set. It introduces Iterative Refinement Network (IRNet), a plug-in framework with two modules—noisy sample detection and label correction—augmented by smoothness constraints to stabilize learning and improve reliability. The authors provide theoretical analysis showing that multi-round refinement reduces dataset noise and approaches the Bayes optimal classifier, and demonstrate consistent empirical gains across CIFAR-10, CIFAR-100, Kuzushiji-MNIST, and real-world RAF-DB benchmarks. IRNet also offers practical benefits by being compatible with existing PLL methods and exhibiting robustness to parameter choices and augmentation strategies.

Abstract

Partial label learning (PLL) is a typical weakly supervised learning, where each sample is associated with a set of candidate labels. Its basic assumption is that the ground-truth label must be in the candidate set, but this assumption may not be satisfied due to the unprofessional judgment of annotators. Therefore, we relax this assumption and focus on a more general task, noisy PLL, where the ground-truth label may not exist in the candidate set. To address this challenging task, we propose a novel framework called ``Iterative Refinement Network (IRNet)'', aiming to purify noisy samples through two key modules (i.e., noisy sample detection and label correction). To achieve better performance, we exploit smoothness constraints to reduce prediction errors in these modules. Through theoretical analysis, we prove that IRNet is able to reduce the noise level of the dataset and eventually approximate the Bayes optimal classifier. Meanwhile, IRNet is a plug-in strategy that can be integrated with existing PLL approaches. Experimental results on multiple benchmark datasets show that IRNet outperforms state-of-the-art approaches on noisy PLL. Our source code is available at: https://github.com/zeroQiaoba/IRNet.

IRNet: Iterative Refinement Network for Noisy Partial Label Learning

TL;DR

This work extends partial label learning (PLL) to the more challenging noisy PLL setting where the ground-truth label may lie outside the candidate set. It introduces Iterative Refinement Network (IRNet), a plug-in framework with two modules—noisy sample detection and label correction—augmented by smoothness constraints to stabilize learning and improve reliability. The authors provide theoretical analysis showing that multi-round refinement reduces dataset noise and approaches the Bayes optimal classifier, and demonstrate consistent empirical gains across CIFAR-10, CIFAR-100, Kuzushiji-MNIST, and real-world RAF-DB benchmarks. IRNet also offers practical benefits by being compatible with existing PLL methods and exhibiting robustness to parameter choices and augmentation strategies.

Abstract

Partial label learning (PLL) is a typical weakly supervised learning, where each sample is associated with a set of candidate labels. Its basic assumption is that the ground-truth label must be in the candidate set, but this assumption may not be satisfied due to the unprofessional judgment of annotators. Therefore, we relax this assumption and focus on a more general task, noisy PLL, where the ground-truth label may not exist in the candidate set. To address this challenging task, we propose a novel framework called ``Iterative Refinement Network (IRNet)'', aiming to purify noisy samples through two key modules (i.e., noisy sample detection and label correction). To achieve better performance, we exploit smoothness constraints to reduce prediction errors in these modules. Through theoretical analysis, we prove that IRNet is able to reduce the noise level of the dataset and eventually approximate the Bayes optimal classifier. Meanwhile, IRNet is a plug-in strategy that can be integrated with existing PLL approaches. Experimental results on multiple benchmark datasets show that IRNet outperforms state-of-the-art approaches on noisy PLL. Our source code is available at: https://github.com/zeroQiaoba/IRNet.
Paper Structure (30 sections, 2 theorems, 37 equations, 9 figures, 8 tables)

This paper contains 30 sections, 2 theorems, 37 equations, 9 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Non-deep Partial Label Learning
  4. Deep Partial Label Learning
  5. Noisy Partial Label Learning
  6. Learning from Noisy Labels
  7. Methodology
  8. Problem Definition
  9. Observation and Motivation
  10. Noisy Sample Detection
  11. Label Correction
  12. Iterative Refinement Network (IRNet)
  13. Choice of Correction Epoch
  14. Noisy Sample Detection
  15. Label Correction
  16. ...and 15 more sections

Key Result

Theorem 1

Assume that there exists a boundary $\epsilon<m<1$ such that $L(m)$ is pure for $(f,S)$. Let $\epsilon<m_{\text{new}}<1$ be the new boundary. For all $x \in \mathcal{D}_{S}$ satisfying $max_{j \notin S(x)}f_j(x) - max_{j \in S(x)} f_j(x)\geq m_{\text{new}}-\epsilon$, we move $argmax_{j \notin S(x)}f $L(m_{\text{new}})$ is pure for $(f^{\text{new}},S_{\text{new}})$ when $m_{\text{new}}$ satisfies:

Figures (9)

  • Figure 1: Typical applications of noisy PLL. (a) Candidate labels can be provided by crowdsourcing. However, the ground-truth label may not be in the candidate set. (b) Candidate names can be extracted from the text caption. However, there are general cases of faces without names.
  • Figure 2: IRNet contains two core modules: noisy sample detection and label correction. First, we use the metric $\tau$ (see Eq. \ref{['eq-tau']}) to identify noisy samples. Then, we use the non-candidate label with the highest probability to update the candidate set for the detected noisy samples.
  • Figure 3: Visualization of the Bayes error rate, hit accuracy, and validation accuracy on CIFAR-10 ($q=0.3, \eta=0.3$). We mark the minima in (a), the maxima in (b), and the local maxima in (c) with red dots.
  • Figure 4: Empirical PDF and estimated GMM models on CIFAR-10 ($q=0.3, \eta=0.3$) with increasing training epochs.
  • Figure 5: Transductive performance under different ambiguity levels. "w/" indicates the model with IRNet. The noise level is fixed to $\eta=0.3$.
  • ...and 4 more figures

Theorems & Definitions (3)

  • Definition 1: Pure $(m,f,S)$-level set
  • Theorem 1: One-round refinement
  • Theorem 2: Multi-round refinement