Pre-train to Gain: Robust Learning Without Clean Labels

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Szczecina, David, Pellegrino, Nicholas, Fieguth, Paul
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914171280424960
author Szczecina, David
Pellegrino, Nicholas
Fieguth, Paul
author_facet Szczecina, David
Pellegrino, Nicholas
Fieguth, Paul
contents Training deep networks with noisy labels leads to poor generalization and degraded accuracy due to overfitting to label noise. Existing approaches for learning with noisy labels often rely on the availability of a clean subset of data. By pre-training a feature extractor backbone without labels using self-supervised learning (SSL), followed by standard supervised training on the noisy dataset, we can train a more noise robust model without requiring a subset with clean labels. We evaluate the use of SimCLR and Barlow~Twins as SSL methods on CIFAR-10 and CIFAR-100 under synthetic and real world noise. Across all noise rates, self-supervised pre-training consistently improves classification accuracy and enhances downstream label-error detection (F1 and Balanced Accuracy). The performance gap widens as the noise rate increases, demonstrating improved robustness. Notably, our approach achieves comparable results to ImageNet pre-trained models at low noise levels, while substantially outperforming them under high noise conditions.
format Preprint
id arxiv_https___arxiv_org_abs_2511_20844
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Pre-train to Gain: Robust Learning Without Clean Labels
Szczecina, David
Pellegrino, Nicholas
Fieguth, Paul
Machine Learning
Artificial Intelligence
Neural and Evolutionary Computing
68T05
I.2.6
Training deep networks with noisy labels leads to poor generalization and degraded accuracy due to overfitting to label noise. Existing approaches for learning with noisy labels often rely on the availability of a clean subset of data. By pre-training a feature extractor backbone without labels using self-supervised learning (SSL), followed by standard supervised training on the noisy dataset, we can train a more noise robust model without requiring a subset with clean labels. We evaluate the use of SimCLR and Barlow~Twins as SSL methods on CIFAR-10 and CIFAR-100 under synthetic and real world noise. Across all noise rates, self-supervised pre-training consistently improves classification accuracy and enhances downstream label-error detection (F1 and Balanced Accuracy). The performance gap widens as the noise rate increases, demonstrating improved robustness. Notably, our approach achieves comparable results to ImageNet pre-trained models at low noise levels, while substantially outperforming them under high noise conditions.
title Pre-train to Gain: Robust Learning Without Clean Labels
topic Machine Learning
Artificial Intelligence
Neural and Evolutionary Computing
68T05
I.2.6
url https://arxiv.org/abs/2511.20844