April 30, 2007
Mouse-sized neural network created:
Despite what the headlines might say, no-one has simulated a mouse brain. What has been created is still quite impressive though.
Scientists from IBM have created an artificial neural network which contains the simulated equivalent of the number of neurons in an actual mouse cortex, but with less synapses.
The mouse cortex contains about 8 million neurons, each of which has an average of 8,000 synaptic connections.
The simulation used the same number of 'neurons', but used an average of only 6,300 synaptic connections per brain cell, and each neuron fires about ten times slower than in real life.
Crucially, the simulated neurons are only vague approximations of the actual thing.
This is no reflection on the researchers, but really a result of the fact that we just don't know enough about how single neurons work to create truly accurate simulations.
Also, the model was made up of simulated neurons of one particular type only to make things a little more straightforward.
Finally, there was no attempt to recreate the 'architecture' of the mouse cortex - that is, the division of the model into sections which do different functions, and no attempt to account for the function of non-neuronal brain cells.
The sheer scale of the model is impressive though, and shows that these large scale models are becoming technically feasible.
Previously, the technical restrictions of dealing with the computations and moving the data about quickly enough had not been overcome for a simulation of this size.
The project was run on a BlueGene/L supercomputer to make it possible.
IBM have released a short technical report on the project which is available at the link below.
pdf of report 'Towards Real-Time Mouse Scale Cortical Simulations'.
Link to Wikipedia page on artificial neural networks.
