Nonstationary source separation using sequential and variational Bayesian learning

Area of Science:

• Signal Processing
• Machine Learning
• Statistical Modeling

Background:

• Independent Component Analysis (ICA) assumes stationary signals and fixed mixing processes.
• Real-world applications often involve nonstationary source signals and time-varying mixing systems.
• Existing ICA methods struggle with dynamic signal environments.

Purpose of the Study:

• To develop an online learning algorithm for Gaussian Process (GP) based nonstationary Independent Component Analysis (ICA).
• To enable blind source separation in scenarios with time-varying mixing coefficients and source signals.
• To accurately capture and adapt to evolving statistical properties of signals.

Main Methods:

• Utilized online Bayesian learning to capture sequential signal statistics.
• Employed Gaussian Processes (GP) to model temporal structures of nonstationary mixing coefficients and sources.
• Developed variational Bayesian inference for parameter estimation and latent source activity characterization.
• Incorporated automatic relevance determination to track source activity dynamically.

Main Results:

• The proposed algorithm effectively handles nonstationary and temporally correlated mixing coefficients and source signals.
• Demonstrated superior performance compared to other ICA methods in experiments involving audio signal separation under nonstationary conditions.
• Showcased the ability to reflect and propagate source activity across frames.

Conclusions:

• The developed online GP-based ICA algorithm provides a robust solution for blind source separation in dynamic environments.
• This approach accurately models and adapts to time-varying characteristics of signals and mixing systems.
• The method offers significant improvements over existing techniques for real-world nonstationary signal separation problems.