[ExI] A small step towards brain emulation

Anders Sandberg anders at aleph.se
Tue Nov 20 11:37:39 UTC 2012


Dharmendra Modha's IBM team have announced a rather impressive brain 
simulation.
http://www.modha.org/blog/SC12/RJ10502.pdf

The good news is that they have simulated a system with 53*10^10 neurons 
(near cortex size) and 1.37*10^14 synapses at a mere 1542 slowdown, 
using an overall network structure borrowed from the macaque brain. It 
also scales well with increasing number of processors. The bad news is 
that the model is a fairly abstract model; not quite chunking all 
neurons in a minicolumn into one, but still using abstracted 
integrate-and-fire neurons. And of course, it does not have any fine 
structure based on any real brain.

The real paper is here, with more technical details: 
http://www.modha.org/blog/SC12/SC2012_Compass.pdf (this was run just on 
the Blue Gene, the press stuff is a later run on Sequoia)

The cores are described here: http://www.modha.org/papers/IJCNN%202012.pdf

OK, we seem to be on track when it comes to computing power for big 
neural networks, and the connectomics people are charging ahead (but 
they need to get onto an exponential track of scan-to-neuron conversion: 
right now algorithms are too slow even for the small scan volumes we 
got). It feels like the ball is now in the computational neuroscientist 
court: we better start working on how to turn imaging data into 
something that can run on something like the Compass.

-- 
Anders Sandberg,
Future of Humanity Institute
Oxford Martin School
Faculty of Philosophy
Oxford University




More information about the extropy-chat mailing list