IBM’s Almaden Research Cen­ter announced in Novem­ber that it had pro­duced a “cor­ti­cal sim­u­la­tion” of the scale and com­plex­ity of a cat brain. This sim­u­la­tion ran on one of IBM’s Blue Gene super­com­put­ers, in this case at the Lawrence Liv­er­more National Lab­o­ra­tory (LLNL):

Sci­en­tists, at IBM Research — Almaden, in col­lab­o­ra­tion with col­leagues from Lawrence Berke­ley National Lab, have per­formed the first near real-time cor­ti­cal sim­u­la­tion of the brain that exceeds the scale of a cat cor­tex and con­tains 1 bil­lion spik­ing neu­rons and 10 tril­lion indi­vid­ual learn­ing synapses.

The fig­ure presents the results of our weak scal­ing study, where the prob­lem size is increased with increas­ing amount of mem­ory. The plot demon­strates nearly per­fect weak scal­ing in terms of mem­ory, since twice the model size could be sim­u­lated when the amount of mem­ory is dou­bled. The largest sim­u­lated model cor­re­sponds to a scale larger than the cat cere­bral cor­tex, reach­ing 4.5% of the human cere­bral cor­tex.

Scal­ing of Cor­ti­cal Sim­u­la­tions. Copy­right © 2009, ACM, Inc.

Using a state-of-the-art Blue Gene/P with 147,456 proces­sors and 144 TB of main mem­ory, authors were able to sim­u­late a thal­a­m­o­cor­ti­cal model at an unprece­dented scale of 10⁹ neu­rons and 10¹³ synapses. Com­pared to the human cor­tex, this sim­u­la­tion has a scale that is roughly 1–2 orders smaller and has a speed that is 2–3 orders slower than real-time. This work opens the doors for bottom-up, actual-scale mod­els of the thal­a­m­o­cor­ti­cal sys­tem derived from biologically-measured data. More­over, fur­ther progress in super­com­put­ing, real­time human-scale sim­u­la­tions are not only within reach, but indeed appear inevitable:

Growth of Top500 super­com­put­ers over­laid with obtained result and a pro­jec­tion for real­time human-scale cor­ti­cal sim­u­la­tion. Copy­right © 2009, ACM, Inc.

Finally, this isn’t a sim­u­la­tion of a cat brain, it’s a sim­u­la­tion of a brain struc­ture that has the scale and con­nec­tion com­plex­ity of a cat brain. It doesn’t include the actual struc­tures of a cat brain, nor its actual con­nec­tions; the var­i­ous exper­i­ments in the project filled the mem­ory of the cor­ti­cal sim­u­la­tion with a bunch of data, and let the sys­tem cre­ate its own sig­nals and con­nec­tions. Put sim­ply, it’s not an arti­fi­cial intel­li­gence, it’s a plat­form upon which an AI could con­ceiv­ably be built. The same team cre­ated a mouse-scale brain sim­u­la­tion a few years ago. One of the researchers, Dhar­men­dra Modha, runs a blog on cog­ni­tive com­put­ing, and has posted a PDF of the research paper on this project.

R. Anan­tha­narayanan, S. K. Esser, H. D. Simon, & D. S. Modha (2009). The Cat is Out of the Bag: Cor­ti­cal Sim­u­la­tions with 109 Neu­rons, 1013 Synapses Pro­ceed­ings of the Con­fer­ence on High Per­for­mance Com­put­ing Net­work­ing, Stor­age and Analy­sis , 1–12 DOI: 10.1145÷1654059.1654124

2nd December, 2009 1 Comment