Source-linked AI summary
Deep Self-Learning From Noisy Labels
Jiangfan Han, Ping Luo, Xiaogang Wang
TL;DR
Noisy labels substantially hinder ConvNet training, while clean supervision is costly and prior methods rely on restrictive assumptions or auxiliary resources. The paper proposes SMP, an iterative self-learning framework that corrects labels with multiple prototypes while jointly training one network. SMP achieves state-of-the-art performance across the evaluated real noisy datasets without extra supervision.
Problem
Noisy labels degrade ConvNet performance, while clean annotations and extra clean supervision are costly for real-world datasets.
Method
SMP iteratively trains one ConvNet with original and corrected labels, selecting multiple class prototypes for label correction without noise-distribution assumptions or auxiliary networks.
Results
SMP achieves state-of-the-art performance on the evaluated real noisy datasets.
Takeaways & Limitations
Iterative multi-prototype self-learning provides an effective end-to-end framework for training on real noisy datasets without extra supervision.
Takeaways & Limitations
Approaches using additional supervision remain expensive because they require extra clean samples for large-scale real-world applications.
Abstract
from arXiv · showhide
ConvNets achieve good results when training from clean data, but learning from noisy labels significantly degrades performances and remains challenging. Unlike previous works constrained by many conditions, making them infeasible to real noisy cases, this work presents a novel deep self-learning framework to train a robust network on the real noisy datasets without extra supervision. The proposed approach has several appealing benefits. (1) Different from most existing work, it does not rely on any assumption on the distribution of the noisy labels, making it robust to real noises. (2) It does not need extra clean supervision or accessorial network to help training. (3) A self-learning framework is proposed to train the network in an iterative end-to-end manner, which is effective and efficient. Extensive experiments in challenging benchmarks such as Clothing1M and Food101-N show that our approach outperforms its counterparts in all empirical settings.
1. Introduction
Noisy labels make real-world ConvNet training difficult because clean annotation is costly and existing correction methods rely on restrictive assumptions or extra supervision. SMP addresses these constraints with iterative self-learning and multi-prototype label correction, achieving state-of-the-art performance on real noisy datasets.
- Large-scale clean annotations are expensive and time-consuming, while human labeling can introduce mistakes and data bias.
- Internet-collected image tags reduce annotation cost but contain noise, which degrades deep-network performance and motivates robust training methods.
- Transition-matrix and noise-tolerant methods assume sample-independent noise probabilities, limiting performance on real noisy datasets where sample appearance matters.
- Methods using additional supervision can improve robustness but require expensive clean samples, reducing their practicality for large-scale real-world applications.
- SMP iteratively trains one ConvNet with original and corrected labels, selecting multiple class prototypes without extra supervision or an auxiliary network.
- SMP achieves state-of-the-art performance across the evaluated real noisy datasets.
2. Related Work
Prior noisy-label methods use transition models, robust losses, auxiliary networks, or extra supervision, but their assumptions and costs limit real-world applicability. SMP instead combines label correction with efficient end-to-end training and achieves state-of-the-art results on Clothing1M and Food101-N.
- ConvNet performance degrades with noisy labels, while producing large-scale clean and unbiased annotations is expensive and time-consuming.
- Transition-matrix methods model probabilities from true to corrupted labels, often requiring assumptions or cleanly labeled subsets for estimation.
- Robust loss functions improve tolerance to selected noise patterns but remain constrained and perform poorly on real-world noisy data.
- Auxiliary-network, meta-learning, and curriculum-learning approaches address noisy labels, but some require additional clean supervision or other extra mechanisms.
- SMP corrects noisy labels without extra clean supervision and trains end-to-end efficiently, achieving state-of-the-art performance on Clothing1M and Food101-N.
- Unlike semi-supervised pseudo-labeling, noisy-label learning assigns labels to all samples, although some assigned labels may be incorrect.
3. Our Approach
SMP iteratively trains a ConvNet and corrects noisy labels using multiple class prototypes, combining original and corrected supervision without extra information. It selects prototypes from deep features using cosine similarity and density, then relabels samples based on prototype similarity.
- Iterative Framework: SMP alternates network training with label correction, using corrected labels from the second phase as supervision in the first.The training phase uses both original noisy labels and corrected labels; the two phases continue iteratively until convergence.
- Training Phase: The training objective combines cross-entropy losses for the original noisy label and corrected label, weighted by α.The original label remains part of supervision because corrected labels may still misclassify hard samples; α controls the two terms, initially set to 0 before becoming positive.
- Prototype Selection: SMP selects multiple prototypes per class from deep features extracted by the preliminary network, rather than relying on a single prototype.The method uses the ResNet representation before the fully connected layer and selects prototypes from images sharing the same noisy label.
- Prototype Selection: Cosine similarity between deep features forms the class similarity matrix used during prototype selection.Because a class may contain up to n =70k images in Clothing1M, the method randomly samples m images with m < n to reduce computation.
- Prototype Selection: Density ρ favors prototypes surrounded by similar images, since correctly labeled samples are more likely to have high density while noisy samples are often isolated.The threshold Sc determines relative density, and its concrete value is reported not to influence the final result.
- Prototype Selection: Similarity η prevents selected prototypes from clustering too closely by measuring each candidate’s relation to higher-density samples.For the highest-density image, η is the smallest; otherwise, η is the maximum cosine similarity to an image with higher density.
4. Experiments
Experiments on Clothing1M and Food101-N evaluate SMP under noisy-only, verification-assisted, and clean-supervision settings, alongside analyses of label correction, prototype selection, and training weights. SMP improves label correction and classification, with multiple prototypes and joint use of original and corrected labels performing favorably.
- Clothing1M comparison: 74.45% accuracy improves on 69.54% in the noisy-only Clothing1M setting, exceeding Joint Optimization by 2.22% and MLNT-Teacher by 0.98%.SMP also achieves the best result when verification labels are available indirectly and remains ahead of other methods when clean labels are used, although CurriculumNet reports 81.5% with a different backbone.
- Label correction: 74.38% corrected-label accuracy is reached after the first iterative cycle, improving the original 61.74% by 12.64%; final correction accuracy reaches 77.36%.The correction accuracy is measured on the original noisy set and improves further by the end of training.
- Class-level correction: For most classes with original accuracy below 50%, SMP raises accuracy above 60%; the Sweater class improves by 10% from about 30%.The paper also shows example Clothing1M and Food101-N samples whose labels were corrected by the method.
- Prototype analysis: Using two prototypes outperforms one by 2.04%, while p = 1 gives sub-optimal accuracy; multiple prototypes provide more comprehensive class representation.The method selects images as prototypes using density and similarity rather than an additional network.
- Weight-factor analysis: Joint training with original and corrected labels performs best at α = 0.5; using only noisy or only corrected labels is sub-optimal.The result supports retaining original labels because corrected labels can misrecognize hard samples as noise.
- Ablation study: Performance is not sensitive to the number of sampled images m, and sampling 2% of Clothing1M's 70k images per class represents the class distribution well.Changing the prototype-selection clustering method has little impact, although the proposed selection method performs best.
- Food101-N: On Food101-N, SMP achieves state-of-the-art performance and outperforms CleanNet by 1.16%.This extends the reported empirical advantage beyond Clothing1M.
5. Conclusion
The paper presents SMP, an iterative self-learning framework that jointly corrects noisy labels and trains a network on real noisy datasets without extra supervision. Experiments show effective performance, including state-of-the-art results across real noisy datasets.
- SMP iteratively corrects labels using several class prototypes while jointly training with corrected and original noisy labels.The framework combines relabeling and network training in an end-to-end procedure.
- A single prototype is insufficient to represent a class distribution, making multi-prototypes necessary.The conclusion identifies multi-prototypes as necessary for representing the distribution of a noisy class.
- SMP trains on real noisy datasets without an accessorial network or extra supervision.The framework jointly uses corrected and original noisy labels without adding another network or external supervision.
- The methods achieve state-of-the-art performance on different real noisy datasets.