Correntropy-Based Improper Likelihood Model for Robust Electrophysiological Source Imaging.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Signal Processing

Background:

• Bayesian learning is a unified framework for electrophysiological source imaging.
• Current methods assume Gaussian noise, which is often inaccurate due to artifacts in brain activity measurements.
• This assumption leads to suboptimal performance in real-world scenarios.

Purpose of the Study:

• To develop a novel, robust likelihood model for Bayesian source imaging that accounts for non-Gaussian noise.
• To improve the accuracy of electrophysiological source reconstruction in the presence of artifacts.
• To provide a more reliable tool for analyzing noisy brain signals.

Main Methods:

• Proposed a new improper distribution model for noise, inspired by the maximum correntropy criterion.
• Developed a robust likelihood function using this new noise model.
• Integrated the robust likelihood with hierarchical priors and employed variational inference for source activity estimation.
• Utilized score matching to determine hyperparameters for the improper likelihood model.

Main Results:

• Simulations with known ground-truth demonstrated more precise source reconstruction compared to the conventional Gaussian model.
• Evaluation on a real-world dataset for a visual perception task confirmed the superiority of the proposed method.
• The new approach effectively handles non-Gaussian noise, outperforming standard techniques.

Conclusions:

• The proposed robust likelihood model offers a significant advancement for Bayesian source imaging.
• This method provides a more accurate and reliable approach for analyzing electrophysiological data affected by artifacts.
• The study establishes a new foundation for applying Bayesian source imaging to real-world noisy brain signals.