ManifoldFormer: Geometric Deep Learning for Neural Dynamics on Riemannian Manifolds
ManifoldFormer: Geometric Deep Learning for Neural Dynamics on Riemannian Manifolds
Authors: Yihang Fu, Lifang He, Qingyu Chen Date: 2026-04-21 Paper ID: openalex:2511.16828
Summary
ManifoldFormer addresses the mismatch between generic Euclidean-based time series models and the intrinsic manifold structure of neural dynamics in EEG signals. The architecture combines RiemannianVAE for manifold embedding, a GeometricTransformer with geodesic-aware attention, and Neural ODEs for manifold-constrained temporal evolution. This approach enables more robust cross-subject generalization and aligns model representations with neurophysiological principles. Empirical results on four datasets demonstrate that incorporating these geometric constraints significantly outperforms traditional Euclidean-based state-of-the-art methods.
Key Contributions
- Introduces ManifoldFormer, a geometric deep learning framework that explicitly enforces Riemannian manifold constraints on EEG signal representations.
- Proposes a geodesic-aware attention mechanism within a transformer architecture to better capture intrinsic neural dynamics compared to Euclidean-space models.
- Achieves 4.6-4.8% accuracy gains and 6.2-10.2% improvements in Cohen’s Kappa across four public EEG datasets.
Open Questions & Future Work
Key Concepts
- manifoldformer: A geometric deep learning framework that leverages Riemannian manifold embedding and geodesic-aware attention for temporal dynamics modeling.
Archivist Review
The paper introduces a coherent framework for handling non-Euclidean signal data by combining manifold embedding, geodesic-aware attention, and neural ODEs. I approved the framework itself and the open question regarding the automation of manifold discovery, as these reflect the core architectural shift away from Euclidean-only time series modeling. Specific subcomponents like the geodesic-aware attention mechanism were rejected as they are integral parts of the primary framework rather than standalone concepts.
Approved Concepts
- ManifoldFormer: It is the central framework of the paper, demonstrating a novel integration of geometric deep learning with neural signal analysis to overcome Euclidean limitations.
Approved Open Questions
- Optimal Neural Manifold Discovery: The geometric structure significantly influences model performance, and transitioning from manual selection to automated/learned manifold discovery is critical for model scalability and robustness.
Rejected Candidates
- [concept] Geodesic-aware Attention Mechanism (
geodesic-aware-attention-mechanism) - subcomponent_of_broader_mechanism: This is a specific subcomponent of the ManifoldFormer framework.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2511.16828
- paper_id
- 2511.16828
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- manifoldformer
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-24T05:52:27Z