Fair Compression of Machine Learning Vision Systems - Thesis Abstract

Aug 3, 2026

Neural network pruning compresses models by removing least-influential parameters, often improving speed and size with minimal performance change. Research shows, however, that pruning can worsen fairness: underrepresented or complex subgroups may be disproportionately affected, producing significant biases in real-world image classification systems. This thesis analyzes fairness effects across multiple vision models and introduces a method to improve fairness in existing pruning approaches, using fairness behavior under varied datasets and pruning conditions. The study finds dataset-dependent fairness outcomes and proposes a performance-weighted cross-entropy tweak to enhance post-pruning fairness.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages63

Reading Time2 h 39 min

AccessFree to download

1 Introduction

1.1 Problem Definition

1.2 Contributions and Outline

2 Background

2.1 Convolutional Neural Networks

2.3 Convolutional Neural Network Compression

2.4 Fair CNN Compression

3 An Empirical Analysis of Fair Pruning

3.1 Analysis Protocol

3.2 Pruning Biased CNNs

3.3 Adjusting the Attribute and Class Balance

Fair Compression of Machine Learning Vision Systems by Robbie Meyer A thesis presented to the University of Waterloo in fulfillment of the thesis requirement for the degree of Master of Applied Science in Systems Design Engineering Waterloo, Ontario, Canada, 2023 © Robbie Meyer 2023 Author’s Declaration This thesis consists of material all of which I authored or co-authored: see Statement of Contributions included in the thesis. This is a true copy of the thesis, including any required final revisions, as accepted by my examiners. I understand that my thesis may be made electronically available to the public. Statement of Contributions The following two papers were used in this thesis. For all papers, I was a co-author with major contributions on the design, development, evaluation and writing of the papers’material. R. Meyer and A. Wong,“From Intention to Action: The Fair AI Toolbox,” Computational Vision and Imaging Systems, 2022 . This paper is incorporated in Chapter 2. R. Meyer and A. Wong,“A Fair Loss Function for Network Pruning,” Trustworthy and Socially Responsible Machine Learning, NeurIPS, 2022 . This paper is incorporated in Chapters 2 , 3 and 4. Abstract Model pruning is a simple and effective method for compressing neural networks. By identifying and removing the least influential parameters of a model, pruning is able to transform networks into smaller, faster networks with minimal impact to overall performance. However, recent research has shown that while overall performance may not be significantly changed, model pruning can exacerbate existing fairness issues. Subgroups that are underrepresented or complex may experience a greater than average impact from pruning. Machine learning systems that use compressed neural networks may consequently exhibit significant biases that could limit their effectiveness in many real world situations. To address this issue, we analyze the effect on fairness of pruning a variety of image classification models and propose a novel method for improving the fairness of existing pruning methods. By analyzing the fairness impact of pruning in a variety of situations, we further our understanding of the theoretical fairness impact of pruning could manifest in real-world conditions. By developing a method for improving the fairness of pruning methods, we demonstrate that the fairness impact of pruning can be influenced and enable machine learning practitioners to improve the post-pruning fairness of their models. Our analysis revealed that the fairness impact of pruning can be observed in many, but not all, image classification systems that utilize deep learning and pruning. The dataset used to train each model appears to influence how pruning affects the fairness of each model. Models trained and pruned using the CelebA dataset did see a negative impact on fairness while models trained and pruned using the Fitzpatrick17k dataset did not. Manipulating the CelebA and CIFAR-10 datasets to remove or introduce potential sources of bias does affect the fairness impact of pruning. The effect does not appear to be limited to a single pruning method, but different pruning methods do not experience the effect equally. The fairness impact of data-driven pruning can be improved through a simple tweak to the cross-entropy loss. The performance weighted loss function assigns weights to samples based on the performance of the unpruned model and uses the corrected output of the unpruned model as classification targets. These small changes improve the fairness of existing pruning methods with some models. The performance weighted loss function does not appear to be universally beneficial, but it is a useful tool for machine learning practitioners who seek to compress models in fairness sensitive contexts. Acknowledgements I would first like to thank my supervisor Prof. Alexander Wong for his support and guidance throughout my degree. Your passion is ins

FAQs

Why can model pruning harm fairness even when overall accuracy stays similar?

Pruning removes least-influential parameters, which can exacerbate fairness issues by disproportionately impacting underrepresented or complex subgroups, leading to significant bias in the resulting systems.

What factors influence how pruning affects fairness in image classification models?

The dataset used to train and prune a model strongly affects the observed fairness impact. The thesis also finds that different pruning methods do not experience the effect equally and that manipulating bias sources in datasets can change fairness outcomes.

How does the thesis improve fairness of existing pruning methods?

It proposes improving fairness through a tweak to the cross-entropy loss using a performance weighted loss function that assigns sample weights based on the unpruned model’s performance and uses corrected unpruned outputs as targets.

Aug 3, 2026

Upload by Lucas Martin