Construction and analysis of a spatially organized cortical network model

The cerebral cortex is one of the most intricate natural systems known, due to the multitude and heterogeneity of interconnected cells and its operation on different temporal and spatial scales. Cortical activity on the mesoscopic scale, spanning square millimeters to centimeters of cortical surface area, can be recorded with multi-electrode arrays implanted in neural tissue. Such extracellular recordings provide simultaneous access to population signals like local field potentials (LFPs) as well as spiking activity of individual neurons, and expose spatiotemporal activity patterns emerging parallel to the cortical surface. Local neuronal connectivity is specific with respect to cortical layers and neuron types, and the probability that two neighboring neurons are connected decays with distance. Computational models of neuronal networks with corresponding spatial extents and signal predictions are needed to infer the relationship between connectivity structure and experimentally recorded activity. This thesis focuses on the development of mesoscopic spatially organized cortical network models with cellular resolution. We develop a multi-layer network model with realistic neuron density and distance-dependent connectivity covering 4x4 mm², a similar area as covered by multi-electrode arrays in use today. The model comprises excitatory and inhibitory spiking neuron populations in four cortical layers, integrates experimentally obtained connectivity data, and reproduces features of observed in-vivo spiking statistics. As a finding, the model reconciles the seemingly contradictory experimental observations of weakly correlated spike trains and strong, distance-dependent correlations of LFPs. Experimental data on the structure and dynamics of cortical networks are only known within certain margins of error and severe simplifications need to be made. Therefore, mean-field theory is required to explore regimes of biologically realistic activity and uncover mechanisms governing the network dynamics. This thesis advances the theory of spatially organized networks to a point where predictions are in quantitative agreement with direct simulations of spiking neuronal networks. Since conventional high-performance computing architectures are not optimized for accelerated and massively parallel neuroscientific simulations, the community develops dedicated neuromorphic hardware. We compare the performance of the software simulator NEST to the neuromorphic hardware system SpiNNaker in terms of accuracy, runtime, and energy consumption. To capture spatiotemporal patterns in simulated activity data, we design concepts for visual data analysis and provide the interactive web-based tool VIOLA (VIsualization Of Layer Activity) as a reference implementation. Moreover, we assess the integration of collaborative and interdisciplinary simulation-analysis workflows into online platforms. This thesis discusses the foundations of a model platform for the stepwise refinement of mesoscopic spatially structured network models and paves the way towards tackling further questions on the brain’s function, learning, and diseases