On Symmetries in Convolutional Weights

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Alsallakh, Bilal, Wroge, Timothy, Miglani, Vivek, Kokhlikyan, Narine
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909588710752256
author Alsallakh, Bilal
Wroge, Timothy
Miglani, Vivek
Kokhlikyan, Narine
author_facet Alsallakh, Bilal
Wroge, Timothy
Miglani, Vivek
Kokhlikyan, Narine
contents We explore the symmetry of the mean k x k weight kernel in each layer of various convolutional neural networks. Unlike individual neurons, the mean kernels in internal layers tend to be symmetric about their centers instead of favoring specific directions. We investigate why this symmetry emerges in various datasets and models, and how it is impacted by certain architectural choices. We show how symmetry correlates with desirable properties such as shift and flip consistency, and might constitute an inherent inductive bias in convolutional neural networks.
format Preprint
id arxiv_https___arxiv_org_abs_2503_19215
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle On Symmetries in Convolutional Weights
Alsallakh, Bilal
Wroge, Timothy
Miglani, Vivek
Kokhlikyan, Narine
Computer Vision and Pattern Recognition
We explore the symmetry of the mean k x k weight kernel in each layer of various convolutional neural networks. Unlike individual neurons, the mean kernels in internal layers tend to be symmetric about their centers instead of favoring specific directions. We investigate why this symmetry emerges in various datasets and models, and how it is impacted by certain architectural choices. We show how symmetry correlates with desirable properties such as shift and flip consistency, and might constitute an inherent inductive bias in convolutional neural networks.
title On Symmetries in Convolutional Weights
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2503.19215