Riemannian Geometry for Brain Computer Interfaces

91 concepts66 papers1 proposal

Traditional brain-computer interfaces (BCIs) often struggle with the messy reality of electroencephalogram (EEG) data. Raw neural signals are highly non-stationary, prone to environmental noise, and heavily corrupted by biological artifacts like eye blinks or muscle tension. Furthermore, these signals exhibit massive variability across different recording sessions and individual subjects.

Historically, signal processing pipelines treated EEG data as flat vectors in Euclidean space, focusing on raw temporal waveforms or spectral power bands. However, as Ce Ju et al. observe, conventional deep learning models frequently fail to capture the inherent spatial relationships of brainwaves when treating them as standard Euclidean inputs, resulting in unstable or suboptimal feature representations. To overcome these limitations, a powerful paradigm shift models brain activity using the non-Euclidean geometry of covariance matrices.

The Geometry of Brain Connectivity

Instead of analyzing individual channel amplitudes, we can represent multi-channel neural signals by their spatial covariance matrices. A spatial covariance matrix captures how strongly the electrical activity across different electrodes synchronizes. Mathematically, these matrices are Symmetric Positive Definite (SPD).

SPD matrices do not conform to flat Euclidean geometry. If you perform standard linear operations on them—such as adding or subtracting matrices directly—you can easily destroy their mathematical properties, yielding matrices that are no longer positive definite. Instead, SPD matrices naturally reside on a curved Riemannian manifold (Ce Ju et al.; David Sabbagh et al.).

By treating covariance matrices as points on this curved manifold, algorithms can compute distances along the manifold's actual curvature (geodesics) rather than assuming a straight line through flat space. As Hadas Benisty et al. demonstrate, accounting for this curvature is essential for capturing the true dynamics of neural communication, significantly improving decoding accuracy compared to standard Euclidean approaches.

Why the Manifold Filters Noise

Operating on a curved Riemannian manifold provides a natural defense against the vulnerabilities of raw EEG data. When biological noise—such as muscle tension or eye blinks—corrupts an EEG recording, it alters the signal amplitudes drastically but shifts the spatial covariance in predictable geometric ways.

Emmanuel K. Kalunga et al. show that representing brain signal covariance matrices on a curved Riemannian manifold inherently filters out these biological artifacts. Because the geometric framework respects the mathematical structure of the data, it separates the true neural interactions from the noise. Furthermore, Zirui Wang et al. establish that representing neural signals on Riemannian manifolds provides inherent invariance to signal scaling and non-stationarity, making models significantly more robust to sensor noise or signal corruption than traditional Euclidean deep learning methods.

Overcoming Session and Subject Variability

The greatest hurdle in deploying real-world BCIs is distribution shift: a model trained on one person, or even on the same person on a Tuesday, often fails on a Wednesday. Riemannian geometry addresses this by focusing on the underlying, invariant geometry of the brain states.

In a study on cross-session workload estimation, Simon Ladouce notes that Riemannian geometry methods demonstrate superior robustness to signal distribution shifts across different recording sessions compared to standard deep learning. This resilience extends to cross-subject transfer learning. Dongrui Wu explains that mapping brain signals to a curved Riemannian manifold handles the inherent noise and non-stationarity of EEG data more effectively than traditional Euclidean methods, preserving the underlying structure of the signals across different users and sessions. Similarly, Lezhi Chen et al. demonstrate that using SPD matrices as an EEG representation reduces the distributional gaps between different subjects, leading to more stable and informative cross-subject classification.

Data Efficiency and Real-Time Inference

Because the Riemannian manifold structure does the heavy lifting of filtering noise and preserving spatial relationships, models do not need to learn these properties from scratch. This leads to exceptional data efficiency.

Ho Jin Choi demonstrates that representing electrode relationships as covariance matrices on a Riemannian manifold enables high classification accuracy with minimal training data. This geometric approach allows simple linear classifiers to outperform complex deep learning models in real-time processing speed and data efficiency, making it highly feasible for real-time robot assistive control.

To perform classification or regression, algorithms typically map the covariance matrices onto a flat "tangent space" touching the manifold or calculate the affine-invariant distance between points and the geometric mean of each class (Dongrui Wu et al.; Alexandre Hippert-Ferrer et al.). This strategy has been successfully applied to diverse tasks, ranging from continuous regression for user reaction times (Dongrui Wu et al.) to predicting biological traits without requiring expensive source modeling (David Sabbagh et al.).

Continuous Tracking of Brain States

Traditional BCI evaluations rely heavily on discrete classification accuracy. However, Camille Jeunet et al. argue that classification accuracy is a poor metric for studying user learning because it obscures continuous progress.

By representing EEG signals as covariance matrices on a Riemannian manifold, researchers can calculate precise, continuous distances between brain states. This geometric approach provides a continuous measure of signal evolution that reveals user learning progress—such as improved brain-signal stability and distinctiveness from rest—that traditional discrete scores completely miss.

Integrating Manifolds with Deep Learning

Modern research does not choose between geometry and deep learning; instead, it merges them. Daniel Wilson et al. introduce Deep Riemannian Networks (such as EE(G)-SPDNet) that combine the strengths of both paradigms by operating directly on the covariance matrices of EEG data, leveraging their geometric properties to preserve inherent data structures.

To optimize these networks for practical use, Igor Carrara et al. developed a geometric deep learning architecture that unfolds brain signal dynamics through phase space reconstruction. This model processes data on curved geometric manifolds, allowing for high decoding accuracy even when using minimal sensor arrays, making BCIs more practical and explainable.

Key Foundational Work

EEG-Based User Reaction Time Estimation Using Riemannian Geometry Features (2017) — contributed 2 concept elements to this direction.
Online SSVEP-based BCI using Riemannian geometry (2016) — contributed 1 concept elements to this direction.
Improving EEG Neurofeedback with Advanced Machine Learning and Signal Processing tools from Brain-Computer Interfaces Research (2017) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 68 papers
  2. Version 1 · 2026-10-04 · 85 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_riemannian_manifold_representation_e759c990,
  title = {Riemannian manifold representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/riemannian-manifold-representation/rdr_e759c990e55c45f4bd4024dfadc56ee1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 2, frozen text):

@misc{lacuna_riemannian_manifold_representation_e759c990_v2,
  title = {Riemannian manifold representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/riemannian-manifold-representation/rdr_e759c990e55c45f4bd4024dfadc56ee1/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake