Partitioning Neural Co-Variability.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Neural Data Analysis

Background:

• Trial-to-trial variability in neural responses is crucial for understanding neural computation and population dynamics.
• Existing overdispersion models often assume independent neuron gain, failing to capture essential network statistics.
• Network-level gain covariance in neural populations remains largely unstudied due to modeling limitations.

Purpose of the Study:

• To develop a novel statistical model capable of capturing structured spiking gain modulation across neural populations.
• To introduce the Poisson matrix-normal latent variable (PMNLV) model for analyzing population-level neural variability.
• To investigate how population co-variability changes across different cortical areas in the mouse visual hierarchy.

Main Methods:

• Developed the Poisson matrix-normal latent variable (PMNLV) model with a matrix-normal prior on latent gain and Kronecker-factored covariance.
• Implemented two estimation algorithms: variational EM (VEM) and Kernel Tournament Method (KTM).
• Applied the VEM algorithm to Neuropixel recordings from mouse visual cortex.

Main Results:

• The PMNLV model successfully recovers inter-neuron and temporal covariance factors and accurate tuning curves on simulated data.
• Analysis of mouse visual cortex data replicated findings of stable single-neuron variability across areas.
• Demonstrated that shared population co-variability peaks in primary visual cortex and declines in higher visual areas.

Conclusions:

• The PMNLV framework effectively models population-level neural variability, including structured gain covariance.
• Shared population co-variability is a dynamic neural network property that varies across cortical hierarchy.
• This approach is broadly applicable to analyzing structured gain covariance in any simultaneously recorded neural population.