Summary
A new arXiv paper (2611.09878) by Tanush Shaska introduces graded convolutional networks built on the Gaussian jet of an image, treated as a graded vector space where dilation by λ multiplies order-k derivatives by λ^{-k}. The author shows that local graded convolutions are exactly homogeneous differential operators, and that a local feature is dilation-invariant if and only if it factors through the scale-normalized jet, or—without using the scale—through a retraction onto the weighted sphere of the grading. Under unknown Gaussian blur, this retraction changes only by a scale factor, while the scale-normalized jet acquires a grading factor that no radial normalization can remove. At the zero jet, the retraction separates image families that radial normalizations and continuous classifiers cannot. Networks based on the retraction are universal for invariant features. Experiments: on MNIST the network retains 88.4% accuracy under a two-pixel blur versus 66.2% with scale-normalized derivatives, and improves Lindeberg's Gaussian-derivative network from 91.5% to 95.1% under the same blur.
Paper Overview
Research area: Computer Vision
Author: Tanush Shaska
Published: 2026-11-06
arXiv: 2611.09878
Summary
The Gaussian jet of an image is a graded vector space: a dilation by \(λ\) multiplies derivatives of order \(k\) by \(λ^{-k}\). Local graded convolutions are exactly homogeneous differential operators, and a local feature is dilation-invariant if and only if it factors through the scale-normalized jet or, without the scale, through the retraction onto the weighted sphere of the grading.
Under unknown Gaussian blur, the retraction changes only by a change of scale, whereas the scale-normalized jet acquires a grading factor that no radial normalization removes. At the zero jet, the retraction separates image families that radial normalizations and continuous classifiers cannot.
Key Results
- Networks built on the retraction are universal for invariant features.
- On MNIST under a two-pixel blur:
- Retraction-based network: 88.4% accuracy
- Same network using scale-normalized derivatives: 66.2%
- Applied to Lindeberg's Gaussian-derivative network, the retraction raises accuracy under that blur from 91.5% to 95.1%.
Links
- arXiv: https://arxiv.org/abs/2611.09878
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178636980