Toward a spiking Multi-area network model of macaque visual cortex

The investigation of large neural network models incorporating several spatial and/or temporal scales has been limited to date by the available computational resources. The availability of JUQUEEN and of the K supercomputer in Kobe, Japan, and recent progress in the simulation technology of NEST [1] have partly lifted this barrier. The multi-scale neural network model of the macaque visual cortex presented here will make use of these facilities to derive aspects of cortical dynamics that were hitherto inaccessible to computational exploration. The model extends a recent microcircuit model of primary visual cortex [2] to 32 visual cortical areas of the macaque, taking into account their layered structure. By expanding the areas to realistic sizes we will account for a large majority of the synapses onto each neuron, in contrast to the local microcircuit model which only accounts for about half of the synapses.We compile the connectivity map from multiple datasets including data from electro-physiological measurements [3], anatomical studies collected in CoCoMac [4] and quantitative tract tracing [5]. These extensive but incomplete data are complemented with estimates based on empirical connectivity rules, such as a decrease of connection probability with distance, and a dependence of layer specificity on the different architectures of the areas. In modeling the single-neuron dynamics, only excitatory and inhibitory neuron types are distinguished, and their morphology is disregarded. These simplifications enable us to study the influence of connectivity itself on cortical dynamics, independent of detailed single-neuron properties. The ability of the structure to account for known aspects of cortical dynamics will be tested using layer- and area-specific firing rates, regularity and synchrony of spiking, and frequency spectra.Since cortical function only arises through the interaction of multiple areas, the extension to multi-area networks is an important step toward modeling functional circuits. The model also has a prominent integrative component, bringing together data from different sources into a unified framework. We present the workflow from the data to the consistent connectivity scheme in which future experimental data can be incrementally incorporated without changes to the framework.In order to study the influence of structural properties of the cortex, we start with a down-scaled version of the model, where the areas are described by a microcircuit representing the neurons under 1 mm2 of cortical surface. We show that this reduced model can be simulated on a modern high-performance cluster and present preliminary results on the network activity