GeoDynamics: A Geometric State-Space Neural Network for Understanding Brain Dynamics on Riemannian Manifolds

AI summary

Overview

Research area: Geometric deep learning applied to neuroimaging — specifically, modeling brain functional connectivity dynamics as trajectories on the Riemannian manifold of symmetric positive definite (SPD) matrices, using a state-space model (SSM).

Technical level: Advanced. The paper assumes familiarity with state-space models, ordinary differential equations, Riemannian geometry (Stein metric, Fréchet means, group actions), and manifold-valued neural networks.

Scope: The paper introduces GeoDynamics, a geometry-aware SSM that evolves latent brain states directly on the SPD manifold, and validates it on five human brain connectome datasets plus three human action recognition benchmarks.

What This Paper Is About

Most models of brain dynamics treat the brain as a set of loosely connected regions or use oversimplified network priors, which misses the fact that functional connectivity (FC) at any moment is an SPD matrix living on a curved Riemannian manifold rather than in flat Euclidean space. This paper asks what happens if you build a state-space model whose inputs, hidden states, and outputs are all SPD matrices, so that state transitions respect the manifold's geometry. The goal is to track how whole-brain connectivity patterns evolve over time, and to use those trajectories to decode task states and detect disease.

Key Contributions

1. A geometric deep model combining SSMs and manifold learning. GeoDynamics replaces the Euclidean algebra of conventional state-space models with Riemannian geometric algebra, so that system dynamics evolve in a way consistent with the intrinsic structure of FC matrices.
2. Manifold-aware state update and observation operators. Temporal aggregation is performed with a weighted Fréchet mean (an intrinsic barycenter along geodesics) and state transitions are performed with orthogonal group actions, which are isometric under the Stein metric and preserve the SPD property.
3. An SPD-preserving attention (SPA) module. Attention is defined directly on the manifold using exponential mapping and elementwise modulation, so the attention-weighted representation remains SPD. The authors report that the learned attention weights are anatomically and temporally interpretable and can highlight candidate disease biomarkers.
4. Computational efficiency. The authors report improved computational efficiency relative to manifold-based deep models, achieved through a global convolutional reformulation of the recurrence, use of the Stein metric (which avoids repeated eigendecompositions), and the SPD-preserving attention mechanism.

Main Findings

• Task-based fMRI (HCP-WM, 1,081 subjects, 360 regions, eight task events): Sequential models achieved markedly higher accuracy than spatial models, with a pairwise t-test of p < 10⁻⁴ and performance gains of up to 30%. GeoDynamics achieved the best overall performance among the sixteen representative methods evaluated.
• Neurodegenerative disease (AD, PD): Spatial models showed slightly better average performance than sequential models, but the difference was not statistically significant (p = 0.37). GeoDynamics substantially outperformed all baselines, with a significant improvement over the second-best method (p < 0.05).
• Neuropsychiatric disorder (ABIDE): Spatial models slightly outperformed sequential ones, and SPDNet consistently ranked second only to GeoDynamics. The authors attribute the shared advantage of SPDNet and GeoDynamics to manifold-based representation learning plus spatio-temporal modeling.
• Multi-class ADNI classification: GeoDynamics reached 56.00 ± 3.36 accuracy, 60.36 ± 7.67 precision, and 50.83 ± 5.73 F1. The next-best accuracy among all listed baselines was GSN at 52.80 ± 5.31 and TF at 52.00 ± 6.93; among spatial baselines, SPDNet and GNN-AK both reached 52.40 accuracy (SPDNet F1 31.63 ± 8.76; GNN-AK F1 43.20 ± 6.55). Mamba reached 47.20 ± 6.14 accuracy.
• Parameter-matched comparison (HCP-WM, N = 360): GeoDynamics used 14.60M parameters and reached 98.29 ± 0.26 accuracy, 98.18 ± 0.34 precision, and 98.16 ± 0.35 F1. The best-tuned Mamba configuration with 132M parameters reached 97.22 ± 0.63 accuracy; a Mamba with 33.71M parameters reached 97.06 ± 0.62; 27.05M reached 96.76 ± 0.86; 14.07M reached 95.92 ± 0.50; 13.93M reached 96.17 ± 0.11.
• Sliding window robustness (PPMI): GeoDynamics was relatively robust to window size, with optimal performance at moderate values (typically 35–45). Accuracy ranged from 71.01 ± 14.23 (window 55) to 72.01 ± 8.51 (window 45); precision from 71.00 ± 7.89 (window 55) to 76.07 ± 7.33 (window 15); F1 from 67.67 ± 7.81 (window 55) to 71.71 ± 7.29 (window 25).
• Interpretable attention patterns (top-20 strongest connections): In HCP-WM, dominant connections were in the default mode network (DMN) and central executive network, both described as essential for working-memory tasks. In AD (OASIS and ADNI), the most affected regions included the DMN and somatosensory cortex. In PD, key connections appeared in the sensorimotor area, frontal lobe, DMN, and cerebellum. In autism, prominent connections emerged in the temporal and visual cortices.
• Action recognition (HAR): The paper states that GeoDynamics "remains competitive" on the Florence 3D Actions, HDM05, and UTKinect-Action3D benchmarks and describes improved robustness to noisy joints, but the specific accuracy values for these benchmarks are not reported in the provided text (they are referenced only via Figure 5).

Methodology in Plain English

The researchers start from the standard state-space model, which describes how a system's hidden state evolves over time and how it produces observable outputs, governed by two equations (the state equation and the observation equation). In the classical version, all quantities and operators are ordinary vectors and matrices in Euclidean space.

GeoDynamics changes the space the model lives in. Instead of representing a brain state as a vector, each state, input, and output is an SPD matrix — the natural mathematical object for a functional connectivity matrix, which is symmetric and positive definite. These matrices live on a curved Riemannian manifold, and the authors measure distances between them with the Stein metric, chosen because it avoids repeated eigendecompositions.

To make the state-space equations work on this curved space, two Euclidean operations are swapped for geometric ones:

• Weighted Fréchet mean replaces the weighted sum. Instead of adding past states and inputs, the model computes their intrinsic average (barycenter) on the manifold, which keeps the result SPD.
• Orthogonal group action replaces additive transitions. State transitions are applied as a multiplicative transform gVgᵀ using an orthogonal transformation inferred from the input, which is an isometry under the Stein metric and preserves the SPD property.

Temporal discretization uses a matrix-exponential scheme to keep integration stable. Because the recurrence can be rewritten as a convolution over time, the model can be trained in parallel rather than step by step. The final SPD output is mapped to the tangent space at the identity via a logarithmic map — decomposing it by eigendecomposition — so it can be fed to a softmax classifier trained with cross-entropy loss.

On top of the backbone, the SPD-preserving attention module computes per-element attention weights by exponentiating the convolution response and normalizing by its maximum, then multiplying the response elementwise. Because the weights are strictly positive and bounded and the convolution is SPD-constrained, the modulated output stays SPD. Attention is recomputed at each time step, which the authors argue lets the model detect evolving abnormal FC trajectories.

For brain data, FC matrices are built by sliding a window over BOLD time courses and computing correlation matrices centered at each time point. For action data, a root joint is fixed at the hip center, relative 3D displacements of the other joints are computed, covariance matrices are formed from those displacements, and a sliding window produces a sequence of SPD matrices.

Why This Matters

Impact on research. The paper argues that treating functional connectivity as a whole-brain SPD matrix on a manifold yields a more holistic, self-organized view of brain dynamics than region-level or loosely connected network models. It also connects SSM theory (well established in vision and NLP) to non-Euclidean data, and the attention module produces anatomically labeled outputs that can be compared against known disease pathology — the authors report a shared DMN involvement across AD and PD, despite disease heterogeneity.

Real-world applications (as implied by the paper's framing):

• Early diagnosis of Alzheimer's disease, Parkinson's disease, and autism spectrum disorder from resting-state fMRI.
• Detection and interpretation of task-driven brain state changes during working-memory paradigms.
• Biomarker discovery, using attention weights to localize abnormal hubs or subnetworks and to flag altered covariance structure in FC trajectories.
• Human action recognition from 3D skeletal joint data, which the authors present as cross-domain validation of scalability and robustness.

Industry relevance. The paper positions GeoDynamics as applicable to manifold-valued time series generally, which extends beyond neuroimaging to any domain where the observations at each time step are SPD matrices. The reported efficiency gains and the parallel convolutional formulation matter for scaling to large connectome datasets — the experiments span over a thousand subjects per dataset. The paper reports funding from the National Institutes of Health (AG091653, AG068399, AG084375) and the Foundation of Hope.

Future Directions

• Reporting full HAR results. The paper validates on UTKinect, Florence, and HDM05 and claims competitive performance, but the numeric results are not present in the provided content. A follow-up would establish how large the gains are relative to LieNet, ST-NBNN, GR-GCN, DMT-Net, and the LSTM baselines.
• Translating attention maps into validated biomarkers. The attention module highlights disease-relevant regions, but the paper presents these as localizations rather than as prospectively validated clinical markers.
• Extending beyond the four-class clinical settings. Multi-class ADNI classification is reported (56.00 ± 3.36 accuracy), and the gap to binary results suggests room for improvement on finer-grained disease staging.
• Generalizing the geometric formulation. The framework depends on the SPD manifold and the Stein metric; whether other manifold geometries (and other metrics) provide better trade-offs between accuracy and efficiency remains an open question the paper raises but does not resolve.
• No dedicated future work section is reported in the provided content, so the above are open questions implied by the paper's results rather than stated plans by the authors.

Target Audience

• Machine learning researchers working on geometric deep learning, Riemannian manifold-valued neural networks, and state-space models.
• Computational neuroscientists and neuroimaging researchers interested in dynamic functional connectivity and disease classification from fMRI.
• Clinical researchers in Alzheimer's, Parkinson's, and autism who want interpretable spatial-temporal models of brain network disruption.
• Computer vision researchers interested in SPD-matrix representations of skeletal motion for action recognition.
• Readers need comfort with Riemannian geometry, ODE-based dynamical systems, and manifold-valued deep learning; the paper is not an introductory read.

Authors’ abstract

State-space models (SSMs) have become a cornerstone for unraveling brain dynamics, revealing how latent neural states evolve over time and give rise to observed signals. By combining the flexibility of deep learning with the principled dynamical structure of SSMs, recent studies have achieved powerful fits to functional neuroimaging data. However, most existing approaches still view the brain as a set of loosely connected regions or impose oversimplified network priors, falling short of a truly holistic and self-organized dynamical system perspective. Brain functional connectivity (FC) at each time point naturally forms a symmetric positive definite (SPD) matrix, which resides on a curved Riemannian manifold rather than in Euclidean space. Capturing the trajectories of these SPD matrices is key to understanding how coordinated networks support cognition and behavior. To this end, we introduce GeoDynamics, a geometric state-space neural network that tracks latent brain-state trajectories directly on the high-dimensional SPD manifold. GeoDynamics embeds each connectivity matrix into a manifold-aware recurrent framework, learning smooth and geometry-respecting transitions that reveal task-driven state changes and early markers of Alzheimer's disease, Parkinson's disease, and autism. Beyond neuroscience, we validate GeoDynamics on human action recognition benchmarks (UTKinect, Florence, HDM05), demonstrating its scalability and robustness in modeling complex spatiotemporal dynamics across diverse domains.