Exploring Symmetry in Parseval CNNs: Smarter Filters for Image Processing
Borna Khodabandeh, External Student at EPFL

Meeting • 2024-09-27

Abstract
In this presentation, I'll go through how symmetric filters can make Parseval Convolutional Neural Networks (CNNs) more efficient, especially for tasks like image denoising. Parseval convolutional operators help maintain energy during transformations, ensuring stability and robustness, which is particularly useful in tasks like denoising. Symmetry further reduces the number of parameters, making models smaller and faster while keeping performance strong. I’ll share some of the theoretical work behind symmetric filters, the challenges we faced, and how we tackled them. We’ll see how these symmetric architectures can outperform conventional ones, even with fewer parameters, and why this approach has potential for improving neural networks in image processing.