Showing 1–1 of 1 results for author: Kurtz, Y
-
Group Orthogonalization Regularization For Vision Models Adaptation and Robustness
Authors:
Yoav Kurtz,
Noga Bar,
Raja Giryes
Abstract:
As neural networks become deeper, the redundancy within their parameters increases. This phenomenon has led to several methods that attempt to reduce the correlation between convolutional filters. We propose a computationally efficient regularization technique that encourages orthonormality between groups of filters within the same layer. Our experiments show that when incorporated into recent ada…
▽ More
As neural networks become deeper, the redundancy within their parameters increases. This phenomenon has led to several methods that attempt to reduce the correlation between convolutional filters. We propose a computationally efficient regularization technique that encourages orthonormality between groups of filters within the same layer. Our experiments show that when incorporated into recent adaptation methods for diffusion models and vision transformers (ViTs), this regularization improves performance on downstream tasks. We further show improved robustness when group orthogonality is enforced during adversarial training. Our code is available at https://github.com/YoavKurtz/GOR.
△ Less
Submitted 18 February, 2024; v1 submitted 16 June, 2023;
originally announced June 2023.
