An efficient, parallel event driven simulator of realistic neuronal networks

Download: File.

“An efficient, parallel event driven simulator of realistic neuronal networks” by Y. Dong, S. Mihalas, and E. Niebur. In Annual Meeting of the Society for Neuroscience, 2008.

Abstract

Among the plethora of neuronal network simulators available to the neuroscience community (Brette et al., J Comput Neurosci, 2007), some (e.g. NEURON) emphasize the realism of the underlying single-unit model, while others use very simple models for the sake of efficiency of network simulations. Our goal is to create a simulator of an intermediate level of biological realism that is efficient, platform-independent, and endowed with a friendly user interface. It is based on an event-driven algorithm (Mattia & Giudice, Neural Comput, 12(10), 2000) for point neurons, which computes spike times approximation-free (i.e., with machine precision). For neuronal model, we used a versatile and efficient generalization of the integrate-and-fire model (Mihalas and Niebur, this volume). Other, custom-written, neuronal models can be integrated as well. Multiple neuronal models can be simultaneously used in the same network. The simulator is written in JAVA, uses the Java Parallel Virtual Machine (JPVM) (Adam Ferrari, Concurrency: Practice and Experience, 10(11-13),1998), and runs in heterogeneous systems of arbitrary size and machine architecture without recompilation. Resources are distributed efficiently over all available machines, with simulations of smaller neuronal networks progressing with different parameter sets (e.g. different experimental settings) in parallel on separate machines, while large neuronal networks are distributed over the machines, with sub-networks communicating by event-based methods. Network description and experimental parameter specification are written in the XML format. Tools for efficient generation of layered networks as well as extensive Java-based visualization tools are provided. In an example, we simulated the primate early visual system using natural images as inputs and including area V1. The model of 1,361,250 neurons and 24,030,000 synapses of realistic firing rate was distributed over eight AMD64 machines, 1GB memory each. The simulation takes 1m54s per second of simulated time, plus 54s setup-time to establish the connectivity. Disclosures: Y. Dong , None; S. Mihalas, None; E. Niebur, None. Support: 5R01EY016281-02 [Authors]. [Abstract Title]. Program No. XXX.XX. 2008 Neuroscience Meeting Planner. Washington, DC: Society for Neuroscience, 2008. Online.

Download: File.

BibTeX entry:

@proceedings{Dong_etal08b,
   author = {Y. Dong and S. Mihalas and E. Niebur},
   title = {An efficient, parallel event driven simulator of realistic
	neuronal networks},
   booktitle = {Annual Meeting of the Society for Neuroscience},
   volume = {Abstract 798.1},
   address = {Washington DE},
   year = {2008},
   organization = {Society for Neuroscience}
}