Most deep learning architectures are built on the assumption of a "flat" world. Operations like addition, matrix multiplication, and convolution implicitly treat data as points in Euclidean space ($\mathbb{R}^n$). However, many real-world signals—such as the orientation of a robot arm, the covariance of brainwaves, or the hierarchical structure of a social network—live on curved surfaces known as Riemannian manifolds.
When we force these non-Euclidean structures into standard neural networks, we risk distorting their intrinsic geometry. Tangent space mapping provides a mathematically rigorous bridge between these worlds. By projecting curved data into a flat, local Euclidean space, researchers can apply the full power of modern deep learning while respecting the underlying geometry of the data.
The Geometry of the Tangent Space
To understand why this works, consider the surface of the Earth. While the planet is curved, a small enough patch of it looks flat. In differential geometry, this flat patch is the tangent space. For every point on a manifold, there exists a unique tangent space that serves as a local linear approximation.
The primary mechanism for moving between these spaces involves two operations:
- The Logarithmic Map ($\log$): This projects a point from the curved manifold into the flat tangent space.
- The Exponential Map ($\exp$): This projects a vector from the tangent space back onto the manifold.
As Huang et al. demonstrate in their work on LieNet, this transformation allows a network to ingest skeletal motion data (which lives on a Lie group) and map it into a Euclidean space where standard fully connected layers and softmax classifiers can operate. Without this mapping, the network would struggle to maintain the physical constraints of human joints.
Bridging Manifolds and Deep Learning
The utility of tangent space projection spans several distinct types of non-Euclidean data:
- Symmetric Positive Definite (SPD) Matrices: Covariance matrices are widely used to describe image textures or EEG signals. However, they form a curved cone-shaped manifold where standard subtraction can lead to "invalid" matrices. Van Gool and Huang introduced SPDNet, which uses a matrix logarithm layer to linearize these matrices, enabling end-to-end learning for emotion and action recognition.
- Hyperbolic Space: For data with hierarchical or tree-like structures, hyperbolic geometry is more efficient than Euclidean space. Ganea et al. developed Hyperbolic Neural Networks (HNNs) by performing linear transformations in the tangent space. They map points to a local Euclidean plane, apply matrix-vector multiplication, and project the result back to the hyperbolic manifold. This same logic is applied to graphs by Liu et al. in Hyperbolic Graph Neural Networks (HGNNs).
- Lie Groups and Robotics: In robotics, orientations are often represented as rotation matrices. Schuck et al. show that wrapping reinforcement learning inputs in Lie algebra (the tangent space of rotations) leads to more stable training and mathematically consistent policies compared to using raw Euler angles or quaternions.
The Log-Euclidean Framework
A particularly powerful version of this idea is the Log-Euclidean framework. By mapping manifold-structured data into a logarithmic domain, the complex Riemannian distance between two points simplifies into a standard Euclidean distance.
As noted by the authors of Kernel Methods on the Riemannian Manifold of Symmetric Positive Definite Matrices, this transformation allows researchers to use classical algorithms like Support Vector Machines (SVMs) or Principal Component Analysis (PCA) on complex descriptors. In medical imaging, this framework has been used to average cardiac diffusion tensor data, as explored in Physics (2016), and to register brain images by simplifying the computation of similarity measures, as shown by Yeo et al..
Learnable Projections
While early methods used a fixed point (like the identity matrix) as the center of the projection, recent work has moved toward making the projection itself learnable. Souza et al. proposed a Log Layer for Grassmann manifolds where the tangency point is a trainable parameter. By optimizing this point via Riemannian Stochastic Gradient Descent, the network can learn the most representative "flat" view of the underlying subspaces.
Similarly, Liu et al. generalized the Log-Euclidean mapping into a Bag-of-Matrix-Summarization (BoMS) framework. Instead of projecting relative to a single point, the model learns a dictionary of reference points, allowing for more flexible transformations that adapt to the specific geometry of the input data.
Why It Works: Intuition
The reason tangent space mapping is so effective is that it preserves intrinsic dynamics while enabling linear analysis.
In Yi et al., high-dimensional curves on manifolds are simplified through a process called "development," which uses parallel transport to flatten the curve into Euclidean space without losing its shape. By operating in the tangent space, we aren't just approximating the data; we are viewing it through a lens that makes its natural relationships compatible with the linear operations that define modern AI. This bridge allows us to maintain the geometric integrity of the physical world while leveraging the computational efficiency of the digital one.