Symmetric Positive Semi-definite Riemmanian Geometry with Application to Domain Adaptation

Or Yair, Almog Lahav, Ronen TalmonarXiv 2020 · arXiv

Abstract

This paper introduces a Riemannian geometric framework for Symmetric Positive Semi-Definite (SPSD) matrices of fixed rank, addressing the mathematical challenges inherent in high-dimensional, low-rank data. By developing approximate geometric maps and a novel canonical representation, the authors propose an unsupervised domain adaptation algorithm. This method aligns dataset means via parallel transport, demonstrating significant performance gains in hyperspectral imaging and motion recognition tasks.

Introduction

In fields such as computer vision and medical imaging, data is frequently represented as Symmetric Positive Definite (SPD) matrices, such as covariance matrices. While the Riemannian geometry of SPD matrices is well-established, these methods typically require matrices to be full-rank. However, real-world high-dimensional data—including gene expressions and hyperspectral images—is often intrinsically low-rank, resulting in Symmetric Positive Semi-Definite (SPSD) matrices.

Standard SPD geometry fails in these cases because SPSD matrices reside on a different manifold where operations like geodesics, parallel transport, and logarithmic/exponential maps lack closed-form solutions. This paper addresses two primary hurdles: the mathematical intractability of SPSD geometric operations and the application of these tools to domain adaptation (DA) to correct for nuisance differences between datasets.

Geometric Framework and Canonical Representation

The authors leverage a "structure space" representation where an SPSD matrix $C$ is decomposed into $C = GPG^T$. In this formulation, $G$ represents a point on a Grassmann manifold (capturing the subspace), and $P$ is an SPD matrix (capturing internal scaling).

Because this decomposition is not unique, the authors introduce a method to "align" a set of SPSD matrices to a common reference point, typically the Riemannian mean. This creates a unique, canonical representation that facilitates joint analysis and allows data to be projected into a "flat" tangent space where standard machine learning tools can operate.

Illustration of the alignment process to achieve a canonical representation of SPSD matrices.

Key Geometric Innovations:

  • Geometric Approximations: The authors derive approximate Logarithmic and Exponential maps, as well as Parallel Transport (PT) mechanisms, to overcome the lack of closed-form solutions on the SPSD manifold.
  • PT-Driven Domain Adaptation: The proposed DA algorithm calculates the Riemannian mean of the source and target datasets. It then uses Parallel Transport to "slide" the source data cloud along the manifold until its mean aligns with the target mean. This rigid movement preserves the internal structure of the data while aligning global statistics.

Experimental Evaluation

The researchers evaluated their approach on two distinct datasets characterized by high-dimensional, low-rank structures:

  1. Greding Dataset (Hyperspectral): This dataset involves images of the same scene taken at different times with varying lighting and acquisition parameters. The goal was to classify land cover using a classifier trained on a different temporal domain.
  2. HDM05 Dataset (Motion Capture): This dataset contains 3D marker data of actors performing various motions. The challenge is actor-independence—training on one individual and recognizing the same motion performed by another.

Key Results and Insights

The results demonstrate that geometric alignment is highly effective for overcoming sensor and subject variability.

  • Hyperspectral Imaging: Using the proposed DA, classification accuracy (measured by Cohen’s kappa) increased from 0.575 to 0.957, outperforming several state-of-the-art methods tailored for hyperspectral data.
  • Motion Recognition: The algorithm improved cross-actor recognition accuracy from 75.91% to 96.73%. Qualitative t-SNE visualizations confirmed that after adaptation, motions from different actors clustered tightly together.

Insights:

  • Low-Rank as a Feature: Rather than treating low-rank matrices as "broken" full-rank matrices (e.g., by adding noise), this framework treats rank as a fundamental geometric property, preserving the true signal.
  • Unsupervised Simplicity: The DA algorithm is entirely unsupervised, requiring only the estimation of geometric means rather than labeled data for alignment.

Limitations and Future Work

While effective, the current method primarily aligns first-order statistics (means). Future work may explore "second-moment alignment" for domains with similar means but different variances. Additionally, the theory currently assumes a fixed rank $r$ for all matrices in a set. The authors also highlight the potential for applying this framework to Graph Laplacians, which are inherently fixed-rank SPSD matrices, opening new avenues for social network analysis and manifold learning.

Summary updated

Abstract

In this paper, we present new results on the Riemannian geometry of symmetric positive semi-definite (SPSD) matrices. First, based on an existing approximation of the geodesic path, we introduce approximations of the logarithmic and exponential maps. Second, we present a closed-form expression for Parallel Transport (PT). Third, we derive a canonical representation for a set of SPSD matrices. Based on these results, we propose an algorithm for Domain Adaptation (DA) and demonstrate its performance in two applications: fusion of hyper-spectral images and motion identification.
BibTeX
@misc{yair2020symmetric,
  author = {Or Yair and Almog Lahav and Ronen Talmon},
  title = {{Symmetric Positive Semi-definite Riemmanian Geometry with Application to Domain Adaptation}},
  year = {2020},
  eprint = {2007.14272},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.14272},
}
Figure