0645094 20260213094513.020260126235959.9 27 s. E cav_un_epca*06451022640-3498San DiegoML Research Press2026 transformers roto-translation geometric deep learning invariance equivariance robustness cav_un_auth*0438860 Karella Tomáš UTIA-B Zpracování obrazové informace Department of Image Processing ZOI ZOI CZ Ústav teorie informace a automatizace AV ČR, v. v. i. cav_un_auth*0439197 Harmanec Adam UTIA-B Zpracování obrazové informace Department of Image Processing ZOI ZOI Ústav teorie informace a automatizace AV ČR, v. v. i. cav_un_auth*0293863 Kotera Jan UTIA-B Zpracování obrazové informace Department of Image Processing ZOI ZOI Department of Image Processing CZ Ústav teorie informace a automatizace AV ČR, v. v. i. cav_un_auth*0254045 Blažek Jan UTIA-B Zpracování obrazové informace Department of Image Processing ZOI ZOI Department of Image Processing CZ Ústav teorie informace a automatizace AV ČR, v. v. i. cav_un_auth*0101209 Šroubek Filip UTIA-B Zpracování obrazové informace Department of Image Processing ZOI ZOI Department of Image Processing Ústav teorie informace a automatizace AV ČR, v. v. i. pdf 11MB https://library.utia.cas.cz/separaty/2026/ZOI/sroubek-0645094.pdf GA24-10069S GA ČR CZ cav_un_auth*0472834 VJ02010029 GA MV CZ cav_un_auth*0449227 Convolutional Neural Networks exhibit inherent equivariance to image translation, leading to efficient parameter and data usage, faster learning, and improved robustness. The concept of translation equivariant networks has been successfully extended to rotation transformation using group convolution for discrete rotation groups and harmonic functions for the continuous rotation group encompassing 360°. We explore the compatibility of the Self-Attention mechanism with full rotation equivariance, in contrast to previous studies that focused on discrete rotation. We introduce the Harmformer, a harmonic transformer with a convolutional stem that achieves equivariance for both translation and continuous rotation. Accompanied by an end-to-end equivariance proof, the Harmformer not only outperforms previous equivariant transformers, but also demonstrates inherent stability under any continuous rotation, even without seeing rotated samples during training. cav_un_auth*0502239 NeurIPS 2024 Workshop on Symmetry and Geometry in Neural Representations 20241214 20241215 Vancouver CA 2027 JD PR RVO:67985556 https://hdl.handle.net/11104/0374942 S https://arxiv.org/pdf/2411.03794 2026 SCOPUS PUBMED WOS cav_un_epca*0645102 NeurIPS 2024 Workshop on Symmetry and Geometry in Neural Representations ML Research Press 2026 San Diego 2640-3498 2640-3498