[Paper Review] Deep Self-Learning From Noisy Labels

Jiangfan Han, Ping Luo+1 more|arXiv (Cornell University)|Aug 6, 2019

Machine Learning and Data Classification37 references15 citations

TL;DR

This paper proposes Self-Learning with Multi-Prototypes (SMP), a novel iterative deep self-learning framework that trains convolutional networks on real noisy datasets without extra supervision or assumptions about label noise distribution. By using multiple class prototypes to correct noisy labels and jointly training on both original and corrected labels, SMP achieves state-of-the-art performance on benchmarks like Clothing1M and Food101-N, outperforming prior methods by up to 1.16% accuracy.

ABSTRACT

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.

Motivation & Objective

• To address the challenge of training deep networks on real-world noisy datasets where labels are unreliable due to human error or web-crawled sources.
• To eliminate reliance on extra clean supervision or auxiliary networks, which are costly and impractical for large-scale applications.
• To develop a self-learning framework that corrects noisy labels iteratively without assuming a specific noise distribution.
• To investigate whether multiple prototypes per class improve representation and label correction over single-prototype methods.
• To achieve state-of-the-art performance on real noisy benchmarks like Clothing1M and Food101-N without additional supervision.

Proposed method

• The framework uses a single shared ConvNet for both feature extraction and label correction in an iterative end-to-end manner.
• In each iteration, the model generates corrected labels by comparing sample features to multiple class prototypes selected via clustering.
• Class prototypes are selected using a novel clustering method that identifies representative samples from each noisy class.
• The model is jointly trained using both the original noisy labels and the corrected labels, with a weighted loss where α=0.5 optimizes performance.
• The label correction process avoids overfitting by not replacing all noisy labels, preserving useful information from original labels.
• The method is robust to clustering method choice, as shown by consistent performance across K-means++, density peak, and the proposed method.

Experimental results

Research questions

• RQ1 Can a self-learning framework effectively correct noisy labels in real-world datasets without extra supervision?
• RQ2 Does using multiple prototypes per class improve label correction and model generalization compared to a single prototype?
• RQ3 How sensitive is the framework to the number of samples used to compute class prototypes?
• RQ4 Does the joint training of original and corrected labels outperform replacing noisy labels entirely?
• RQ5 Can the framework achieve state-of-the-art performance on real noisy benchmarks like Clothing1M and Food101-N?

Key findings

• On Clothing1M, SMP achieves 74.45% test accuracy using only 2% of training samples (2560 out of 128,000 per class) to compute prototypes, outperforming prior methods.
• With 25,000 additional verified labels, SMP reaches 76.44% accuracy, demonstrating strong generalization and scalability.
• The method is robust to the choice of clustering algorithm, with performance within 0.3% across K-means++, density peak, and the proposed method.
• Using multiple prototypes per class significantly improves label correction and model accuracy compared to single-prototype methods like CleanNet.
• On Food101-N, SMP achieves 85.11% accuracy, surpassing CleanNet (83.95% with soft weighting) by 1.16%.
• The joint training with both original and corrected labels (α=0.5) yields the best performance, confirming that original labels retain useful inductive bias.