University of Twente Student Theses


Fourier Insights in Machine Learning : Bridging the Augmentation Gap through Frequency-basis Functions

Vaish, Puru (2024) Fourier Insights in Machine Learning : Bridging the Augmentation Gap through Frequency-basis Functions.

[img] PDF
Abstract:For neural networks, challenges arise when deploying models in real-world scenarios, as unforeseen changes in inputs can lead to diminished performance. While data augmentation is a common remedy to bridge the gap between training and test data, its efficacy in enhancing the robustness of computer vision models is not guaranteed. This paper introduces Auxiliary Fourier-basis Augmentation (AFA), a novel approach that extends beyond visual augmentations to address this limitation by focusing on neural networks. AFA leverages Fourier-basis additive noise as a complementary technique in the frequency domain, filling the robustness gap left by conventional visual augmentations. Our method demonstrates its effectiveness in an adversarial setting, showcasing its utility in enhancing model robustness. Notably, AFA contributes to reducing the impact of common corruptions, facilitates out-of-distribution (OOD) generalisation, and ensures consistent model performance against increasing perturbations. Importantly, it introduces a unique capability to minimise frequency shortcuts, further fortifying the overall resilience of neural network models. The results affirm that AFA seamlessly integrates with existing augmentation techniques, providing a comprehensive enhancement to model performance. This work presents a valuable contribution to the broader pursuit of robust neural networks, extending beyond the conventional focus on computer vision models.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:31 mathematics, 54 computer science
Programme:Applied Mathematics MSc (60348)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page