Instance-Aware Learning in the Presence of Label Noise: An Adaptive Framework for Image Classification
Date
2025
Authors
Garg, Arpit
Editors
Advisors
Carneiro, Gustavo
Felix, Rafael
Nguyen, Cuong
Felix, Rafael
Nguyen, Cuong
Journal Title
Journal ISSN
Volume Title
Type:
Thesis
Citation
Statement of Responsibility
Conference Name
Abstract
In the field of machine learning, the performance of image classification models depends heavily on the quality of training datasets. Such quality is negatively affected by noisy labels, which can significantly degrade the training process of the models and reduce their accuracy and generalisation capabilities. Traditionally, noisy labels have been viewed as random labelling mistakes, in which the resulting noisy label is assumed to be independent of the underlying data it represents. However, this perspective fails to account for the complexities introduced by real-world instance-dependent and open-set label noise that present additional layers of complexity compared with random mistakes, where the likelihood of a label being noisy is contingent on both the features of the instance and its distribution. These types of label noise arise from subjective interpretations, ambiguous images, out-of-distribution samples, or context-specific misconceptions, making them prevalent issues in training datasets commonly employed in contemporary machine-learning applications. Although noisy label learning has seen notable progress, existing state-of-the-art (SOTA) techniques often fail to account for instance-dependent and open-set noisy labels, leading to suboptimal performance in real-world applications. We argue that this limitation stems from the absence of effective mechanisms capable of addressing both types of noise simultaneously. Such mechanisms rely on accurately estimating noise rates, which is crucial for successful learning in the presence of instance-dependent and open-set noisy labels. This thesis aims to investigate the nuanced dynamics of instance-dependent and openset noisy labels in image classification to answer the following research question: How to develop effective image classification models and techniques to identify, adapt to, and rectify instance-dependent and open-set noisy labels via the accurate estimate of noise rates? This thesis focuses on four key contributions: (1) developing an innovative graphical model framework (InstanceGM) to capture complex dependencies in instance-dependent noisy labels; (2) introducing a peer-based sample selection framework (PASS) to enhance label reliability through mutual agreement between classifiers; (3) establishing robust methodologies for accurate noise rate estimation in instancedependent scenarios; and (4) proposing a unified framework (AEON) to jointly handle both closed- and open-set noise. These contributions introduce innovative methodologies to model, identify, and rectify labelling errors through comprehensive frameworks to estimate label noise rates. In particular, this thesis demonstrates how accurate label noise estimation and subsequent label correction can significantly improve the model performance. The methods proposed in this research contribute to the noisy label learning literature by offering improvements in robustness and accuracy compared to several current benchmark models under instance-dependent noise conditions. The findings and methodologies are thoroughly documented and made accessible through a GitHub repository (https://github.com/arpit2412), facilitating further research and application in this critical area of machine learning.
School/Discipline
School of Computer and Mathematical Sciences
Dissertation Note
Thesis (Ph.D.) -- University of Adelaide, School of Computer and Mathematical Sciences, 2025
Provenance
This electronic version is made publicly available by the University of Adelaide in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material you wish to be removed from this electronic version, please complete the take down form located at: http://www.adelaide.edu.au/legals