Related Experiment Videos
Independent component analysis using an extended infomax algorithm for mixed subgaussian and supergaussian sources
T W Lee1, M Girolami, T J Sejnowski
1Computational Neurobiology Lab, The Salk Institute, 10010 North Torrey Pines Road, La Jolla, CA 92038, USA. tewon@salk.edu.
Neural Computation
|February 9, 1999
Summary
This study extends the infomax algorithm to blindly separate mixed signals, even with complex source distributions. The enhanced method effectively isolates artifacts from brain signals in electroencephalographic recordings.
Area of Science:
- Signal Processing
- Machine Learning
- Computational Neuroscience
Background:
- Independent Component Analysis (ICA) is crucial for blind signal separation.
- Traditional ICA methods struggle with non-Gaussian source distributions.
- The infomax algorithm provides a framework for ICA.
Purpose of the Study:
- To extend the infomax algorithm for improved blind signal separation.
- To handle mixed signals with sub- and super-Gaussian source distributions.
- To apply the enhanced algorithm to real-world electroencephalographic (EEG) data.
Main Methods:
- Utilized a learning rule based on negentropy as a projection pursuit index.
- Employed parameterized probability distributions to derive a general learning rule.
- Incorporated natural gradient optimization and stability analysis for regime switching.
- Extended the Bell and Sejnowski (1995) infomax architecture.
Main Results:
- Successfully separated 20 sources with diverse distributions.
- Demonstrated effective separation of artifacts (eye blinks, line noise) from EEG signals.
- Showcased the algorithm's ability to handle high-dimensional neurophysiological data.
Conclusions:
- The extended infomax algorithm offers robust blind source separation for complex distributions.
- This method is effective for artifact removal in EEG analysis.
- The approach enhances the utility of infomax for practical signal processing applications.