Probabilistic learning vector quantization on manifold of symmetric positive definite matrices

Area of Science:

• Machine Learning
• Riemannian Geometry
• Data Science

Background:

• Data represented by symmetric positive definite matrices reside on Riemannian manifolds.
• Non-Euclidean geometry of these manifolds challenges traditional Euclidean machine learning algorithms.
• Existing methods often fail to capture the intrinsic geometry of manifold-valued data.

Purpose of the Study:

• To develop a novel classification method for manifold-valued data.
• To generalize probabilistic learning vector quantization for data on the manifold of symmetric positive definite matrices.
• To address the limitations of Euclidean approaches in non-Euclidean spaces.

Main Methods:

• Generalization of probabilistic learning vector quantization (PLVQ).
• Development of a probabilistic learning Riemannian space quantization algorithm.
• Utilizing the Riemannian natural metric (affine-invariant metric) and induced Riemannian distance.
• Employing Riemannian gradient descent for learning rule derivation.

Main Results:

• The proposed method demonstrates superior performance compared to traditional algorithms.
• Empirical validation on synthetic, image, and motor imagery electroencephalogram (EEG) data.
• Effective classification of manifold-valued data by respecting its non-Euclidean geometry.

Conclusions:

• The probabilistic learning Riemannian space quantization algorithm offers a powerful new tool for classifying manifold-valued data.
• This approach effectively handles the geometric complexities of symmetric positive definite matrices.
• The method shows significant promise for applications in image analysis and biomedical signal processing.