Neuromanifold-Regularized KANs for Shape-fair Feature Representations
October 9, 2025 · View on GitHub
Official Code Repository for ICCV2025
⚠️ This is the official repository and is under active development.
We will push code, pretrained weights, and additional documentation as the project matures.
Stars ⭐ and watches 👀 help us prioritize public releases!
Table of Contents
Overview
Traditional deep networks often over-fit to fine-scale textures, while unconstrained Kolmogorov-Arnold Networks (KANs) suffer the same issue because their adaptive nonlinearities remain overly expressive. Our work tackles this by constraining KANs to a low-degree neuromanifold and pairing two feature-extractor branches with a novel Style Decorrelation Loss. The resulting model—NeuroManifold-Regularized KAN (NMR-KAN)—learns shape-fair representations that separate local and global cue processing, boosting shape bias by 14.8 pp over baseline convolutional KANs and delivering extra resilience to common corruptions and adversarial attacks.
Citation
If you find our work useful, please cite:
@inproceedings{arslan2025neuromanifold,
title = {Neuromanifold-Regularized KANs for Shape-fair Feature Representations},
author = {Mazlum Ferhat Arslan and Weihong Guo and Shuo Li},
booktitle = {Proceedings of the International Conference on Computer Vision},
year = {2025},
url = {https://arxiv.org/abs/}
}
Contact
For questions or collaboration:
- Ferhat Arslan – mxa1328@case.edu
- Issues ⇒ use the GitHub issue tracker
License
Distributed under the MIT License. See LICENSE for more information.
TODO / Roadmap
- Clean up training scripts 🧹
- Release full inference pipeline ⚡
- Push pretrained weights 📦
Acknowledgements
Template inspired by pytorch-cv, OpenMMLab, and several ICCV-2023 repos.