Standard deep learning architectures, most notably Convolutional Neural Networks (CNNs), are fundamentally designed for Euclidean spaces. They operate on flat, regular grids where every pixel or data point has a fixed number of equidistant neighbors. While this works exceptionally well for images and audio, it creates a severe architectural mismatch when applied to physical sensor arrays, such as electroencephalography (EEG) or magnetoencephalography (MEG) caps.
As noted by Yiluan Guo et al. in Deep neural networks on graph signals for brain imaging analysis, these sensors are distributed irregularly across curved, three-dimensional surfaces. Forcing this non-Euclidean spatial distribution into a flat 2D matrix or a 1D vector introduces a "Euclidean bias" that distorts the underlying spatial relationships.
Andac Demir et al. point out in EEG-GNN: Graph Neural Networks for Classification of Electroencephalogram (EEG) Signals that grid-based models inherently assume equidistant relationships between adjacent elements in a matrix. In reality, two sensors placed next to each other in a flattened data array might sit on opposite sides of the scalp, while physically adjacent sensors are treated as completely unrelated. This structural mismatch prevents models from accurately capturing how signals propagate across the physical geometry of the system.
Physical Proximity versus Functional Connectivity
The problem runs deeper than just physical coordinates. In biological networks like the human brain, the most critical relationships are often defined by functional connectivity—how strongly two regions synchronize their activity—rather than their physical distance in the skull.
As Jingcong Li et al. observe in STGATE: Spatial-temporal graph attention network with a transformer encoder for EEG-based emotion recognition, the brain operates as a dynamic network where distant regions constantly communicate to process information. Traditional convolutional filters slide locally over a grid, meaning they can only capture interactions between sensors that happen to be adjacent in the data structure.
This limitation is explicitly addressed by Kanghan Oh et al. in Diagnosis of schizophrenia with functional connectome data: a graph-based convolutional neural network approach. The authors argue that in a functional connectome, the importance of a connection is defined by functional relationships rather than physical proximity. Treating network correlations as a flat vector or a rigid image patch completely discards the underlying network topology, blinding the model to the complex, non-local synchronization patterns characteristic of neurological states.
Mapping Sensors to Graphs
To resolve this mismatch, a rapidly growing body of research shifts the paradigm from Euclidean grids to graph-based representations. Instead of an image, the sensor array is modeled as a graph $G = (V, E)$, where:
- Nodes ($V$) represent the individual sensors or anatomical regions of interest.
- Edges ($E$) define the relationships between them.
Crucially, these edges are not constrained by a flat grid. They can be constructed using physical 3D distances, anatomical wiring, or statistical correlations (such as Pearson correlation or phase-locking values) calculated directly from the signal timelines.
By structuring data this way, models can perform geometric deep learning. As explained by Fotis P. Kalaganis et al. in A Data Augmentation Scheme for Geometric Deep Learning in Personalized Brain-Computer Interfaces, graph-based architectures allow operations to adapt directly to the irregular, non-grid patterns of the sensors, preserving both spatial and functional topology.
Euclidean Approach (CNN) Non-Euclidean Approach (GNN)
[O]—[O]—[O] [Node A] —— (Functional Link) —— [Node F]
| | | / \ /
[O]—[O]—[O] [Node B] [Node C] ——————————————— [Node E]
| | | \ /
[O]—[O]—[O] [Node D]
(Assumes rigid, flat grid; (Preserves true 3D topology and
distorts irregular spacing) dynamic, long-distance correlations)
Why Graph Convolutions Succeed
The core reason graph neural networks (GNNs) outperform traditional architectures on sensor data lies in how they redefine the convolution operation.
In a standard CNN, a convolutional kernel computes a localized, weighted average of a pixel's spatial neighbors. In a Graph Convolutional Network (GCN), the convolution is mathematically redefined using the graph Laplacian, enabling an operation known as message passing. During a forward pass, each node updates its internal state by aggregating features from its connected neighbors, regardless of where those neighbors sit in physical space.
This mechanism allows the network to dynamically learn the "wiring" of the system. For example, if a specific cognitive task causes the frontal and occipital lobes to synchronize, a graph-based model can establish an edge between those distant nodes. The message-passing mechanism will then directly fuse their features, capturing a macro-level network phenomenon that a localized 2D convolutional kernel would entirely miss.
Beyond Graphs: Manifolds and Riemannian Geometry
While graphs excel at capturing topological connections, some facets of this research direction take geometric deep learning a step further by modeling the data on continuous, curved mathematical surfaces known as manifolds.
When analyzing multi-channel signals, researchers frequently calculate covariance matrices to capture how different sensors vary together over time. These matrices naturally reside on a Symmetric Positive Definite (SPD) manifold, which is inherently non-Euclidean.
Reinmar J Kobler et al. demonstrate in SPD domain-specific batch normalization to crack interpretable unsupervised domain adaptation in EEG that processing these signals directly on their natural geometric surface—rather than flattening them into Euclidean space—drastically improves the model's ability to generalize across different human subjects.
Similarly, Zirui Wang et al. leverage this concept in EEG-ReMinD: Enhancing Neurodegenerative EEG Decoding through Self-Supervised State Reconstruction-Primed Riemannian Dynamics, showing that respecting Riemannian dynamics allows deep learning models to capture non-linear functional connectivity with significantly higher data efficiency than traditional supervised methods. By aligning the mathematics of the neural network with the non-Euclidean geometry of the physical and functional sensors, these models extract cleaner, more robust representations of complex systems.
Key Foundational Work
Spatial and anatomical regularization of SVM for brain image analysis (2010) — contributed 1 concept elements to this direction.
A Nonparametric Bayesian Approach to Detecting Spatial Activation Patterns in fMRI Data (2006) — contributed 1 concept elements to this direction.