Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TMLC-Net: Transferable Meta Label Correction for Noisy Label Learning

Mengyang Li

TL;DR

TMLC-Net addresses the challenge of noisy labels by learning a transferable label-correction strategy via a meta-learner trained across datasets. It integrates Normalized Noise Perception, Time-Series Encoding with an LSTM, and Subclass Decoding to output a probabilistic corrected label distribution, enabling robust learning without task-specific retraining. Empirical results on CIFAR-10/100, Clothing1M, and WebVision show state-of-the-art accuracy and strong transferability under symmetric, asymmetric, and real-world noise, with ablations validating each component. The approach reduces the burden of meta-label correction in large-scale settings and holds promise for robust learning under label noise across domains.

Abstract

The prevalence of noisy labels in real-world datasets poses a significant impediment to the effective deployment of deep learning models. While meta-learning strategies have emerged as a promising approach for addressing this challenge, existing methods often suffer from limited transferability and task-specific designs. This paper introduces TMLC-Net, a novel Transferable Meta-Learner for Correcting Noisy Labels, designed to overcome these limitations. TMLC-Net learns a general-purpose label correction strategy that can be readily applied across diverse datasets and model architectures without requiring extensive retraining or fine-tuning. Our approach integrates three core components: (1) Normalized Noise Perception, which captures and normalizes training dynamics to handle distribution shifts; (2) Time-Series Encoding, which models the temporal evolution of sample statistics using a recurrent neural network; and (3) Subclass Decoding, which predicts a corrected label distribution based on the learned representations. We conduct extensive experiments on benchmark datasets with various noise types and levels, demonstrating that TMLC-Net consistently outperforms state-of-the-art methods in terms of both accuracy and robustness to label noise. Furthermore, we analyze the transferability of TMLC-Net, showcasing its adaptability to new datasets and noise conditions, and establishing its potential as a broadly applicable solution for robust deep learning in noisy environments.

TMLC-Net: Transferable Meta Label Correction for Noisy Label Learning

TL;DR

TMLC-Net addresses the challenge of noisy labels by learning a transferable label-correction strategy via a meta-learner trained across datasets. It integrates Normalized Noise Perception, Time-Series Encoding with an LSTM, and Subclass Decoding to output a probabilistic corrected label distribution, enabling robust learning without task-specific retraining. Empirical results on CIFAR-10/100, Clothing1M, and WebVision show state-of-the-art accuracy and strong transferability under symmetric, asymmetric, and real-world noise, with ablations validating each component. The approach reduces the burden of meta-label correction in large-scale settings and holds promise for robust learning under label noise across domains.

Abstract

The prevalence of noisy labels in real-world datasets poses a significant impediment to the effective deployment of deep learning models. While meta-learning strategies have emerged as a promising approach for addressing this challenge, existing methods often suffer from limited transferability and task-specific designs. This paper introduces TMLC-Net, a novel Transferable Meta-Learner for Correcting Noisy Labels, designed to overcome these limitations. TMLC-Net learns a general-purpose label correction strategy that can be readily applied across diverse datasets and model architectures without requiring extensive retraining or fine-tuning. Our approach integrates three core components: (1) Normalized Noise Perception, which captures and normalizes training dynamics to handle distribution shifts; (2) Time-Series Encoding, which models the temporal evolution of sample statistics using a recurrent neural network; and (3) Subclass Decoding, which predicts a corrected label distribution based on the learned representations. We conduct extensive experiments on benchmark datasets with various noise types and levels, demonstrating that TMLC-Net consistently outperforms state-of-the-art methods in terms of both accuracy and robustness to label noise. Furthermore, we analyze the transferability of TMLC-Net, showcasing its adaptability to new datasets and noise conditions, and establishing its potential as a broadly applicable solution for robust deep learning in noisy environments.

Paper Structure

This paper contains 34 sections, 12 equations, 5 figures, 6 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Noisy Label Learning
  4. Loss Correction
  5. Sample Selection
  6. Label Correction
  7. Meta-Learning
  8. Transfer Learning
  9. Method
  10. Problem Formulation and Notation
  11. TMLC-Net Architecture
  12. Normalized Noise Perception (NNP)
  13. Time-Series Encoding (TSE)
  14. Subclass Decoding (SD)
  15. Training Algorithm
  16. ...and 19 more sections

Figures (5)

  • Figure 1: An overview of the pipeline of our TMLC-NET.
  • Figure 2: An overview of the pipeline of our TSE.
  • Figure 3: t-SNE visualization of the LSTM hidden states in TMLC-Net. Points are colored according to their true labels (even though the model is trained with noisy labels). (a) Before training TMLC-Net. (b) After training TMLC-Net.
  • Figure 4: Analyzing the Factors Influencing Transferability Performance.
  • Figure 5: Examples of failure cases. (a) A mislabeled image that TMLC-Net fails to correct. (b) Another example.