A research area and set of techniques that generalize deep learning methods to non-Euclidean domains (graphs, meshes, manifolds, point clouds) by building architectures and layers that respect underlying geometric and topological structure, such as permutation invariance, equivariance to group actions, and locality defined by the geometry.

Principle

Impose geometric priors (invariance, equivariance, locality, message passing consistent with adjacency or geodesic neighborhoods) to design architectures whose operations commute with relevant symmetries and thus generalize across different structures and discretizations.

Demonstration

Graph convolutional networks (GCNs) that aggregate neighbor features respecting permutation invariance for node classification; spectral or spatial convolutions on meshes that use geodesic kernels; graph attention networks that weight edges according to learned geometric relations.

Misapplication

Applying standard Euclidean CNNs directly to graph-structured data without enforcing permutation equivariance, leading to models that are sensitive to node ordering; or using flat approximations of curved manifolds for convolution, ignoring curvature and producing artifacts.

Consequence

Models that respect geometry often show better sample efficiency, transfer across different discretizations, robustness to deformations, and theoretical guarantees tied to invariance/equivariance properties, enabling principled handling of non-Euclidean data.

Reversal

Classical deep learning assumes grid-structured Euclidean input and translational invariance; without geometric adaptation, it fails to exploit structure of graphs/manifolds and generalizes poorly on non-Euclidean tasks.

Boundary

Concerns models and algorithms for structured non-Euclidean data but does not exclusively require deep neural networks—kernel methods and message-passing frameworks also fall under the umbrella; computational and scalability limits and discretization choices constrain applicability.

Semantic Tension

Tension between handcrafted geometric priors and purely learned representations: strong priors improve generalization and interpretability but may limit expressivity; tension also with purely topological methods (TDA) which focus on global shape rather than local equivariant operations.

Synthesis

Geometric deep learning unites geometric and topological principles with learning architectures to construct models that respect symmetries and locality of non‑Euclidean domains, yielding better generalization and robustness for structured data.