[Paper Review] Attention-Aware Noisy Label Learning for Image Classification

Zhenzhen Wang, Chunyan Xu+2 more|arXiv (Cornell University)|Sep 30, 2020

Machine Learning and Data Classification24 references4 citations

TL;DR

This paper proposes A²NL, an attention-aware noisy label learning framework for image classification that uses a Noise-Attention model with multiple noise-specific units to capture diverse label noise distributions across classes and images. It further employs a recursive learning strategy to distill knowledge from previous iterations, significantly improving robustness on both synthetic and real-world noisy datasets, achieving up to 4% error reduction over strong baselines without requiring clean data.

ABSTRACT

Deep convolutional neural networks (CNNs) learned on large-scale labeled samples have achieved remarkable progress in computer vision, such as image/video classification. The cheapest way to obtain a large body of labeled visual data is to crawl from websites with user-supplied labels, such as Flickr. However, these samples often tend to contain incorrect labels (i.e. noisy labels), which will significantly degrade the network performance. In this paper, the attention-aware noisy label learning approach ($A^2NL$) is proposed to improve the discriminative capability of the network trained on datasets with potential label noise. Specifically, a Noise-Attention model, which contains multiple noise-specific units, is designed to better capture noisy information. Each unit is expected to learn a specific noisy distribution for a subset of images so that different disturbances are more precisely modeled. Furthermore, a recursive learning process is introduced to strengthen the learning ability of the attention network by taking advantage of the learned high-level knowledge. To fully evaluate the proposed method, we conduct experiments from two aspects: manually flipped label noise on large-scale image classification datasets, including CIFAR-10, SVHN; and real-world label noise on an online crawled clothing dataset with multiple attributes. The superior results over state-of-the-art methods validate the effectiveness of our proposed approach.

Motivation & Objective

• To address the degradation in deep learning performance caused by noisy labels in web-crawled image datasets.
• To model heterogeneous label noise distributions across different classes and individual images, rather than assuming uniform noise per class.
• To enhance the learning capacity of noisy-label-robust networks without increasing model parameters.
• To develop a recursive training strategy that leverages soft predictions from prior iterations to improve generalization.
• To validate the method on both synthesized label noise and real-world noisy data from online sources.

Proposed method

• The Noise-Attention (NA) model consists of multiple noise-specific units, each learning a distinct noisy label distribution for a subset of images.
• Each noise-specific unit is trained to model specific types of label confusion, such as misclassifying dogs as cats or horses.
• A recursive learning strategy combines ground-truth labels with soft predictions from a previously trained attention network to supervise the current iteration.
• The recursive process iteratively refines the network by leveraging high-level knowledge from prior iterations, improving feature representation without adding parameters.
• The framework is plug-and-play, compatible with any standard CNN architecture, and can be applied to both synthetic and real-world noisy datasets.
• The method is trained end-to-end with cross-entropy loss, using the combined supervision of true labels and soft outputs from the previous iteration.

Experimental results

Research questions

• RQ1 Can modeling multiple, distinct noisy label distributions per class improve robustness to label noise in image classification?
• RQ2 How does the recursive learning strategy that distills soft predictions from prior iterations affect model generalization on noisy data?
• RQ3 Does the proposed attention-aware framework outperform existing methods on both synthetic and real-world noisy label benchmarks?
• RQ4 To what extent can the Noise-Attention model capture intra-class variation in label noise, such as different misclassification patterns for similar classes?
• RQ5 Can the method achieve strong performance without requiring additional clean data for pretraining or fine-tuning?

Key findings

• The proposed A²NL framework achieved a 4% error reduction on the real-world clothing dataset with multiple attributes compared to the base model, with M=6 noise-specific units.
• The recursive learning strategy improved test error by over 1% after just one iteration, with further gains up to 5 iterations, after which improvements became marginal.
• The method outperformed the Bottom-up model [20] by 2% in error rate before recursive fine-tuning, and by 4% after recursive learning, despite using only noisy training data.
• Even without access to clean data, the A²NL model with recursive learning matched or surpassed the performance of methods like Distillation [15], which require 2,000 manually purified test images for pretraining.
• The ablation study confirmed that increasing the number of noise-specific units improved performance up to a saturation point, with M=6 being optimal for the clothing dataset.
• The method generalized well across both synthetic label noise (CIFAR-10, SVHN) and real-world noisy data, demonstrating robustness to diverse noise patterns.