Independent component analysis based on marginal density estimation using weighted Parzen windows.

Area of Science:

• Signal Processing
• Machine Learning
• Statistical Inference

Background:

• Independent Component Analysis (ICA) is crucial for blind source separation.
• Existing ICA algorithms face challenges with high-dimensional data and accurate density estimation.
• Effective demixing requires robust modeling of source signal distributions.

Purpose of the Study:

• To propose a novel ICA algorithm leveraging marginal density estimation.
• To develop an effective demixing matrix and weighted Parzen window (WPW) representations.
• To achieve a factorial joint density for high-dimensional observations.

Main Methods:

• The algorithm minimizes Kullback-Leibler (KL) divergence using Potts encoding for an approximating form.
• Tractable optimization of WPW parameters and the demixing matrix.
• Utilizes annealed expectation-maximization and natural gradient descent for minimization.

Main Results:

• Numerical simulations validate the algorithm's accuracy and reliability.
• Performance is demonstrated on linear mixtures and real-world signals (fetal ECG, facial images, ERPs).
• Outperforms several popular ICA algorithms in comparative tests.

Conclusions:

• The proposed ICA algorithm offers an accurate and reliable approach to signal separation.
• Marginal density estimation provides an effective framework for ICA.
• The method shows promise for diverse applications in signal processing and data analysis.