In traditional machine learning, we represent an object—a face, a document, or a signal—as a single point in a high-dimensional Euclidean space. However, real-world data is rarely static. A single face appears differently under various lighting conditions; a video of a moving object contains a sequence of correlated frames; and a brain scan captures a complex, deformable geometry.
When we collapse these rich collections of data into a single average vector, we lose the internal structure and variance that define them. As Harandi et al. argue, representing image sets or videos as single vectors fails to capture the underlying structural relationships. Instead, a more powerful approach is to model these collections as linear subspaces. By treating a set of observations as the "span" of its variations, we can represent data as points on a Grassmann manifold.
The Geometry of Subspaces
A Grassmann manifold $\mathcal{G}(k, n)$ is the set of all $k$-dimensional linear subspaces within an $n$-dimensional vector space. While a vector is a point in flat space, a subspace is a point on this curved, non-Euclidean manifold.
This shift in perspective allows us to handle data variability naturally. For instance, Farnell et al. explain that modeling a class as a subspace captures inherent fluctuations, such as lighting changes in hyperspectral imagery, by treating the signature as a collection of vectors rather than a single fixed point. Similarly, Luo demonstrates that motion trajectories in video can be effectively segmented by treating them as points on a smooth manifold, allowing for the separation of complex, overlapping motions.
Measuring Similarity on the Manifold
To perform classification or clustering, we need a way to measure the distance between two subspaces. Because the Grassmann manifold is curved, standard Euclidean distances are insufficient. Researchers have developed several geometric metrics to address this:
- Projection Distance: Wang et al. represent subspaces as projection matrices, allowing the distance between them to be calculated via the Frobenius norm. This serves as a computationally efficient proxy for the true geodesic distance along the manifold's surface.
- Principal Angles: Jayasumana et al. show that valid kernels for algorithms like SVMs can be constructed using projection distances derived from principal angles—the angles between the closest vectors in two subspaces.
- Chordal and Geodesic Distances: For tasks like ordinal embedding, Hou et al. utilize chordal Grassmannian distance to measure semantic separation, while Gatto et al. use weighted geodesic distances to capture the natural geometric structure of multi-dimensional tensor data.
Averaging and Prototyping
In Euclidean space, the "average" of a set of points is their arithmetic mean. On a manifold, the concept of an average is more complex. Xi et al. propose using the Karcher Mean—the geometric center of gravity on the manifold—to serve as a class prototype for image-set recognition. This captures the geometric relationships of the subspaces more accurately than linear averaging.
This geometric perspective also revitalizes classical algorithms. Feragen et al. and Chakraborty reformulate Principal Component Analysis (PCA) as an optimization problem on the Grassmann manifold. By computing "Grassmann Averages," they create robust, scalable versions of PCA that can process massive datasets that would otherwise cause traditional matrix-based methods to fail.
Bridging Domains and Modalities
One of the most compelling reasons to use Grassmannian representations is their ability to unify disparate types of data. Because a subspace is defined by its dimensionality rather than its raw coordinates, it can act as a common language between different domains.
Gopalan demonstrates this by using parallel transport—a way of moving vectors along a manifold without changing their geometric properties—to bridge heterogeneous domains. By mapping local data clusters from different sources (like high-res images and text) to subspaces on a Grassmann manifold, the author builds "geometric bridges" that allow models trained on one domain to generalize to another.
Why It Works
The effectiveness of the Grassmannian approach stems from its ability to preserve invariance. Whether it is Bakry using manifold kernels to remain invariant to pixel-level variance in lipreading, or Dong et al. finding representative subspaces across multi-layer graphs, the core mechanism remains the same: by treating data as a subspace, we focus on the structure of the information rather than the noise of individual observations. As Luo et al. conclude, representing data as linear subspaces captures internal variance more effectively than single-vector representations, leading to more robust classification in complex, real-world environments.
Key Foundational Work
Boosting Constrained Mutual Subspace Method for Robust Image-Set Based Object Recognition (2009) — contributed 2 concept elements to this direction.
Subspace-based feature representation and learning for language recognition (2012) — contributed 1 concept elements to this direction.
Spectral Curvature Clustering (SCC) (2009) — contributed 1 concept elements to this direction.