#40 · ML / AI methods

Explicit latent and geometric models of shared variability in neural populations

The shared thread is the explicit modeling of co-variability in neural populations. The multi-scale information geometry preprint derives a Riemannian metric from coarse-graining principles, extending Fisher information and linking that metric to population mutual information. The PMNLV preprint models neural co-variability with a Kronecker-factored latent gain covariance, estimated on Neuropixel data, with the effect peaking in V1. The third preprint extends the theme to selectivity: a closed-loop framework using digital twins of macaque V1 and V4 together with generative models produces verifiable language descriptions of neuron tuning. What is new across the set is the move toward explicit, testable parameterizations of variability and selectivity rather than summary statistics of correlation.