« Google and AI | Main | Cognitive Computing 2007 »

Towards real-time, mouse-scale cortical simulations

Last week, we presented a poster on a fascinating new result at the CoSyNe 2007 conference:

Neurobiologically realistic, large-scale cortical and sub-cortical simulations are bound to play a key role in computational neuroscience and its applications to cognitive computing. One hemisphere of the mouse cortex has roughly 8,000,000 neurons and 8,000 synapses per neuron. Modeling at this scale imposes tremendous constraints on computation, communication, and memory capacity of any computing platform.

We have designed and implemented a massively parallel cortical simulator with (a) phenomenological spiking neuron models; (b) spike-timing dependent plasticity; and (c) axonal delays.  

We deployed the simulator on a 4096-processor BlueGene/L supercomputer with 256 MB per CPU. We were able to represent 8,000,000 neurons (80% excitatory) and 6,300 synapses per neuron in the 1 TB main memory of the system. Using a synthetic pattern of neuronal interconnections, at a 1 ms resolution and an average firing rate of 1 Hz, we were able to run 1s of model time in 10s of real time!

I believe that such cortical simulators are the linear accelerators of neuroscience. We are already able to study extremely large-scale cortical dynamics. This is a developing story...please stay tuned in!

Reference:

James Frye, Rajagopal Ananthanarayanan, and Dharmendra S. Modha, "Towards real-time, mouse-scale cortical simulations," CoSyNe: Computational and Systems Neuroscience, Salt Lake City, Utah, Feb 22-25, 2007 PDF

TrackBack

TrackBack URL for this entry:
http://p9.hostingprod.com/@modha.org/blog-mt/mt-tb.fcgi/32

Listed below are links to weblogs that reference Towards real-time, mouse-scale cortical simulations:

» Mouse brain simulated at 1/10 of real-time from Miron's Weblog
Dharmendra S Modha posts an article about a recent result presented at CoSyNe 2007. We deployed the simulator on a 4096-processor BlueGene/L supercomputer with 256 MB per CPU. We were able to represent 8,000,000 neurons (80% excitatory) and 6,300 syna... [Read More]

» Mouse cortex simulated at 1/10 of real-time from Miron's Weblog
[Cross posted from my blog ] Update: the BlueGene/L instance used here is only 1/32 of the size of the [Read More]

» Musimulation from Notes from Classy's Kitchen
Som rapporteret over det hele har en gruppe IBM forskere simuleret et nervesystem på størrelse med en mus i nogenlunde... [Read More]

Comments

This is great news, looks like a real step up in simulation. Question: I looked over the Izhikevich and Brette articles in the bibliography of your presentation paper, and your paper itself. Will there be the possibility to model dendritic timings and topology, or only axonal timings? The reasons for asking are that STDP in real neurons is very dependent on timing of currents within the dendritic tree, as well as modulations of these currents (spec. the backpropagated action potentials).

It is known that the precise actuation of any STDP is critically dependent on arbor shape, distribution of calcium ion channels, and the irregularity of the backpropagating trains, in addition to the more frequently modeled timing between the input signal and the backpropagating signal.

It is even possible to construct a model in which the tree itself is the storage system for a complex shaped signal, which stores by stabilizing a region on a possibly chaotic attractor.

Different STDP profiles have been cataloged in different types of cells, allowing the possibility of a range in different locations. They allow a large number of gradients between pure synchronization and pure desynchronization when interpreted as moving the excitability of the synapse, and therefore its firing timing, of which the neutral position, in which synchronization is unimportant, is represented by Hebbian learning. (this interpretation is dependent on findings that excitability is also changed with STDP, and that synaptic failure rates may be higher in vivo, due to the fact that synapses fail at higher rates when separated from glial support). Since synchronization simplifies the signal, but also reduces penetration of the backpropagating wave, and desynchronization increases the detail and complexity, and increases the penetration to more of the neuron, it is possible to make the argument that these operations create flat spots and peaks in a bifurcation surface, thereby working to stabilizing part of an attractor.

Just a beginning of an idea, but it has backing in data, and it argues hard for dendritic arbor shape and timing modeling.

Sources:

Abbott, L. F. and S. B. Nelson (2000). "Synaptic plasticity: taming the beast." Nature Neuroscience 3: 1178-1183.

Pfrieger, F. W. (2002). The role of glia in the development of synaptic contacts. The Tripartite Synapse: glia in synaptic transmission. A. Volterra, P. J. Magistretti and P. G. Haydon. New York, Oxford University Press: 24-34.

Sjöström, P. J. and M. Häusser (2006). "A Cooperative Switch Determines the Sign of Synaptic Plasticity in Distal Dendrites of Neocortical Pyramidal Neurons." Neuron 51(2): 227-238.

Williams, S. R. and G. J. Stuart (2000). "Backpropagation of Physiological Spike Trains in Neocortical Pyramidal Neurons: Implications for Temporal Coding in Dendrites." The Journal of Neuroscience 20(22): 8238-8246.

Forgive me for gushing, but I've just stumbled upon this extraordinary news, and I'm still tachycardic! This is quite the most remarkable and exciting thing I've seen for ages, and I want to thank and congratulate you most sincerely for the work.

I'll now start the - presumably long and tedious, but happy - task of trying to dig up as much of the background to this experiment as I can find. As little more than an AI and neural network groupie, it won't be easy, but I'm pretty sure I'll enjoy it all. I'm especially interested to see how much detail you will have provided about the logical architecture of the system.

From my position of dangerous ignorance I've always assumed that a cortical simulation would require, on the one hand, the sort of massive parallelism that you have succeeded in assembling; and on the other a great deal of information about how wetware neural 'machines' are constructed, in order to duplicate and simulate them.

I'm, aware that some success has been achieved with, for example, visual cortex circuitry, I was under the impression that knowledge of the overall circuitry of the brain was still very much at functional block level. I assume that you've yet to duplicate this sort of circuitry at a fine granularity, but this almost becomes a configuration issue when you've reached the stage you have.

I'm babbling. Sorry. I'll start my hunt now. Thanks very much indeed.

CD

Good morning dott. Modha,
I have read a news about IBM Research of Almanden Lab about Mouse Cortical Simulations. From my position of dangerous ignorance, I would like to know from what philosophical theory you and your staff have postulated your experiment... I'm really interested and I want to thank and congratulate you most sincerely for the work.

I'm a Phd semiotic student, from University of Turin. I'm interested about problem of consciousness and identity and I support the theory of Pierce that Subject is a sign about itself, and all of the results this old theory presuppose.

Is it possible to have some references more beyond those I can read in your reports, please?

Thank you for your job,
Daniela Ghidoli

Post a comment

(If you haven't left a comment here before, you may need to be approved by the site owner before your comment will appear. Until then, it won't appear on the entry. Thanks for waiting.)