Combating Label Noise in Deep Learning Using Abstention

Sunil Thulasidasan; Tanmoy Bhattacharya; Jeff Bilmes; Gopinath Chennupati; Jamal Mohd-Yusof

Source-linked AI summary

Combating Label Noise in Deep Learning Using Abstention

Sunil Thulasidasan, Tanmoy Bhattacharya, Jeff Bilmes, Gopinath Chennupati, Jamal Mohd-Yusof

arXiv:1905.10964v2May 27, 2019stat.MLcs.LG

TL;DR

Deep networks need large labeled datasets, but those labels can be noisy, and training-time noise identification has been underexplored. The paper introduces the DAC, which learns with abstention to handle structured and arbitrary label noise, reporting robust learning, feature-based abstention, and data-cleaning improvements across image benchmarks. Its scope excludes adversarially perturbed samples, and abstention does not generally prevent memorization.

Problem
Large-scale training data can contain erroneous labels, while identifying and ignoring such noise during deep-model training has received limited attention.
Method
The DAC adds an abstention-capable loss that lets a DNN abstain during training and inference while continuing to learn on non-abstained samples.
Results
The DAC supports feature-based abstention for structured noise and acts as a data cleaner for arbitrary noise, improving results over existing methods on multiple image benchmarks.
Takeaways & Limitations
The simple loss-function change works with existing DNN architectures and makes the DAC applicable to real-world deep-learning pipelines.
Takeaways & Limitations
The paper does not study adversarially perturbed samples, and abstention appears to counter but does not generally prevent memorization.

Abstract

from arXiv · show

We introduce a novel method to combat label noise when training deep neural networks for classification. We propose a loss function that permits abstention during training thereby allowing the DNN to abstain on confusing samples while continuing to learn and improve classification performance on the non-abstained samples. We show how such a deep abstaining classifier (DAC) can be used for robust learning in the presence of different types of label noise. In the case of structured or systematic label noise -- where noisy training labels or confusing examples are correlated with underlying features of the data-- training with abstention enables representation learning for features that are associated with unreliable labels. In the case of unstructured (arbitrary) label noise, abstention during training enables the DAC to be used as an effective data cleaner by identifying samples that are likely to have label noise. We provide analytical results on the loss function behavior that enable dynamic adaption of abstention rates based on learning progress during training. We demonstrate the utility of the deep abstaining classifier for various image classification tasks under different types of label noise; in the case of arbitrary label noise, we show significant improvements over previously published results on multiple image benchmarks. Source code is available at https://github.com/thulas/dac-label-noise

1 Introduction

The paper introduces training-time abstention as a mechanism for combating label noise in deep neural networks. The DAC learns to abstain on confusing or unreliable samples while continuing to learn on non-abstained examples, supporting structured-noise representation learning and arbitrary-noise data cleaning.

Research gap: Training-time abstention addresses the under-explored problem of identifying and ignoring label noise during deep-model training.Earlier abstention work primarily focused on post-training selective classification during inference.
Core approach: The DAC abstains on confusing training samples while continuing to learn true classes and progressively reducing abstention as classification improves.The model ultimately abstains mainly on the most confusing samples.
Structured noise: For structured noise, the DAC learns features associated with unreliable labels and abstains on those features with high precision.This representation learning supports both eliminating structured noise and interpreting abstention decisions.
Unstructured noise: For arbitrary label noise, the DAC functions as a data cleaner and improves learning on CIFAR-10, CIFAR-100, and Fashion-MNIST over existing methods.The method changes only the loss function and can use existing DNN architectures.
Scope: The paper does not evaluate whether the abstaining classifier reliably abstains on adversarially perturbed samples.Adversarial settings are left for future exploration.

2 Loss Function for the Deep Abstaining Classifier

The DAC modifies cross-entropy with an abstention output and penalty, then adapts the penalty during training. The resulting dynamics encourage abstention on poorly learned samples while preserving learning on their true classes.

Loss formulation: The DAC adds a (k+1)th output p_k+1 representing abstention to a k-class DNN classifier.The other k outputs represent class probabilities for the input.
Loss formulation: The modified loss normalizes cross-entropy over non-abstaining classes and adds an abstention penalty weighted by α ≥ 0.When p_k+1 = 0, the loss recovers standard cross-entropy; extreme α values suppress or permit abstention.
Theoretical result: The loss guarantees that gradient descent continues learning the true class even when that class is not the winning prediction.The proof establishes persistence of true-class learning in the presence of abstention.
Abstention dynamics: When little probability mass lies on the true class, the DAC pushes probability into abstention if α satisfies the derived threshold condition.This behavior links abstention to the model’s current cross-entropy error on the true class.
Auto-tuning α: The analytical threshold enables automatic tuning of α during training using a smoothed moving average updated at mini-batch iterations.The tuning procedure uses the threshold to control abstention behavior as learning progresses.
Auto-tuning α: Training begins with L abstention-free warm-up epochs, then α starts below the threshold and is linearly increased toward α_final.The schedule initially encourages broad abstention and reduces it as learning on true classes progresses.

3 The DAC as a Learner of Structured Noise

The DAC learns features correlated with structured label noise and uses them to abstain on confusing samples. In smudged-image and randomized-monkey experiments, it identifies unreliable examples while preserving learning on non-abstained data.

Motivation: Structured label noise can arise when confusing features or annotator unreliability make some classes more likely to be mislabeled.Such correlations may be present in inputs without being initially obvious, making manual curation costly or impractical.
Experimental setup: The experiment randomizes labels on 10% of training images while adding a distinguishing smudge, then compares DAC, baseline DNN, and post-DAC retraining.The post-DAC model is trained after removing samples abstained on at the DAC’s best-performing epoch.
Smudged images: Over smudged, label-randomized data, baseline accuracy falls from over 82% on clean training data to under 75% at 100% coverage.The DAC eventually abstains on about 10% of training data, corresponding to the smudged images.
Smudged images: At its best validation epoch, the DAC abstains with high precision and recall on precisely the smudged images.The model associates the smudge with unreliable data and assigns those samples to the abstention class.
Smudged images: Removing DAC-abstained samples and retraining produces significantly higher accuracy and consistently better risk-coverage curves than the baseline DNN.The DAC’s own curve closely tracks the baseline because both models learn similarly on non-abstained samples.
Randomized monkey class: With randomized labels for all monkey images, the DAC abstains on most monkey test images while abstaining on far fewer images from other classes.The loss redirects the learned mapping from monkey features toward the abstention class rather than eliminating the underlying representation.
Randomized monkey class: Baseline softmax confidence is unreliable for randomized monkey images: many predictions exceed p >= 0.9 even though most are incorrect.The DAC has a small but consistent risk-versus-coverage advantage over softmax and SGR comparisons.
Visual explanations: Filter visualizations show smudges dominating saliency during abstention, while monkey features are correctly detected before monkey samples are rejected.Removing the smudge makes actual class features more salient and yields a correct prediction.

4 Learning in the Presence of Unstructured Noise: The DAC as a Data Cleaner

For arbitrary label noise, the DAC identifies likely noisy samples, removes them, and enables downstream training on a cleaner set. Across image benchmarks, this data-cleaning procedure usually improves performance over comparison methods while avoiding detailed label-flip modeling.

Problem and assumptions: Unstructured noise consists of labels corrupted arbitrarily on some fraction of data, degrading classification performance.The method assumes a fraction of labels are uniformly corrupted rather than modeling class-to-class flipping probabilities in detail.
Evaluation: Table 1 compares DAC filtering plus downstream DNN training with baseline, related noisy-label methods, and a hypothetical perfect oracle cleaner.Parenthetical DAC values indicate removed-data fraction and remaining noise level.
Data-cleaning procedure: The DAC identifies samples for elimination by monitoring the non-abstaining classifier’s validation performance and selecting the best validation-error point.Lemma 1 supports continued learning on true classes despite abstention.
Evaluation: Experiments use CIFAR-10, CIFAR-100, and Fashion-MNIST with increasing fractions of arbitrarily randomized labels and matched downstream architectures and training settings.The downstream DNN trains on the cleaner set using the same architecture and optimization settings as comparison models.
Results: DAC cleaning yields noticeable and often significant improvements over comparison methods in most cases, and it is consistently better than the noisy baseline.At 80% randomization on CIFAR-100, comparable performance may reflect substantial data removal leaving few samples per class.
Abstention behavior: Fixed abstention penalties can produce all-or-nothing abstention, while zero abstention matches baseline non-abstaining DNN accuracy.The paper therefore examines abstention behavior as training progresses rather than relying only on a fixed penalty.
Practical implications: The approach is presented as simpler than mentor-student and graphical-model methods and requires changing only the loss function for existing DNN architectures.It does not require estimating a label confusion matrix or assuming the noise amount or a trusted clean dataset.

5 Abstention and Memorization

The DAC’s abstention can counter, but does not generally prevent, memorization of random labels. With fixed abstention parameters, training may eventually reduce abstention and harm generalization.

Abstention and memorization: The DAC’s abstention appears to counter the tendency to memorize random labels, but it does not generally prevent memorization.In structured-noise experiments, the DAC often abstained near-perfectly on label-randomized samples by learning their shared features.
Abstention and memorization: As training time increases with fixed α, the abstention rate γ approaches either 0 or 1.The loss-function analysis establishes this asymptotic behavior for a fixed α.
Abstention and memorization: If α is close to 0, the DAC can abstain on all samples, driving loss and gradients to zero and preventing further learning.This is the exceptional case identified in the proof sketch.
Abstention and memorization: With sufficiently small learning rates and long schedules, abstention eventually converges to zero as the DAC enters a memorization phase.Learning-rate decay coincides with steep abstention reductions and subsequent generalization degradation in the random-smudging experiments.

6 Conclusions

The paper presents the DAC as a robust approach to label noise across structured and arbitrary settings. It reports improved downstream classification, simple implementation, and compatibility with existing DNN architectures.

Conclusions: The DAC supports robust learning under multiple label-noise types, serving as a representation learner for structured noise and a data cleaner for arbitrary noise.These are the two principal uses identified in the conclusion.
Conclusions: Data-cleaning with the DAC significantly improves classification performance for downstream training.
Conclusions: The DAC changes only the loss function, making it simple to implement with any existing DNN architecture.
Conclusions: The paper addresses an area with little prior work on abstention in deep learning and even less work targeting label noise.

Appendix A Proof of Lemma 1

The appendix proves that abstention does not stop learning on the true class: under the stated loss, the true-class gradient is nonpositive during gradient descent.

Proof of Lemma 1: Learning on the true class occurs when the gradient with respect to its pre-activation is negative.Here, j denotes the true class and a_j is the pre-activation entering its softmax unit.
Proof of Lemma 1: The proof shows ∂L/∂a_j ≤ 0, so gradient descent continues learning toward the true class even when abstention is present.The argument uses α ≥ 0 and defines q as the probability mass in the remaining real classes.

Appendix B Noisy Labels associated with a data transformation

In a blurred-image experiment, 20% of training and test images receive Gaussian blur, with randomized labels in training. The DAC abstains on these degraded samples while preserving accuracy on the remaining validation images.

Noisy labels associated with a data transformation: 20% of train and test images are Gaussian-blurred across all classes, and labels are randomized on the blurred training set.The transformation creates unreliable labels associated with feature degradation without a conspicuous added feature.
Noisy labels associated with a data transformation: The DAC abstains remarkably well on blurred test images while maintaining approximately 79% accuracy on the remaining validation samples.
Noisy labels associated with a data transformation: The baseline DNN reaches 63% overall accuracy, while its accuracy on smudged images alone is approximately 9.8%, no better than random.
Noisy labels associated with a data transformation: After abstention begins, the DAC initially abstains on many samples, then settles near 20%, matching the proportion of blurred training images.The persistent abstention indicates an association between blurring and unreliable labels.
Noisy labels associated with a data transformation: The training dynamics show that the DAC continues learning while abstaining, then retains abstention on the blurred subset after learning the non-randomized samples.

Appendix C Results on Non-Uniform Label Noise

This section reports CIFAR-10, CIFAR-100, and Fashion-MNIST results for class-dependent label noise, using the setup described in Zhang and Sabuncu (2018). Table 1 compares DAC with related methods, with DAC using abstention-based noise elimination before downstream training.

Table 1 compares DAC with related methods for class-dependent label noise.For DAC, an abstaining classifier identifies and eliminates label noise before an identical DNN performs downstream training.
Experiments evaluate class-dependent label noise on CIFAR-10, CIFAR-100, and Fashion-MNIST.The experimental setup is described as identical to Zhang and Sabuncu (2018).
CIFAR-10 noise flips selected classes, including TRUCK to AUTOMOBILE, BIRD to AIRPLANE, DEER to HORSE, and CAT with DOG.The flips occur with probability η, with CAT and DOG mutually flipping.