Neural Sphere: Spherical Structures in Neural Systems

• Neural Sphere is a research program that incorporates spherical geometry into neural models, defining signal domains, decision surfaces, and parameter spaces with explicit geometric structure.
• It refines methodologies by applying spherical convolutions, graph-based approaches, and explicit surface scaffolding to improve rotation equivariance and spatial reasoning.
• By integrating geometric principles into network design, Neural Sphere enhances computational accuracy in applications from neuroimaging to atmospheric forecasting and neural rendering.

In the literature surveyed here, Neural Sphere functions less as the name of a single architecture than as a family of constructions in which the sphere is a mathematically operative object rather than a visualization aid. Depending on the problem, the sphere may serve as a signal domain for equivariant learning on $S^2$, a decision surface for individual neurons, a coarse explicit scaffold for neural implicit geometry, a seamless parameter domain for genus-0 neural surfaces, a registered coordinate system for cortical analysis, or a geometric substrate for explicit model-based reasoning (Melnyk et al., 2021, Dogaru et al., 2022, Williamson et al., 2024, Dong et al., 2024). This suggests that the most precise use of the term is taxonomic: it denotes research programs that insert spherical geometry into the representation, operator, optimization, or inference mechanism of a neural system.

1. Conceptual scope

Across recent work, the sphere appears in several technically distinct roles.

Sphere role	Representative formulation
Signal domain	Spherical CNNs, graph spherical CNNs, spherical neural operators (Feng et al., 2018, Defferrard et al., 2020, Bonev et al., 2023)
Computational primitive	Spherical neurons with sphere decision boundaries (Melnyk et al., 2021)
Auxiliary geometric scaffold	Sphere clouds for empty-space skipping in neural implicit training (Dogaru et al., 2022)
Explicit scene primitive	Sphere-based differentiable rendering and neural shading (Lassner et al., 2020)
Surface parameter domain	Spherical neural surfaces for genus-0 geometry processing (Williamson et al., 2024)
Neuroanatomical coordinate system	Cortical spherical maps, spherical tokenization, spherical embeddings of brain networks (Feng et al., 2018, Chung et al., 2022, Yu et al., 24 May 2026)
Reasoning substrate	Sphere or circle configurations for syllogistic reasoning (Dong et al., 2024, Dong et al., 1 Jan 2026)

What unifies these lines is not a shared network template but a shared commitment: spherical structure is treated as part of the model’s inductive bias. In some cases this bias is group-theoretic, as in $SO(3)$-equivariant convolution; in others it is geometric, as in sphere-based decision surfaces or sphere clouds; in still others it is symbolic-geometric, as in reasoning by constructing sphere configurations. The resulting field spans geometric deep learning, neural rendering, scientific machine learning, neuroimaging, and neuro-symbolic reasoning.

2. Sphere as a computational domain

A major strand of Neural Sphere research concerns signals defined on the two-sphere $S^2$ and operators that respect spherical geometry. In cortical analysis, spherical CNNs replace translational equivariance with rotational equivariance by defining correlations on $S^2$ and $SO(3)$. For a signal $f:S^2\to\mathbb R$ and filter $\psi:S^2\to\mathbb R$, the first-layer spherical correlation is

$(f\star \psi)(R)=\int_{S^2} f(x)\psi(R^{-1}x)\,dx,$

and subsequent layers operate on $SO(3)$-valued features; weighted global average pooling then corrects for nonuniform sampling density with weights proportional to $\sin\beta$ (Feng et al., 2018). The point is not merely to rasterize a sphere, but to preserve the action of rotations through the network.

A more implementation-oriented branch replaces harmonic constructions with graph or pixel methods. The HEALPix CNN of Krachmalnicoff and Tomasi uses a pixel-centered neighborhood consisting of the pixel itself plus its eight ordered neighbors, reducing spherical convolution to a standard 1D convolution on a reordered vector; pooling exploits the hierarchical HEALPix relation $SO(3)$0 (Krachmalnicoff et al., 2019). DeepSphere instead models a sampled sphere as a sparse graph and uses polynomial graph filters

$SO(3)$1

with approximate rotation equivariance controlled by graph resolution, neighborhood count, and kernel width (Defferrard et al., 2020). In that formulation, equivariance is not exact but becomes a tunable property of the discretization.

Operator learning on the sphere pushes the same idea into dynamical systems. Spherical Fourier Neural Operators replace the Euclidean FFT with the spherical harmonic transform and use the spherical convolution theorem

$SO(3)$2

so that learned filters depend on spherical harmonic degree $SO(3)$3 rather than Euclidean wavevectors (Bonev et al., 2023). In atmospheric forecasting this geometry-aware construction produced stable autoregressive rollouts for 1,460 steps, corresponding to a year of simulated time, whereas FFT-based models developed polar artifacts, spurious waves, and excessive diffusion over long horizons (Bonev et al., 2023). A related but PDE-oriented line appears in physics-informed neural networks for the shallow-water equations on the sphere, where latitude–longitude coordinates are embedded through

$SO(3)$4

so that periodicity in longitude and continuity at the poles are hard-encoded in the input representation (Bihlo et al., 2021).

3. Sphere as a computational primitive

A second interpretation treats the sphere not as the domain of a field but as the basic decision object of a neuron. In the 3D spherical-neuron framework of Melnyk and Koniusz, Euclidean points are conformally embedded so that a learned sphere becomes a linear classifier in the embedding space. In the $SO(3)$5 coordinate form,

$SO(3)$6

and the neuron response is

$SO(3)$7

The decision surface is therefore the Euclidean sphere $SO(3)$8, positive inside and negative outside (Melnyk et al., 2021). For point clouds, a geometric neuron sums such responses over points,

$SO(3)$9

yielding a higher-order sphere-based decision geometry (Melnyk et al., 2021).

The same paper derives an exact 3D steerability result. Because the response of a rotated spherical neuron depends only on spherical harmonics up to degree $S^2$0, a four-function tetrahedral basis suffices in 3D, with

$S^2$1

A learned sphere can be converted after training into a four-element steerable bank $S^2$2, and the response under a known rotation $S^2$3 can be recovered by interpolation coefficients $S^2$4 satisfying

$S^2$5

This yields an exact rotation-invariant construction for the known-rotation case and shows that sphere-based neurons can support group-theoretic structure without being trained as constrained equivariant filters from the outset (Melnyk et al., 2021).

4. Sphere as scaffold, primitive geometry, and surface parameterization

In 3D reconstruction and neural rendering, the sphere often appears as an explicit geometric scaffold rather than the final implicit field itself. Dogaru, Rosen, and Paris introduce a jointly learned sphere cloud for neural implicit surface training. The final surface remains

$S^2$6

but the sphere cloud restricts volumetric sampling to the union of ray–sphere intersection intervals, thereby excluding empty volume before neural evaluations are performed (Dogaru et al., 2022). Sphere centers are optimized by

$S^2$7

and the common radius follows

$S^2$8

Because ray–sphere intersections are analytic, the method excludes empty space without additional forward passes of the neural surface network and improves Chamfer distance and rendering metrics across UNISURF, VolSDF, NeuS, and NeuralWarp baselines (Dogaru et al., 2022).

Pulsar takes a more literal sphere-based route by representing a scene as a set of learned spheres

$S^2$9

where $S^2$0 is center, $S^2$1 radius, $S^2$2 opacity, and $S^2$3 a feature or color vector (Lassner et al., 2020). Rendering is differentiable and uses a soft visibility rule

$S^2$4

combining depth, opacity, and orthogonal ray distance (Lassner et al., 2020). The system is reported to render and optimize representations with millions of spheres, with forward and backward times well below those of competing differentiable renderers in the reported benchmarks (Lassner et al., 2020). Here the sphere is both the scene primitive and the differentiable rendering substrate.

A complementary development is the use of the sphere as a seamless parameter domain for genus-0 surfaces. Spherical Neural Surfaces define an overfitted map

$S^2$5

and compute geometric quantities directly from this neural map without meshing (Williamson et al., 2024). The first fundamental form is

$S^2$6

the second fundamental form is

$S^2$7

and curvatures follow from

$S^2$8

From these quantities the paper derives surface gradient, divergence, Laplace–Beltrami operator, spectral modes, heat flow, and mean curvature flow directly on the neural surface (Williamson et al., 2024). This suggests a broader Neural Sphere pattern: the sphere can serve as a global, seam-free latent domain on which differential geometry is carried out natively.

5. Neural Sphere in neuroimaging and brain science

Neuroimaging provides some of the most natural uses of sphere-based neural modeling because the cortex is routinely reconstructed and registered on spherical coordinates. In Alzheimer’s disease diagnosis, Zhao, Dey, and Hong formulate classification directly on left- and right-hemisphere cortical thickness maps sampled on a bandwidth-$S^2$9 spherical grid, with each hemisphere processed by a shared spherical CNN trunk and combined only after weighted global average pooling (Feng et al., 2018). Their architecture

$SO(3)$0

outperformed a matched planar CNN baseline on both tasks studied. For AD vs CN, the spherical CNN achieved AUC 0.915 versus 0.895, with accuracy $SO(3)$1 versus $SO(3)$2; for MCI progression, it achieved AUC 0.707 versus 0.657, with accuracy $SO(3)$3 versus $SO(3)$4 (Feng et al., 2018). The underlying claim is geometric: cortical thinning is a signal on a sphere-like manifold, and planar flattening either distorts geometry or alters topology.

A distinct but related use of the sphere appears in functional connectomics. Chung and colleagues observe that centered, unit-norm regional time-series vectors already lie on a sphere, so Pearson correlation becomes an angular quantity, and the natural distance is

$SO(3)$5

not the non-metric quantity $SO(3)$6 (Chung et al., 2022). They then derive a spherical multidimensional scaling objective

$SO(3)$7

and a spectral approximation yielding an $SO(3)$8 embedding of large correlation networks (Chung et al., 2022). On an HCP-derived $SO(3)$9 functional network, the correlation between original and embedded distances was reported as 0.51 for spherical MDS and 0.0501 for hyperbolic MDS (Chung et al., 2022). In this setting, the sphere is the manifold induced by correlation geometry.

Diffusion MRI adds a different neural-spherical factorization. Neural Spherical Harmonics model the signal

$f:S^2\to\mathbb R$0

where the coefficient field $f:S^2\to\mathbb R$1 is predicted by an MLP from spatial coordinates and the angular dependence is analytic in a spherical harmonics basis (Hendriks et al., 2023). This yields a continuous field of spherical signals over space, trained from a single subject, and the reconstructed signal was reported to show smoother fiber-orientation distributions and more structurally coherent variation across neighboring voxels than voxelwise SH interpolation, even when RMSE to noisy measured data was not always lower (Hendriks et al., 2023).

The most recent sphere-based brain decoding framework in the corpus, NeurIPS, uses the sphere as a cortical alignment and tokenization substrate rather than as a harmonic basis. Inputs are GLM beta weights on visual cortical ROIs registered to fsaverage6, and the model introduces a Selective ROI Spherical Tokenizer together with a Structure-Guided Mixture of Experts conditioned on cortical morphology (Yu et al., 24 May 2026). The default NSD-General ROI contains 9,488 vertices total rather than the 81,924 vertices of full fsaverage6 cortex, and the paper reports that the resulting surface decoder reaches a new state of the art among surface methods while converging dramatically faster in adaptation settings, namely 10 vs. 600 epochs (Yu et al., 24 May 2026). A central claim is that cortical thickness, surface area, sulcal depth, and curvature should be treated not as nuisance variation but as predictive inductive priors.

6. Reasoning, manifold-constrained learning, and broader extensions

The most literal use of the phrase Sphere Neural Networks appears in neuro-symbolic reasoning. In the 2024 SphNN framework, a concept is represented by a sphere $f:S^2\to\mathbb R$2 with center $f:S^2\to\mathbb R$3 and radius $f:S^2\to\mathbb R$4, and syllogistic relations are mapped to spatial relations such as parthood

$f:S^2\to\mathbb R$5

and disconnectedness

$f:S^2\to\mathbb R$6

(Dong et al., 2024). Validity is defined by countermodel non-existence: premises are conjoined with the negation of the conclusion, and the system attempts to construct a sphere configuration satisfying them. The paper states that for any three satisfiable syllogistic statements, the method can determine satisfiability in the first epoch, and for long-chained reasoning it reports computational complexity $f:S^2\to\mathbb R$7 (Dong et al., 2024). Here the sphere is explicitly the representational unit of rational inference.

The 2026 extension moves from spheres in Euclidean space to circles on the surface of an $f:S^2\to\mathbb R$8-dimensional sphere, chiefly to support negation via complement circles (Dong et al., 1 Jan 2026). Atomic membership, universal implication, and universal negation become circle-containment statements such as

$f:S^2\to\mathbb R$9

and disjunctive syllogism is handled by constructing candidate circle configurations for each disjunct and filtering out unsatisfiable ones (Dong et al., 1 Jan 2026). The paper reports that this extended Sphere Neural Network solved 16 syllogistic reasoning tasks, determined all 64 tested reasoning types correctly in one experiment, and preserved 100\% performance on 256 classical syllogism types in another (Dong et al., 1 Jan 2026). The underlying philosophy is that reasoning should proceed by explicit model construction rather than by statistical pattern matching.

A less symbolic but still sphere-centric line uses the sphere as a parameter manifold. In hierarchical classification, child classifiers are constrained to lie on spheres centered at parent classifiers through

$\psi:S^2\to\mathbb R$0

so that the final layer becomes

$\psi:S^2\to\mathbb R$1

with $\psi:S^2\to\mathbb R$2 encoding the label hierarchy and $\psi:S^2\to\mathbb R$3 a depth-dependent radius schedule (Scieur et al., 2021). This replaces a flat classifier with a network of connected sphere manifolds and improved performance on CIFAR100, CUB200, Stanford Dogs, Stanford Cars, and Tiny-ImageNet in the reported experiments (Scieur et al., 2021). In scale-invariant optimization, normalized networks are trained directly on a sphere of fixed radius $\psi:S^2\to\mathbb R$4 via

$\psi:S^2\to\mathbb R$5

which reveals three regimes—convergence, chaotic equilibrium, and divergence—depending on the effective learning rate $\psi:S^2\to\mathbb R$6 (Kodryan et al., 2022). In these works, the sphere is not a data domain but the intrinsic geometry of the parameter space.

A peripheral but instructive extension appears in communications, where deep learning based sphere decoding uses a DNN to predict a short list of radii for a classical sphere decoder rather than replacing the decoder itself (Mohammadkarimi et al., 2018). The learned module approximates the distances to the nearest lattice points and thereby reduces the number of lattice points inside the decoding hypersphere while keeping performance close to maximum-likelihood decoding (Mohammadkarimi et al., 2018). This use is conceptually narrower than the others: the sphere is a search region, not the representational substrate. Even so, it reinforces the general pattern that once spherical structure is made explicit, it can become the locus of learned control.

Taken together, these strands show that Neural Sphere is best understood as a geometric research program rather than a single model family. The sphere may encode topology, symmetry, neighborhood structure, parameter scale, explicit spatial extent, or logical complementarity. What varies from paper to paper is the level at which spherical structure enters; what remains constant is the claim that neural computation improves when that structure is made native rather than incidental.