New geometric framework improves signal processing models
Using geometry-driven models can lead to better results in tasks like speaker separation and neuroimaging.
Researchers developed a geometry driven framework for directional statistics in signal processing. This framework connects signal representations to their natural manifolds or quotient spaces.
Many signal processing objects cannot be modeled in ordinary Euclidean space. This is because they have periodicity, normalization, sign or global-phase ambiguity, basis invariance, or positive-semidefinite constraints.
The work reviews circular, toroidal, spherical, projective, Stiefel, Grassmann, and covariance-valued distributions. It also describes modeling and estimation strategies like analytical and gradient-based maximum likelihood.
Other strategies include structured low-rank parameterizations, mixtures and hidden Markov models (HMMs), score-based and contrastive estimation, and geometry-aware K-means. These methods consider statistical assumptions and computational requirements.
The study compares toroidal phase vectors to projective, Grassmann, and positive-semidefinite representations in functional neuroimaging. It also tests speaker separation in a car environment.
Directional representations outperform a standard Euclidean model in the car environment. Improved models may occur when geometric representation, statistical distribution, inference method, and computational regime are considered together.
Key facts
- Directional representations outperform standard Euclidean models for speaker separation in a car environment.
- The framework connects signal representations to natural manifolds or quotient spaces.
- The research studies multivariate phase coherence in synthetic data and functional neuroimaging.