Building a Silicon Brain

prostoalex tips us to an article in MIT's Technology Review on a Stanford scientist's plan to replicate the processes inside the human brain with silicon. Quoting: "Kwabena Boahen, a neuroengineer at Stanford University, is planning the most ambitious neuromorphic project to date: creating a silicon model of the cortex. The first-generation design will be composed of a circuit board with 16 chips, each containing a 256-by-256 array of silicon neurons. Groups of neurons can be set to have different electrical properties, mimicking different types of cells in the cortex. Engineers can also program specific connections between the cells to model the architecture in different parts of the cortex."

obligatory

[intthis (525681)]

that's great, but will it run linux?

so...

[President_Camacho (1063384)]

prostoalex tips us to an article in MIT's Technology Review on a Stanford scientist's plan to replicate the processes inside the human brain with silicon.

So how long until we get AI that's addicted to World of Warcraft?

2^^20 neurons? That's wayyyy too many

[Eternal Vigilance (573501)]

...to accurately model most American thought processes.

Gotta go - American Idol's back on.


Dave, my mind is going. I can feel it...

Go to Hollywood

[syousef (465911)]

Lots of silion in Hollywood....oh you said BRAINS not BREASTS.

Hardly something new...

[by Anonymous Coward]

This is hardly something new. Intel had a chip a number of years ago, called ETANN that was a pure-analog neural network implementation. Another cool aspect of this chip was that the weight values were stored in EEPROM-like cells (but analog) so the training of the chip would not be erased if it lost power.

But the whole technology of neural networks almost pre-dates the Von Neumann architecture. Early analog neural networks were constructed in the late 40's.

Not only are these simulations nothing new but they are in every-day products. One of the most common examples is the misfire detection mechanism in Ford vehicle engine controllers. Misfire detection in spark ignition engines is based on so many variables that neural networks often perform better than hard-coded logic (although not always, just like the wetware counterparts, they can be "temperamental").

There are several other real-world neural network applications (autofocusing of cameras for example).

Ahh the hidden magic of embedded systems...

What'll be new?

[wanax (46819)]

I have to wonder what the purpose is.. You can model simplified 'point' neurons, and various aggregates that can be drawn from them (eg, McLoughlin's PDEs)... or you can run a simplified temporal dynamic (eg. Grossberg's 3D LAMINART), and easily include 200k+ neurons in the model easily to capture a broad range of function. For those would like running more detailed models of individual neuronal dynamics, you have Markram's project simulating a cortical column with compartmental models, or what Izhikevich is doing with delayed dynamic models.

Although this setup may be able to run ~1mil neurons, in total, it would seem that with 16 chips of 256x256 each, the level of interaction would be limited, and the article has no indication that these are the more complicated (and realistic) compartmental models of neurons that can sustain realistic individual neuronal dynamics (and for example Izhikevich, Markram and McLoughlin have spent a lot of time trying to simplify), or whether this is just running point style neurons a bit faster than is traditional.. and I have to wonder here, whether if these chips can't do compartmental models, why not just run this on a GPU?

I checked out this guy's webpage, and he seems smart.. but this project is years away from contributing.. I wonder, especially with the Poggio paper yesterday, when the best work being done just at MIT in Neuro/AI right now is probably in the Torralba lab, whether slashdot editors may want to find some people to vet the science submissions just a tad.

Naturally Intelligent Systems

[TheCouchPotatoFamine (628797)]

For those interested in this field, may i suggest a book, Naturally Intelligent Systems? It's slightly older, but it explains a wide gamut of neural networks without a single equation, and manages to be funny and engaging at the same time. it is one of the three books that changed my life (by it's content and ideas alone - i'm not otherwise into AI). highly recommended: Naturally Intelligent Systems on amazon [amazon.com]

Why?

[Quiet_Desperation (858215)]

Look at the rubbish the human brain generates. Ideology. Irrationality. Depression. Religion. Politics. Reality TV.

You really want processors that need weekly visits from an Eliza program and iZoloft patches?

"Sorry, Bob. I can't run those projections now. The supercomputing cluster is in a funk over the American Idol results."

Y'all think AI is going to be so great and a bag of chips, too.

brain does not use math logic,but pattern matching

[master_p (608214)]

The human brain generates so much rubbish because it does not use mathematical logic, but pattern matching.

In many cases, mathematical logic can not be used to prove the absolute truth of a proposition; therefore the brain uses pattern matching to 'prove' the 'truth' of a proposition to the degree that is useful for the survival of the entity that carries it.

Take, for example, the proposition that 'prime numbers are infinite'. We all think they are infinite, but there is no mathematical proof for it yet. Wh

More than just modeling the brain

[AndOne (815855)]

Having been a fan of neuromorphic engineering for several years now(Note I'm not an active researcher but I pretend somedays :) ) one of the major advantages of neuromorphic functionality isn't necessarily it's ability to model biological systems but the fact that the devices are extremely low power. When modeling neurons in silicon(at least back in the day of Carver Mead's work and for cochlea and retina stuff and I'm doubting it's changed too bunch but I could be wrong) the transistors would run in sub threshhold mode(basically leakage currents so OFF) since the power curves modeled the expected neuro response curves. One of Boahen's stated goals(at least on his website when he was at Penn) was to reduce power consumption and improve processing power for problem solving via these techniques. His lab has been in Scientific America a couple times in the last few years for work in accurately modeling Neuronal spiking in hardware too. I have them but not at hand so I can't cite them at the moment but they were fun reads.

So in summary, it's more than just modeling the brain. It's about letting biology inspire us to make better and more efficient computing systems.

Re:One Million Neurons ;)

[tehdaemon (753808)]

As far as I know, brains do not use back-propagation at all. Each neuron changes it's own weights based on things like timing of inputs vs output, and various neurotransmitters present.

If all you want are more neural nets like we have been doing then sure - back-propagation algorithms matter. That does not seem to be the goal here though.

T

Re:

[tehdaemon (753808)]

back-propagation is a mathematical simplification of neurotransmitters.

No. Correct me if I am wrong, but back-propagation works by comparing the output of the whole net to the desired output, and tweeking the weights one layer at a time back up the net. In real brains, neurotransmitters either do not travel up the chain more than one neuron, or they simply signal all neurons physically close, whether they are connected by synapses or not. (like a hormone) Further, since real brains are recurrent networks (

Re:Depends on What Consciousness Is

[QuantumG (50515)]

I, like many other engineers, don't give a shit. We just want to solve problems to which there are no simple solutions and "AI" offers some approaches that work.

Leave the philosophy till after we have the science.

Re:Depends on What Consciousness Is

[Venik (915777)]

How can you build a software model of a process you don't understand? The best hope is to build a hardware approximation of a human brain and hope that, somehow, the same processes start occurring, quantum or otherwise. And if that doesn't work, then you'll have to do some real science.

Re:Depends on What Consciousness Is

[kestasjk (933987)]

Read "How Brains Think" by William H. Calvin; he's a neurologist and the book goes into lots of detail about how brains think (dur), how they evolved, and the possibility of AI.
He's an expert in the field and you can feel his bitter dislike of "quantum consciousness" proponents through his writing. He writes that it's just saying "we don't know how X works, and we don't know how Y works, but if we say that Y depends upon X then we have one problem instead of two".

Consciousness is built on the interactions of neurons. We understand how neurons work at interact at a low level (from studying the ~50 neuron brains of snails etc), and we understand on a large level which regions of the brain do what, but we don't understand the "middle ground".

It's as if we understand the transistor, and logic gates, and we can recognize which part of a chip is the ALU and which is the cache, but we can't recognize an adder circuit or microinstruction translator for what it is.

Quantum physics is certainly involved in the action of transistors but it doesn't explain how they combine to process data.

(On a similar note some I saw, in a documentary, one crackpot explain away "spontaneous human combustion" with an unknown quantum particle.)

The reverse seems more interesting.

[Kadin2048 (468275)]

One thing you don't hear much about, is what progress, if any, is being made in interfacing electronic systems into biologic ones, and growing biologic circuits. Perhaps our understanding of biological computation and storage simply isn't complete enough to make such a system practical, even if we were able to somehow interface a clump of neurons to the outside world electronically, but it certainly seems like the data storage capacity of biologic systems is far greater (per mass/volume) than anything devised artificially. Although, I suppose it's impossible to equate, since it's not clear how 'compressed' information is, when it's encoded by the mammalian brain as memories.

Re:

[QuantumG (50515)]

I thought Interface [wikipedia.org] was a remotely interesting read.. at least the technological aspects.. the commentary on media dominated elections was just depressing. They extract some neural tissue from a subject, grow a bunch of neurons, interface them to chip with a wireless transmitter, then reinsert them into the brain. Then, with some training, the chip can replace functions of the brain destroyed by stroke or cancer or whatever. The data dump of the communication between the neurons and the chip is the reall

Re:

[NixieBunny (859050)]

An interesting aspect of the brain is that it may be possible to build circuitry that mimics its behavior without understanding that behavior. There are many complex systems (collections of simple parts) that exhibit surprisingly coherent behavior that you just wouldn't expect. Swarms of locusts are one example. The insect robots that learn how to walk every time you turn on their power is another example.

Re:Only one mibiNeuron?

[tehdaemon (753808)]

Are you sure about that? FTA:

"We can currently do small simulations of hundreds to thousands of neurons, but it would be great to be able to scale that up," he says.

A 2.0GHz dual-core CPU running 2^20 neurons in the net at 100Hz gets about 40 clock cycles per neuron per cycle...Somebody check my math please.

T

BOINC

[Gary W. Longsine (124661)]

I would think a BOINC project might produce enough muscle to get a really big brain going. Imagine a BOINC [berkeley.edu] cluster of...

;-)

Re:

[wall0159 (881759)]

The calculations involve adjusting the weight of connections between neurons, which generally scale exponentially with the number of neurons. This is because each neuron typically has connections to many other neurons.

So, your math might be right, but your assumptions are wrong. :-)

Re:Only one mibiNeuron?

[naoursla (99850)]

Now add a bunch of connections between all of those neurons. As you approach fully connecting the network, the time complexity to compute one time-step approaches O(N^2) where N is the number of neurons.

2^20 * 2^20 == 2^40. Ignore memory cache constraints for a moment and say each update takes 1 clock cycle. Since we are dual core we can get 2 updates per cycle. Each clock cycles takes 500pS. 2^40*500ps/2 means each complete brain update takes 274s on your computer.

Re:Hello, world?

[by Anonymous Coward]

A soul...

Re:

[Patrik_AKA_RedX (624423)]

Same thing that's already in a human brain at birth:
void SuckAtNipple();
void CryForAttention();
void Shit();

Re:Hello, world?

[Alsee (515537)]

void SuckAtNipple();
void CryForAttention();
void Shit();

I think Shit() has a return type...

-