Scalable Gamma-Driven Multilayer Network for Brain Workload Detection Through Functional Near-Infrared Spectroscopy.

Area of Science:

• Computational neuroscience
• Machine learning
• Signal processing

Background:

• Functional near-infrared spectroscopy (fNIRS) is a key neuroimaging technique for monitoring brain activity.
• Accurate analysis of fNIRS data requires sophisticated computational models to interpret complex neural signals.
• Non-negative matrix factorization (NMF) methods are widely used for dimensionality reduction and feature extraction in signal processing.

Purpose of the Study:

• To propose a novel scalable gamma non-negative matrix network (SGNMN) for enhanced brain fatigue detection.
• To investigate the relationship between network architecture (width and depth) and detection accuracy.
• To establish a method for learning SGNMN parameters through up-down sampling layers.

Main Methods:

• Utilizing Poisson randomized Gamma factor analysis to initialize the first layer neurons.
• Employing Gamma distribution for shape parameters inferring subsequent layer neurons and weights.
• Applying Dirichlet distribution for upsampling connection weights and Gamma distribution for downsampling hidden units.
• Implementing an iterative up-down sampling process across network layers to learn parameters.

Main Results:

• Demonstrated that the width and depth of the SGNMN are significantly interrelated.
• Showcased that optimizing network width, depth, and parameters leads to accurate brain fatigue detection.
• Validated the efficacy of the proposed SGNMN model using functional near-infrared spectroscopy data.

Conclusions:

• The scalable gamma non-negative matrix network (SGNMN) offers a robust framework for brain fatigue detection.
• Network architecture design, specifically width and depth, is critical for optimizing performance in fNIRS analysis.
• The proposed parameter learning method effectively optimizes the SGNMN for neuroimaging applications.