Table of Links

• Abstract and 1 Introduction
• 2 The Effect Of Data Augmentation-Induced Class-Specific Bias Is Influenced By Data, Regularization and Architecture
• 2.1 Data Augmentation Robustness Scouting
• 2.2 The Specifics Of Data Affect Augmentation-Induced Bias
• 2.3 Adding Random Horizontal Flipping Contributes To Augmentation-Induced Bias
• 2.4 Alternative Architectures Have Variable Effect On Augmentation-Induced Bias
• 3 Conclusion and Limitations, and References
• Appendices A-L

2 The Effect Of Data Augmentation-Induced Class-Specific Bias Is Influenced By Data, Regularization and Architecture

This section details our study’s data-centric and model-centric analysis of the phenomena originally observed in (Balestriero, Bottou, and LeCun 2022). Firstly, we establish a practical framework for replicating such experiments in Section 2.1. Following this, we use a ResNet50 model trained from scratch with the Random Cropping and Random Horizontal Flip DA to provide the data-centric analysis of DA-induced class-specific bias on three datasets (Fashion-MNIST, CIFAR-10 and CIFAR-100) in Section 2.2. We then take a step back in Section 2.3 to evaluate the potential side effects of including the Random Horizontal Flip augmentation, as done in the original study. Finally, we conclude by demonstrating how alternate computer vision architectures interact with the phenomenon illustrated in previous sections. These findings are key as they serve to deepen our understanding of the potential pitfalls of introducing DA to computer vision tasks in order to improve overall model performance, while showing how the problem of class-specific bias can be alleviated or forestalled.