Parameter estimation in a model for multidimensional recording of neuronal data: a Gibbsian approximation approach

Abstract.

This article proposes improved numerical procedures for estimating parameters in a spatiotemporal lattice model introduced for the analysis of cortical activities monitored from arrays of diodes. The numerical algorithms are based on approximations inspired by statistical physics. Both Gibbsian and mean-field approximations are used; they allow for computing local conditional probabilities inside the lattice. The statistical procedures rely on the computation of pseudomaximum-likelihood estimators. The estimators are evaluated on the basis of Monte Carlo simulations. These simulations show that mean-field approximations are useful for reducing the variance of estimators when the data are recorded from arrays of 144 diodes (which are in accordance with standard practice). In light of these improved methods, we give new interpretations for a data set obtained from optical recording of a Guinea pig's auditory cortex in response to pure tone stimulations.