Beyond GLMs: a generative mixture modeling approach to neural system identification.

Area of Science:

• Computational neuroscience
• Statistical modeling

Background:

• Generalized linear models (GLMs) are widely used for neural spike response characterization.
• GLMs have strong assumptions, limiting the discovery of complex stimulus-response relationships.
• Alternative methods struggle with high-dimensional data.

Purpose of the Study:

• To develop a novel approach that bridges the gap between GLMs and assumption-light methods.
• To create a model that can capture complex dependencies in high-dimensional stimulus spaces.
• To improve the characterization of neural spike responses.

Main Methods:

• Extended linear and quadratic generalized linear models (GLMs) using Gaussian mixtures.
• Derived the model from a generative perspective for interpretable components.
• Addressed the non-concave log-likelihood challenge in mixture models.

Main Results:

• The Gaussian mixture-based model captures complex stimulus-response dependencies.
• It requires fewer parameters than histogram-based methods for high-dimensional data.
• The proposed model outperformed standard GLMs in practice.

Conclusions:

• Gaussian mixture models offer a flexible and powerful alternative for neural spike response analysis.
• This approach effectively models complex relationships in high-dimensional neural data.
• The method provides interpretable components while maintaining computational feasibility.