AdS-GNN - a Conformally Equivariant Graph Neural Network

Maksim Zhdanov AMLab, University of Amsterdam [email protected] Nabil Iqbal ∗ Department of Mathematical Sciences, Durham University AMLab, University of Amsterdam [email protected] Erik Bekkers AMLab, University of Amsterdam [email protected] &Patrick Forré AMLab, AI4Science Lab, University · 2025

Read AdS-GNN - a Conformally Equivariant Graph Neural Network by Maksim Zhdanov AMLab, University of Amsterdam [email protected] Nabil Iqbal ∗ Department of Mathematical Sciences, Durham University AMLab, University of Amsterdam [email protected] Erik Bekkers AMLab, University of Amsterdam [email protected] &Patrick Forré AMLab, AI4Science Lab, University (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-12880.txt.

Introduction

Angle-keeping maps matter in physics and vision, and a neural network that respects them could learn faster. Maksim Zhdanov and co-authors build such a network in a 2025 paper. Their trick lifts flat data into a curved space called anti-de Sitter space. In that space, angle-keeping changes of the flat data act as distance-keeping moves, which are easier to build into a network.

Layers pass messages based on true distances in the curved space, keeping the cost of computing low. Tests came from vision tasks and from statistical physics setups. The network performed strongly and generalized well to new cases. It could even read off key scaling numbers learned from data. The authors frame it as a bridge between geometry and learning.

Catalog record

CREATORMaksim Zhdanov AMLab, University of Amsterdam [email protected] Nabil Iqbal ∗ Department of Mathematical Sciences, Durham University AMLab, University of Amsterdam [email protected] Erik Bekkers AMLab, University of Amsterdam [email protected] &Patrick Forré AMLab, AI4Science Lab, University
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE114 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-12880

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

AdS-GNN - a Conformally Equivariant Graph Neural Network Maksim Zhdanov AMLab, University of Amsterdam [email protected] Nabil Iqbal ∗ Department of Mathematical Sciences, Durham University AMLab, University of Amsterdam [email protected] Erik Bekkers AMLab, University of Amsterdam [email protected] &Patrick Forré AMLab, AI4Science Lab, University of Amsterdam [email protected] Abstract Conformal symmetries, i.e. coordinate transformations that preserve angles, play a key role in many fields, including physics, mathematics, computer vision and (geometric) machine learning. Here we build a neural network that is equivariant under general conformal transformations. To achieve this, we lift data from flat Euclidean space to Anti de Sitter (AdS) space. This allows us to exploit a known correspondence between conformal transformations of flat space and isometric transformations on the AdS space. We then build upon the fact that such isometric transformations have been extensively studied on general geometries in the geometric deep learning literature. We employ message-passing layers conditioned on the proper distance, yielding a computationally efficient framework. We validate our model on tasks from computer vision and statistical physics, demonstrating strong performance, improved generalization capacities, and the ability to extract conformal data such as scaling dimensions from the trained network. 1 Introduction The notion of symmetry is a key tool both in our understanding of nature and for the construction of machine learning systems that perceive nature. The construction of equivariant neural network architectures that encode specific symmetries has powerful advantages both conceptual and computational. In particular, much work has been dedicated to building networks that are equivariant under familiar symmetries such as rotations and translations. In this work, we will study the symmetry group of conformal transformations i.e. the set of transformations on \mathbb{R}^{d} that preserve angles. This includes translations, rotations, reflections, scalings and so called special conformal transformations. Scale transformations – i.e. rigid rescalings of the whole system – are an extremely important subgroup of the conformal group. Conformal and scale invariance play a central role in many diverse fields. To give some examples: biological visual systems seem to exhibit insensitivity to scale [ 1 , 2 ] . Physical systems undergoing a second-order phase transition have fluctuations at all scales; they are generally conformally invariant at the critical point [ 3 ] , and can usefully be described by conformal field theories [ 4 ] . Indeed in physics most systems exhibiting scale invariance also exhibit conformal invariance “for free” [ 5 ] , [ 6 ] . Diverse applications also exist in computational geometry and computer vision (see e.g. [ 7 , 8 ] ) It is reasonable to believe that a neural network that is equivariant under conformal transformations will be naturally insensitive to scale, instead focusing on robust properties of shape and form, lending to many possible applications. In this work, we construct such a conformally equivariant network. Our approach acts on point clouds and lifts the data into an auxiliary higher dimensional space called Anti de Sitter ( \AdS ) space. As outlined below, this approach is inspired by ideas in conformal field theory in theoretical physics. We validate our construction on some tasks drawn from computer vision and statistical physics. 2 Previous work Construction of equivariant neural networks under general symmetry groups began with [ 9 ] and has since evolved into a rich, well-studied field (see, e.g., [ 10 , 11 ] for reviews). Advancements include treatments of isometric transformations on general Riemannian manifolds [ 11 ] and extensions to semi-Riemannian manifolds [ 12 , 13 ] . The significant benefits of incorporating inherent equivariance, particularly for point cloud data and even at large computational scales, are increasingly demonstrated [ 14 , 15 ] , highlighting advantages in performance and efficiency. Lie group theory has been employed to develop equivariant networks for broad classes of transformations, such as general affine transformations [ 16 ] , or to leverage specific algebraic structures like adjoint actions on Lie algebras [ 17 ] . However, despite progress in handling Euclidean symmetries, scale equivariance, or even these more general Lie groups, these methods do not address the full conformal group, which uniquely includes non-affine special conformal transformations. The pursuit of scale equivariance, a critical component of conformal symmetry, is perhaps the closest related area to our work on full conformal equivariance, with various approaches developed from specialized convolutional architectures [ 18 , 19 ] to techniques like Fourier layers for robust scale handling [ 20 ] . Our use of \AdS space connects conceptually to scale space theories [ 21 , 22 ] , where the extra dimension corresponds to scale. However, a crucial distinction is that the geometry of \AdS space enforces equivariance under the larger group of all conformal transformations by construction, not just scale or isometric transformations. This approach also finds resonance in physics, for instance, with recent explorations of using neural networks to model aspects of conformal field theory [ 23 ] . \mathbf{x}_{i},\mathbf{h}_{i} X_{i},\mathbf{h}_{i} AdS Embedding \psi\left(\mathbf{h}_{i},\mathbf{h}_{j},D(X_{i},X_{j})\right) AdS-GNN Message boundary x^{2} x^{1} z label: 0 Conf. transf-n Figure 1 : AdS-GNN lifts points from Euclidean space to Anti de Sitter space and computes message passing conditioned on the proper distance. 3 Conformal symmetry and Anti de Sitter space We briefly review conformal symmetry and the geometry of \AdS space before describing our neural network. 3.1 Conformal transformations A detailed and self-contained account of conformal transformations is given in Appendix D ; here we present a brief review. Formally, a global conformal transformation of the Euclidean space \mathbb{R}^{d} is an injective smooth map \varphi:\,\mathbb{R}^{d}\setminus\{x_{\varphi}\}\to\mathbb{R}^{d} , x\mapsto\varphi(x) , defined on \mathbb{R}^{d} except on a possible point x_{\varphi} 1 1 1 By allowing \varphi to not be defined on a certain point x_{\varphi}\in\mathbb{R}^{d} , we effectively allow \varphi to map x_{\varphi} to the “points at infinity” \infty in the conformal compactification \mathbb{S}^{d} of \mathbb{R}^{d} . In fact, every global conformal transformation of \mathbb{R}^{d} uniquely extends to an angle-preserving diffeomorphism of \mathbb{S}^{d} for d\geq 2 , see [ 24 ] Thm. 2.6-2.11, or D.4.2 and D.4.3 . This is why we introduce the definition with x_{\phi} in this way. such that local angles are always preserved, i.e. for all x\in\mathbb{R}^{d}\setminus\{x_{\varphi}\} and v_{1},v_{2}\in\mathbb{R}^{d}\setminus\{0\} we require 2 2 2 An equivalent, but more general and abstract, definition is provided in D.1.2 and D.1.1 . : \displaystyle\frac{\langle\varphi^{\prime}(x)v_{1},\varphi^{\prime}(x)v_{2}\rangle}{\|\varphi^{\prime}(x)v_{1}\|\cdot\|\varphi^{\prime}(x)v_{2}\|}=:\cos(\measuredangle(\varphi^{\prime}(x)v_{1},\varphi^{\prime}(x)v_{2}))\overset{!}{=}\cos(\measuredangle(v_{1},v_{2})):=\frac{\langle v_{1},v_{2}\rangle}{\|v_{1}\|\cdot\|v_{2}\|}, (1) where \langle.,.\rangle and \|.\| denote the standard Euclidean scalar product and norm, resp., and \varphi^{\prime}(x) the Jacobian matrix of \varphi at x . Note, that, in contrast to isometric transformations , we do not require that the distances/norms are preserved. The group of all global conformal transformations of \mathbb{R}^{d} is denoted by \GConf(\mathbb{R}^{d}) . D.4.2 (see [ 24 ] Thm. 2.6-2.11) shows that for d\geq 2 the global conformal group \GConf(\mathbb{R}^{d}) is isomorphic to the projective orthogonal group : \POr(d+1,1):=\Or(d+1,1)/\{\pm 1\} , where the role of the quotient is described in Appendix D in D.4.1 . It is instructive to consider the action of the group on points in terms of separated parameters. In particular, a general element G of the group can be written in terms of parameters (\lambda,t,b,M)\in\mathbb{R}_{>0}\times\mathbb{R}^{d}\times\mathbb{R}^{d}\times\Or(d) and acts on a point x\in\mathbb{R}^{d} through the composition of maps defined by: \displaystyle x^{\prime}=x+t\qquad \displaystyle x^{\prime}=Mx\qquad \displaystyle x^{\prime}=\lambda x\qquad \displaystyle\frac{x^{\prime}}{\|x^{\prime}\|^{2}}=\frac{x}{\|x\|^{2}}-b, (2) translations rotations scalings special conformal transformations

Read the complete text