From local to supercomputational spiking models of resting-state dynamics in cerebral cortex

Abstract

Cortical resting-state dynamics is organized on multiple spatiotemporal scales and involves cell-type-specific spike rates, slow and fast fluctuations, clustered inter-area correlations, and inter-area activity propagation mainly in the feedback direction. Simulations of large parts of cortex resolving the individual neurons and synapses enable studying how cortical network structure shapes this multi-scale activity, but have been limited by the available computational resources and simulation technology. I will describe developments in the simulation technology of NEST which have improved the build and run times of large-scale spiking neural network simulations. Using these advances, we simulated a network of 32 vision-related areas of macaque cortex with each area represented by a 1 mm² microcircuit with the full density of neurons and synapses on the supercomputers JUQUEEN and JURECA. In recent work, we also compared the computational performance of NEST on a high-performance compute cluster with that of the neuromorphic hardware SpiNNaker for such a 1 mm² microcircuit. The supercomputer simulations of the multi-area network rely on a recently derived connectivity map for the visual areas of macaque cortex that predicts the connection probability between any two neurons based on their types, areas, and layers. This connectivity map integrates axonal tracing data with predictions from cortical architecture (neuron densities, layer thicknesses), inter-area distances, and neuronal morphologies. In line with models using simplified equations for the individual areas, our model predicts that cortex operates in a metastable state where slow activity fluctuations appear. In this regime, the power spectrum of simulated V1 spiking activity and the distribution of spike rates across V1 neurons agree well with those from parallel spike recordings in lightly anesthetized macaque. Furthermore, the inter-area functional connectivity is similar to that from macaque resting-state fMRI. The simulated neuronal activity propagates across areas mainly in the feedback direction, akin to LFP findings during sleep. A mean-field-based analysis shows that the areas activated first are those with the largest contributions to the most unstable eigenvector of the effective connectivity. Our model reconciles microscopic and macroscopic accounts of cortical neural networks and provides a platform for further developments. To promote reproducibility and enable adaptations and extensions by others, we have made the full model code available at github.com/INM-6/multi-area-model .

Speaker’s bio

Sacha van Albada leads the Theoretical Neuroanatomy group at the Institute of Neuroscience and Medicine (INM) at Research Center Jülich in Germany. Her group studies the architecture and connectivity of brain circuits as the basis for neural network models at the resolution of neurons and synapses that relate structure to dynamics. She leads a task on large-scale spiking network models of cerebral cortex in the Human Brain Project, and in this same project, she is implementation lead for a so-called Co-Design Project (cutting across subprojects) on visuomotor integration. She obtained a Bachelor of Science degree from University College Utrecht, followed by a Master's in Theoretical Physics from Utrecht University, and went on to do a PhD at the School of Physics at the University of Sydney, working with Peter Robinson on mean-field brain models. Before joining her current institute, the INM-6, in 2011, she did a two-year postdoc on tinnitus with Peter Tass at the INM-7. She has her own group since March 2017.