Jacobian Granger causality for count and binary data with applications to causal network inference.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Granger causality is vital for neural network inference.
• Current artificial neural network formulations for Granger causality excel with continuous data but struggle with discrete, sparse neural data.
• Limitations exist in applying existing methods to count and binary neural activity.

Purpose of the Study:

• To adapt Jacobian Granger causality for discrete neural data types (count and binary).
• To address limitations of continuous-data optimized formulations for sparse, discrete neural systems.
• To evaluate the performance of the extended method against existing approaches.

Main Methods:

• Extended Jacobian Granger causality using specialized loss functions for count and binary data.
• Utilized simulated datasets to compare the novel approach with a competing method.
• Applied the enhanced Granger causality method to real neural spiking data from monkey visual cortex.

Main Results:

• The adapted Jacobian Granger causality method demonstrates effectiveness with discrete neural data.
• Simulations confirmed the method's performance against a competing approach.
• Analysis of monkey visual cortex data revealed structured neural activity under natural movie stimuli, including increased positive self-connections in neurons.

Conclusions:

• The extended Jacobian Granger causality provides a robust framework for inferring neural networks from discrete and sparse neural data.
• Natural movie stimuli induce more structured neural activity compared to white noise.
• Positive self-connections in neurons, potentially encoding salient visual information, are more prevalent during naturalistic visual processing.