Evaluating the significance of spike correlations in the neural code by means of analytically solvable models and of Information Theoretical Analysis

F. MONTANI, R SENATORE, S PANZERI

Punta del Este, Uruguay

Conferencia; XVI Conference on Nonequilibrium Statistical Mechanics and Nonlinear Physics; 2008

An important question in mathematical neuroscience is whether the activity of a network of cortical neurons can be described by pairwise interaction, much like the Ising model. An appropriate theoretical framework to study complex firing pattern of activity of a network of many simultaneously recorded neurons is Information Theory. To address the importance of the spike correlations we developed novel estimators, using maximum-entropy response models, to lower and upper bounds to the information loss by a decoder. These methods allowed us to estimate if different order of correlations were sufficient to decode the Information encoded in a given experimental data-set. Pairwise models were sufficient in this data-set to decode almost all the information.  To evaluate further the significance of higher order correlations, we considered the case of a widespread distribution of neuronal activity in which correlations cannot be reduced to pairwise correlations. This widespread distribution was theoretically conceived under the information geometry framework to generate higher-order stochastic interactions. We analytically estimate this widespread distribution in the "finite response model regime" taking into account higher order correlations across the neuronal pool.  This approach allowed us to estimate Fisher analytically information and to characterize higher order correlations with a deformation parameter.