Source-linked AI summary

How does Disagreement Help Generalization against Label Corruption?

Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor W. Tsang, Masashi Sugiyama

arXiv:1901.04215v3cs.LGstat.ML

TL;DR

Noisy labels challenge learning because deep networks can memorize corrupted supervision, while Co-teaching’s two networks can eventually reach consensus. The paper proposes Co-teaching+, which filters disagreement data, selects small-loss examples, and cross-updates peer networks; experiments report superior robustness across noisy-label benchmarks. Its scope includes the stated assumption that the noise rate is known, and the paper notes that disagreement updating alone cannot handle noisy labels.

  • Problem

    Noisy labels degrade classifier accuracy, and Co-teaching’s networks can converge over training, creating a need to preserve useful divergence.

  • Method

    Co-teaching+ keeps prediction-disagreement data, selects small-loss examples among them, and updates each network using its peer’s selected data.

  • Results

    Co-teaching+ was more robust than many state-of-the-art methods, including Co-teaching, MentorNet, and F-correction, across simulated and real-world noisy datasets.

  • Takeaways & Limitations

    Robust noisy-label training combines small-loss selection, cross-updating between networks, and maintaining divergence throughout training.

  • Takeaways & Limitations

    The paper states that Update by Disagreement alone cannot handle noisy labels, and its experimental initialization assumes the noise rate τ is known.

Abstract

from arXiv · show

Learning with noisy labels is one of the hottest problems in weakly-supervised learning. Based on memorization effects of deep neural networks, training on small-loss instances becomes very promising for handling noisy labels. This fosters the state-of-the-art approach "Co-teaching" that cross-trains two deep neural networks using the small-loss trick. However, with the increase of epochs, two networks converge to a consensus and Co-teaching reduces to the self-training MentorNet. To tackle this issue, we propose a robust learning paradigm called Co-teaching+, which bridges the "Update by Disagreement" strategy with the original Co-teaching. First, two networks feed forward and predict all data, but keep prediction disagreement data only. Then, among such disagreement data, each network selects its small-loss data, but back propagates the small-loss data from its peer network and updates its own parameters. Empirical results on benchmark datasets demonstrate that Co-teaching+ is much superior to many state-of-the-art methods in the robustness of trained models.

1. Introduction

Noisy labels can degrade classifier accuracy, especially because deep networks may memorize corrupted labels. Co-teaching uses peer-selected small-loss examples, but its networks can converge; Co-teaching+ combines disagreement-based filtering with cross-updating to preserve divergence.

  • Deep networks can fully memorize noisy labels, worsening the accuracy degradation caused by corrupted supervision.
  • Small-loss training became a promising approach, including MentorNet and Co-teaching, for selecting examples believed to be cleaner.MentorNet uses an extra network for selection, whereas Co-teaching trains two networks symmetrically and exchanges selected examples.
  • Co-teaching’s initially diverged networks gradually approach consensus, reducing its distinction from self-paced MentorNet.The consensus issue motivates maintaining or slowing the loss of divergence during training.
  • Co-teaching+ first retains examples on which two networks disagree, then selects small-loss examples and cross-updates each network with its peer’s selections.This combines a disagreement-update step for data with a cross-update step for parameters.
  • Experiments on simulated and real-world noisy datasets found Co-teaching+ more robust than several state-of-the-art methods, including Co-teaching and MentorNet.The evaluated datasets include noisy MNIST, CIFAR-10, CIFAR-100, NEWS, T-ImageNet, and three Open-sets.
  • The paper identifies small-loss selection, cross-updating, and keeping the two networks diverged as three factors in robust noisy-label training.

2. Co-teaching+: Towards Training of Robust Deep Networks with Noisy Labels

Co-teaching+ combines disagreement-based filtering with Co-teaching’s small-loss and peer-update mechanisms to keep two networks divergent during training. Its schedule gradually reduces the retained small-loss data as networks begin fitting noisy labels.

  • Co-teaching+ algorithm: Co-teaching+ first retains prediction-disagreement data, then each network selects small-loss instances and updates itself using its peer’s selected data.The algorithm combines a disagreement-update step with a cross-update step.
  • Motivation: Disagreement-based updates maintain divergence by restricting further training to data on which the two classifiers disagree.This is intended to prevent Co-teaching from reducing to self-training MentorNet.
  • Motivation: Co-teaching+ is motivated by complementary peer learning, in which diverged peers can provide more useful signals than identical peers.The paper illustrates this with students who have complementary strengths in different subjects.
  • Training schedule: The small-loss retention rate decreases over epochs because deep networks fit clean data before gradually over-fitting noisy data.Co-teaching+ decreases λ(e) quickly early and then either holds it constant or decreases it more slowly after Ek epochs.
  • Motivation: Co-teaching can converge to consensus and functionally reduce to self-training MentorNet as training epochs increase.The initial divergence comes from different parameter initializations, but it diminishes over training.
  • Key factors: The paper identifies three factors for handling noisy labels: exploiting memorization, cross-updating two networks, and keeping them divergent.These factors correspond to the small-loss trick, simultaneous peer learning, and disagreement-based training.

3. Experiments on Simulated Noisy Datasets

Experiments evaluate Co-teaching+ on simulated noisy datasets using standard baselines, multiple architectures, and test accuracy over training. Across MNIST, CIFAR-10, CIFAR-100, NEWS, and T-ImageNet, Co-teaching+ generally remains more robust, especially as noise becomes harder.

  • Experimental setup: Experiments cover MNIST, CIFAR-10, CIFAR-100, NEWS, and T-ImageNet, using simulated symmetry-flipping and pair-flipping label noise.The evaluation includes vision and text datasets, with benchmark methods implemented using default parameters.
  • Experimental setup: The comparisons include MentorNet, Co-teaching, Decoupling, F-correction, and a standard network trained directly on noisy data.These baselines represent small-loss selection, peer updating, disagreement-based updating, transition-matrix correction, and direct noisy-label training.
  • Results: Standard training first achieves high test accuracy and then declines as the network overfits noisy labels, revealing the memorization effect.This trend appears in the MNIST, CIFAR-10, CIFAR-100, and NEWS experiments.
  • Results: Co-teaching+ outperforms the competing methods on the easiest CIFAR-10 and CIFAR-100 symmetry-noise cases.On CIFAR-10, it performs much better than all baselines under Symmetry-20%; on CIFAR-100, it and F-correction outperform Co-teaching, MentorNet, and Decoupling.
  • Results: On T-ImageNet, Co-teaching+ is best in both symmetry cases, while Co-teaching and Co-teaching+ outperform other methods in the pair case.The comparison uses averaged and maximal test accuracy over the last 10 epochs.

4. Experiments on Real-world Noisy Datasets

The experiments evaluate Co-teaching+ on open-set noisy datasets and report stronger robustness than competing methods. They also identify maintaining disagreement between networks as a key ingredient of robust training.

  • Dataset: Open-sets replace some CIFAR-10 training images with outside images while preserving labels and per-class image counts.The outside images come from CIFAR-100, ImageNet-32, and SVHN.
  • Network & Optimizer & Initialization: The experiments use a six-convolutional-layer network with one fully connected layer, batch normalization, ReLU activations, max pooling, and SGD.
  • Results on three Open-sets: Co-teaching+ outperforms other baselines on all three open-set noisy datasets with 40% open-set noise.The reported comparisons include Standard, Iterative, MentorNet, and Co-teaching.
  • Results on three Open-sets: Co-teaching and Co-teaching+ outperform Iterative on CIFAR-100 and ImageNet-32 noise, while Co-teaching+ is significantly better than Iterative on SVHN noise.
  • Reflection of results: Co-teaching+ is framed as cross-training two diverged deep networks with the small-loss trick, and network divergence is identified as a key ingredient of robustness.The paper links this conclusion to the result difference between Co-teaching and Co-teaching+.

5. Conclusion

The paper presents Co-teaching+ as a robust learning paradigm for training deep networks under noisy supervision. Its central procedure combines prediction-disagreement selection with peer cross-training on small-loss data.

  • Co-teaching+ maintains two networks that identify prediction-disagreement data and cross-trains on samples screened by the small-loss criterion.
  • The paper reports that Co-teaching+ trains deep models robustly with extremely noisy supervision beyond Co-teaching and MentorNet.
  • The method is evaluated empirically to demonstrate robust training under noisy supervision.

A. Related literature

The related literature spans statistical learning and deep learning approaches to noisy-label learning. Statistical methods are organized into surrogate loss, noise-rate estimation, and probabilistic modeling.

  • Statistical learning methods: Statistical learning approaches to noisy labels are categorized into surrogate loss, noise rate estimation, and probabilistic modeling.
  • Statistical learning methods: Surrogate-loss methods include unbiased noise-corrected losses and robust non-convex losses.
  • Statistical learning methods: Noise-rate estimation methods use class-probability estimators based on score order statistics or ROC-curve slopes.
  • Statistical learning methods: Probabilistic modeling methods include two-coin models for noisy labels from multiple annotators and extensions with instance-dependent flipping probabilities.

B. Training details

Training schedules vary by dataset: Co-teaching+ starts immediately on MNIST and NEWS, while other datasets use Co-teaching warm-up periods before disagreement updates.

  • Training schedule: Co-teaching+ begins training immediately on MNIST and NEWS but uses warm-up strategies on the other datasets.
  • Training schedule: CIFAR-10 uses 20 epochs of Co-teaching warm-up, and CIFAR-100 uses 5 epochs before Co-teaching+ training.
  • Training schedule: For T-ImageNet, disagreement updates begin in the middle of training.
Loading 1901.04215v3…