Why are we doing this?

Neural network is a large and complex system. Being able to simualte large-scale neural network is essential for us to understand how brain works. With the emergence of open source big data platform such as Hadoop, we are able to achieve this with a low cost.

How are we doing this?

We are running simulations on a Hadoop cluster deployed at the University of St. Thomas, by Department of Graduate Programs in Software.

Basically, from computer science perspective, a neural network is a graph structure, so we are focusing on both neural modeling as well as distributed graph algorithms. We have been working on graph algorithms using MapReduce and Giraph. Finally we settled with Giraph because it is a effective model for large-scale iterative graph processing.

Basic work flow

alt basic work flow

phase one: generating neural network

The neural network to be processed is configurable by a set of XML templates. These templates then go through a Python parser which will create the configurations in XML format and feed it to a MapReduce job to generate the neural network in parallel.

phase two: modeling with Giraph

The neural network generated in phase one is then passed to a Giraph job for simulation.

phase three: post analysis

The results output by phase two are then used for post analysis, such as firing rate computation and firing plots.

That’s it ^_^