Mathematical Connections between Convolutional Neural Networks and the Scattering Transformation

Convolutional neural networks (CNNs) are used in visual recognition tasks because they learn parameters, or weights, that are composed of hidden layers to form one classification in many settings. Standard CNN architectures are surprisingly stable, even though the stability is not ensured a priori. The goal of this paper is to consider a mathematically rigorous architecture that uses steps analogous to those in the neural network architecture. Stéphane Mallat, in his paper “Group Invariant Scattering,” provides one of the earliest attempts by creating the scattering transformation and its finite approximation, the windowed scattering. Importantly, it is invariant to diffeomorphic translation. We review and rederive many of his proofs and constructs, adding context and details where pertinent. We also briefly review a construction of wavelets and basic principles of measure theory, which are used in both the scattering and the windowed scattering. Next, we review extensions of this theory: representing the scattering transformation on a stochastic process, ensuring the scattering works over translations/rotations in a Lie Group, and creating the duality argument, which attempts to avoid the restrictions imposed by the scattering. Finally, we implement the scattering transformation on both the CIFAR-10 dataset and a publicly available dataset from TensorFlow (datasets.bee_dataset). Results include similar (though at times decreased) accuracy but improved stability compared to that of two Keras Conv2D and MaxPooling2D layers.

Grantor dc:publisher

University of Kansas

Year dc:date.issued

2025

Author dc:creator

• Hastings, Adam

Advisor dc:contributor.advisor

• Talata, Zsolt

• Applied mathematics
• Mathematics
• Computer science
• convolutional neural network
• Fourier transform
• harmonic analysis
• scattering
• signal processing
• wavelets

Statement dc:rights

• This item is protected by copyright and unless otherwise specified the copyright of this thesis/dissertation is held by the author.

Language dc:language.iso

en

Dc Identifier Other

https://www.proquest.com/LegacyDocView/DISSNUM/32039306

OAI identifier oai:identifier

oai:kuscholarworks.ku.edu:1808/36243