New PMNLV model reveals neural co-variability hierarchy in mouse visual cortex

A team from Johns Hopkins and other institutions has developed a novel statistical framework for analyzing neural population activity. The Poisson matrix-normal latent variable (PMNLV) model addresses a key limitation of existing overdispersion models: they treat each neuron's gain variability as independent, missing the structured co-variability that arises from network interactions. PMNLV places a matrix-normal prior over latent gains with a Kronecker-factored covariance, enabling it to capture both inter-neuron and temporal covariance without structural assumptions. The authors provide two complementary estimation algorithms—a variational EM (VEM) approach and a Kernel Tournament Method (KTM) for data-driven kernel selection—and validate them on simulated data with accurate recovery of tuning curves and covariance factors.

Applying PMNLV to real Neuropixel recordings across four cortical regions of the mouse visual hierarchy (V1, LM, AL, and higher areas), the team replicated a known finding that single-neuron marginal variability changes little across areas. More importantly, they discovered that shared population co-variability—measured by the model's covariance structure—peaks in primary visual cortex (V1) and progressively declines in higher visual areas. This trend is invisible to traditional scalar summaries like the Fano factor, which only capture per-neuron variability. The finding suggests that early visual cortex maintains higher coordinated gain modulation, potentially as a mechanism for encoding uncertainty or attention, while higher areas rely on more decorrelated activity for specialized processing. The PMNLV framework is broadly applicable to any simultaneously recorded neural population where structured gain covariance is of interest, offering neuroscientists a powerful new tool to probe network-level dynamics.