AMICA-Python: Adaptive Mixture Independent Component Analysis with Anderson Acceleration

Adaptive Mixture Independent Component Analysis (AMICA) is a gold-standard algorithm for blind source separation in EEG research, prized for its ability to learn flexible, non-Gaussian source distributions. However, its historical reliance on a legacy Fortran implementation—accessible primarily through MATLAB—has created a significant barrier for researchers working within the modern scientific Python ecosystem. This paper introduces AMICA-Python, an open-source implementation designed to bridge this gap.

The authors developed AMICA-Python to be fully compatible with the scikit-learn API, allowing it to integrate seamlessly into existing Python-based data pipelines. The implementation closely mirrors the original Fortran algorithm while leveraging modern numerical libraries (NumPy, SciPy) to optimize performance. A key technical contribution is the integration of Damped Anderson Acceleration (DAAREM), an optimization technique that speeds up the convergence of fixed-point iterations like AMICA by using historical information from previous steps to extrapolate the next update.

AMICA-Python demonstrates high numerical fidelity, showing a negligible median relative absolute difference of approximately $1.07\times10^{-8}$ compared to the Fortran reference when tested on 14 open EEG datasets. Beyond accuracy, the implementation is computationally efficient: the standard Python version is 17.7% faster than the Fortran reference, and the version utilizing Anderson acceleration is 34.1% faster. These results confirm that AMICA-Python is a robust, faster, and more accessible alternative for researchers who require the high-quality source separation provided by AMICA without the constraints of the MATLAB environment.

By providing a performant, Python-native version of AMICA, this work lowers the barrier to entry for advanced blind source separation techniques. It enables researchers to incorporate AMICA into modern machine learning workflows, facilitates easier algorithm modification, and improves computational efficiency for large-scale EEG analysis.