Brain-scale simulations with NEST: supercomputers as data integration facilities

Abstract

Event Abstract Back to Event Brain-scale simulations with NEST: supercomputers as data integration facilities Tobias C. Potjans1, 2*, Susanne Kunkel3, 4, Abigail Morrison3, 4, 5, Hans E. Plesser5, 6, Rolf Kötter1, 7 and Markus Diesmann2, 5 1 Research Center Juelich, Institute of Neuroscience and Medicine, Germany 2 RIKEN Computational Science Research Program, Brain and Neural Systems Team, Japan 3 Bernstein Center Freiburg, Germany 4 Faculty of Biology, Albert-Ludwigs-University, Functional Neural Circuits, Germany 5 RIKEN Brain Science Institute, Japan 6 Norwegian University of Life Sciences, Department of Mathematical Sciences and Technology, Norway 7 Radboud University Medical Centre, Donders Institute for Brain, Cognition and Behaviour, NeuroPi.org, Netherlands Neural network simulations are currently making a qualitative leap, leaving the regime of highly specialized “single-scale” models that incorporate only limited aspects of biological data, but focus on the multi-scale nature of the brain. These developments are fueled by the advent of computing power at the peta-scale which is increasingly becoming available to computational (neuro-)scientists all over the world (see e.g. http://www.prace-project.eu/, http://www.nsc.riken.jp/, http://www.ncsa.illinois.edu/BlueWaters/). In order to optimally employ this computing power in the context of neural network modeling, we identify two main requirements: (1) Suitable simulation technology has to be made available that can efficiently represent multi-scale brain networks on supercomputers and rapidly solve the network's activity dynamics. (2) A new class of models needs to be developed comprising multiple scales from the local microcircuit to the macroscopic brain network consistently with available data. In the past years, simulation technology development focused on the representation of the local cortical network consisting of approximately 100,000 neurons and 1 billion synapses on standard HPC clusters (http://www.nest-initiative.org). The construction of models resolving the layer- and type-specific connectivity structure of the cortical microcircuit integrated a large body of experimental data ranging from anatomical and electrophysiological studies to photostimulation and electron microscopy [1]. The comparison of the simulated network activity and experimentally observed in vivo cell-type specific activity reveals the consistency as well as potential shortcomings of the available data and models. We show that this class of models successfully captures prominent microscopic activity features such as layer-specific spike rates and the interplay of excitation and inhibition in the propagation of transient feed-forward inputs as observed in vivo (e.g. [2]). In order to address the activity dynamics and the function of the local network in the context of the embedding in a network of multiple cortical areas, next generation multi-scale neural network simulations have to simultaneously represent the local microcircuit and the macroscopic connectivity structure (e.g. [3]). These brain-scale network simulations approach a regime where the number of cores is larger than the number of synapses per neuron. Therefore, corresponding data structures for the network representation have to make use of the sparseness of connections but nevertheless allow rapid network construction and spike delivery to target neurons. Models on this scale drastically increase their self-consistency and explanatory power because they explicitly incorporate most of the long-range inputs to neurons that were previously modeled as abstract external inputs but make up around 50% of all synaptic inputs. The advanced data integration now not only combines multiple methods but also multiple scales, linking microscopic and macroscopic connectivity. Here, we present the scale-up of the NEST simulation tool (http://www.nest-initiative.org) up to tens of thousands of processors on the JUGENE supercomputer (http://www.fz-juelich.de/jsc/jugene) and quantify time-memory trade-offs. Furthermore, we summarize our efforts in the construction of brain-scale network models that integrate a vast amount of data on multiple scales. In this concept supercomputers are utilized as data integration facilities. Partially supported by the Next-Generation Supercomputer Project of MEXT, Japan, the Helmholtz Alliance on Systems Biology, EU Grant 15879 (FACETS), McDonnell Foundation Collaboration Grant, Research Council of Norway Grant 178892/V30 (eNeuro) and JUGENE Grant JINB33.