Modeling the Brain with NCS and Brainlab
The numarray extension package for Python provides for efficient manipulation and statistical analysis of the large NCS datasets that result from a simulation. For graphs and charts of results, the excellent matplotlib package produces publication quality output through a simple yet powerful MATLAB-like interface (Figure 1). Brainlab also provides a number of convenient interfaces for these packages, making it easier to do the operations commonly needed for neuroscience research. Brainlab also provides interactive examination of 3-D views of the network models using the Python OpenGL binding (Figure 2).
Quite often, some experimentation with a number of network parameters is required in order to find a balanced brain model. For example, if a synaptic strength is too high or too low, the model may not function realistically. We have seen how Brainlab could help a modeler do a search for a good model by repeatedly running the same model with a varying parameter. But an even more powerful technique than that simple search is to use another inspiration from biology, evolution, to do a genetic search on the values of whole set of parameters. I have used Brainlab to do this sort of multiparameter search with a genetic algorithm (GA) module of my own design and also with the standard GA module of the Scientific Python package, SciPy.
Brainlab has made my complex experiments practical, perhaps even possible. At this point I can't imagine doing them any other way. In fact, if NCS were to be reimplemented from scratch, I would suggest a significant design change: the elimination of the intermediate NCS input text file format. This file format is just complex enough to require a parser and the associated implementation complexity, documentation burden and slowdown in the loading of brain models. At the same time, it is not nearly expressive enough to be usable directly for any but the simplest brain models. Instead, a scripting environment such as Python/Brainlab could be integrated directly into NCS, and the scripts could create structures in memory that are accessed directly from the NCS simulation engine. The resulting system would be extremely powerful and efficient, and the overall documentation burden would be reduced. This general approach should be applicable to many different problems in other areas of model building research.
This summer, NCS is going to be installed on a new 4,000-processor IBM BlueGene cluster at our sister lab, the Laboratory of Neural Microcircuitry of the Brain Mind Institute at the EPFL in Switzerland, in collaboration with lab director Henry Markram. Early tests show that we can achieve a nearly linear speedup in NCS performance with increasing cluster size, due to efficient programming and the highly parallel nature of synaptic connections in the brain. We hope that other researchers around the world will find NCS and Brainlab useful in the effort to model and understand the human brain.
Resources for this article: /article/8203.
Rich Drewes (email@example.com) is a PhD candidate in Biomedical Engineering at the University of Nevada, Reno.
- Machine Learning Everywhere
- Own Your DNS Data
- Bash Shell Script: Building a Better March Madness Bracket
- Understanding OpenStack's Success
- Simple Server Hardening
- Understanding Firewalld in Multi-Zone Configurations
- From vs. to + for Microsoft and Linux
- Natalie Rusk's Scratch Coding Cards (No Starch Press)
- The Weather Outside Is Frightful (Or Is It?)
- Ensono M.O.