Slashdot Mirror

IBM recently announced success in simulating 2 billion of their custom designed synaptic cores, 1 trillion synapses apparently. Here's the pdf report

I don't think that phrase invokes the same idea as most of the folks on /. The "neuromorphic" algorithms they allude to are the kind that run on highly specialized hardware (e.g., this beast). This type of hardware really just works similarly to synapses (integrate & fire architecture). Of course you could simulate the algorithm on a more conventional processor, but it would probably lose much of it's low-power attribute.

FWIW, the algorithm they propose is attempt to identify objects that project up from the ground. To do this, they attempt to label parts of the image as obstacle (or not) taking a raw initial guess and filtering it with a pre-trained neural net (using some sort of adjacent region belief propagation technique).

I think they may have "cheated" a bit in that in some papers, they describe decomposing the image with oriented Gabor filters (edge orientation detectors), but they admit that this decompsition doesn't currently work well on their ultra-low-power computing platform.

FYI: MAV=micro aerial vehicle

From TFA, it doesn't sound like they simulated the cerebral cortex of a cat. It sounds like they simulated a neural net with a comparable number of neurons. Not the same thing.

What article did you read? The one linked to in the post clearly says they simulated a portion of cat cortex and, in fact, that's largely what they did. There's more here about some of the specifics. It's not an entirely accurate simulation, but it's pretty close. Not all neuron types are represented and it's largely cortical, thalamus and reticular nucleus neurons. They've created cortical hypercolumns which is the way a real cortex is laid out. They've omitted the layer 1 neurons, but otherwise the cortex is probably pretty functional for what they're doing. I think it's a pretty amazing feat.

Actually, the simulation isn't the big deal. This is: "We have developed a new algorithm, BlueMatter, that exploits the Blue Gene supercomputing architecture to noninvasively measure and map the connections between all cortical and sub-cortical locations within the human brain using magnetic resonance diffusion weighted imaging." So they're also developing techniques to extract the wiring diagram of living brains. That's significant.

Don't read too much into the amount of supercomputer hardware required. They're running what's basically a circuit simulator, and those are inefficient but flexible. When NVidia develops a new graphics chip, they test and debug by compiling the VHDL into C, and running it, slowly, on about thirty racks of 1U servers. When that's working, the VHDL is compiled down to IC masks and the consumer part that's a few centimeters across is fabricated. That kind of shrink ratio should be expected once the R&D effort figures out what to fab.

Here's the actual paper (pdf).

Although, of course, posting the piece of pap that explains how many processors my machine has makes so much more sense.

Wasn't Slashdot supposed to be for a semi-technical audience? Hell, even a semi-literate one.

Also, from the research note, they only simulated a neuron firing rate of 1Hz. I don't know much about mouse brains, but I know that's nowhere near the firing rate found in mammal brains.

Frankly, I don't think a general purpose computer architecture like BlueGene is appropriate for this kind of research. A cluster of lower-power small nodes with small local memory and a dense interconnect would be much better for the purpose. Their simulation had 4096 nodes simulating approximately 2,000 neurons each. I think 65536 nodes with about 120 neurons on each would be more appropriate for this kind of work.

How can it be half a mouse brain if it has 1/1000 the number of a real half mouse brain? Their simulated neurons also had less synapses than the real thing. So is the 8000 a typo, or am I missing something?

It's a typo. See original research note here.

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

(see the actual research description here: http://www.modha.org/papers/rj10404.pdf)
Secondly, it is not necessary for a cortex to have left-right brain functionality in order for it to function. This has been demonstrated in live humans.
And third, the speed, relative to real-time, is irrelevant. It is comparatively a minor task to increase the speed of the simulation by increasing parallelization.
Now, to respond to your somewhat antiquated understanding of the current state of AI:

In addition, everything I have seen in tech press on AI since the rules based AI reasoning failures of the 80's has been neural net simulations looking for patterns, such as the mentioned synchronized firings.

Sounds like you're a couple of years behind (as would be expected on slashdot, which primarily focuses on IT and science, and not neuroscience). Let me bring you up to date a little. Spiking neural networks began to grow in popularity in the mid to late 90's. They are much more biologically realistic then most of the models used in the 80's and early 90's. Also, a lot of research has been done which points to the significance of chaotic attractors, which arising from phase-locked loops in the neuronal structure. The fact that synchronous firing is observed tends to imply similar dynamics are occurring.
Furthermore, you make the assumption that biological brains are somehow superior to simulated brains, just because they are more chemically complex. That assumption has absolutely no research to back it up. For all we know at this point all of that chemical complexity may be superfluous for evolutionary benefits (and this is direction which evidence suggests).

Aren't the neural net rules just tweaked until they get interesting behavior like that?

That's the way it used to be done, so I can understand your confusion here. I think the problem lies in the fact that people are very interested in neuroscience these days. But a remarkable amount of progress has been made. Phenomenological spiking neural networks are quite a bit biologically accurate than the "neural nets" of the 80's and early 90's.

Don't tell me you think they actually have any idea how they would simulate brain functionality.

The cortex is arranged into mircocolumns of neurons, which have a very definite structure repeating structure over the surface of the cortex. Jeff Hawkins has recently presented a very convincing argument for structure of the mind, in relation to the structure of the cortex.

Training neural nets is just something easy to do. Beats actually writing complex code, doesn't it?

If you're implying that the simulation was not complex, consider that each neuron had its own dedicated computer. And, once again, this is much more complicated than a simple neural network.

I've never seen any explanation for how either short term or long term memory works, much less reasoning or any other functionality. And that at least is something that would seem able to be modeled and explained. How does man know anything about something they have never encountered before, for example, to acquire language as a child?

Explanations for both short and long term memory have been out there for quite some time. But neuroscience is not a popular topic of discussion, partly because it can get quite complex. People would much rather be talking about the step in the evolution of Intel processors, or life