Variational mixture of Bayesian independent component analyzers.

Area of Science:

• Data Science
• Machine Learning
• Signal Processing

Background:

• Subspace data modeling is increasingly important, with methods like Principal Component Analysis (PCA) and Independent Component Analysis (ICA) widely used.
• Existing mixture models for PCA and factor analysis often assume Gaussian features, limiting their effectiveness on non-Gaussian or discontinuous data.
• Sophisticated analyses are needed as applications and computing power expand.

Purpose of the Study:

• To extend Gaussian mixture models to an Independent Component Analyzers (ICA) mixture model.
• To develop a novel framework for modeling non-Gaussian and discontinuous data manifolds.
• To automatically determine local dimensionality and optimize the number of ICA components.

Main Methods:

• Proposed an Independent Component Analyzers (ICA) mixture model.
• Utilized variational Bayesian inference and structure determination techniques.
• Developed automatic local dimensionality determination and optimized ICA component selection.

Main Results:

• Successfully modeled non-Gaussian and discontinuous subspace manifolds.
• Demonstrated the framework's effectiveness on complex synthetic data.
• Applied the method to decompose functional magnetic resonance imaging (fMRI) data into medically useful features.

Conclusions:

• The proposed ICA mixture model offers a powerful approach for analyzing complex data with non-Gaussian features.
• This framework enables more meaningful representations and applications in diverse fields, including medical imaging.
• The method provides automatic structure determination for robust subspace modeling.