One Genius or Two Normals

Let’s say you have a puzzle. Who would you rather hire to solve it: two strangers from the street-corner or an intellectual giant (Einstein, Da Vinci, whomever). Let’s say that this puzzle relates to cutting-edge science and technology. How many people of average intellect equal one genius?

Rod Furlan recently published On accelerating the Artificial General Intelligence timeline on his blog. He writes:

There is also the issue of energy consumption, while a human brain runs on about 20 watts, a traditional supercomputer capable of rivalling it would require it’s own power plant to run. From the energy standpoint there is absolutely no doubt we are doing it wrong.

Just appreciate for a moment that when all you have is a Von Neumann hammer, everything looks like a Von Neumann nail. Unfortunately for us the human brain has very little in common with a Von Neumann machine.

… It is very unlikely that traditional computers will deliver human-level cognition with low energy requirements.

The factual points above are spot on. The “what this means” message is way off.

The human brain is truly a magnificent object, and it does a spectacular job at general problem solving given the time, space, and energy allocated to it. When we look at energy efficiency, there’s absolutely no way we’ll be matching the brain this side of the Intelligence Explosion.

If it took an entire power plant to run a computer capable of reproducing the computational power of the brain, then presumably two power plants could run two artificial brains, right? How many power plants would it take to run an artificial brain that is twice as big? My guess, given the parallel nature of neurons, etc, is about two power plants*. (I don’t think that lumping one brain onto another brain makes a mind that’s twice as effective, but I’m guessing that, say, doubling cortex size would be a good start.)

The human brain cannot double in size in a single generation, but who says a machine brain could not? Who says that it could not quadruple in size? Would the resulting mind then be twice as smart as [insert famous brainiac here]?

What would the world look like with a human-level machine? Probably mostly the same.
What would the world look like with a machine significantly more intelligent than the smartest human in history, that didn’t have to eat or sleep? I certainly don’t know, but I do know that it might even be worth spending billions of dollars on meeting energy needs.

* – This is a major claim. If the number of units doubles in a network, the number of possible connections can skyrocket. I think that this sort of combinatorial explosion isn’t likely to be an issue because I’m not talking about connecting every unit to every other unit. I think the neuroscience shows that the brain is quite modular, and that doubling the size of each module probably wouldn’t even double the bandwidth between modules or cause an exponential increase in connectivity within each module. We can debate this if you disagree.

Even if doubling the brain size isn’t an option, I see no reason that the brain of a super-genius should be significantly more energy-expensive than an ordinary human. It may even be as simple/cheap as tuning the software and providing the right development environment.

As a final note: This post is in no ways an endorsement of inefficient hardware. I think that quantum computers and memristors are super-exiting, and I can’t wait to see how they effect the price-point of more brain-like computing systems. I think that energy efficiency is not a roadblock as much as a weight on the progress of AGI. Shedding some of that weight would be most excellent.

6 Comments

1. Forrest
   Posted December 7, 2010 at 2:04 pm | Permalink

   The energy ratio of supercomputer to human brain is even more amazing when you recall that the brain is actually pushing atoms around (!) instead of just electrons as in a computer.

   For the human brain, 20W/1e11 neurons = 2e-10W/neuron. That’s tough.

   Most likely we’d build 10GHz neurons instead of the brain’s 1K Hz neurons. This means we could build 10G Hz/1K Hz = 1e7 fewer neurons by time slicing them. Since dynamic power scales linearly with frequency, we can use this time slicing to make a linear trade off of speed for number of neurons.

   So we build 1e11/1e7 = 10K neurons running at 10G Hz each. To fit this in 20W, we have a power budget of 20W/10K neurons = 2 mW/neuron, which is doable. Then add in static power (doable since static power isn’t effected by frequency and we only have 10K neurons), power for memory for 1e15 synapses (doable using non volatile memory), and power for communication (use the new direct chip-to-chip fiber optics).

   My back of the envelop says we get a human brain in <100W before an intelligence explosion.

2. Raelifin
   Posted December 7, 2010 at 2:28 pm | Permalink

   That seems overly optimistic, given the “processing” needed to dynamically rewire neurons/synapses, and keep the whole thing from melting…

   I’m not an expert though, so it might be that simple.

   Regardless, we agree energy isn’t the biggest problem.

3. anne
   Posted December 7, 2010 at 2:39 pm | Permalink

   But you’re only talking computation, not strategies, not software. What might make Einstein, Einstein is not his computation rate or his brain size, but the strategies he has learned.

4. Raelifin
   Posted December 7, 2010 at 2:46 pm | Permalink

   To a degree this is true, but it’s also probably true that Einstein was genetically gifted with higher capacity for intelligence. Regardless, my point remains that energy efficiency is not a prerequisite for the technology to change the world.

5. Forrest
   Posted December 9, 2010 at 1:51 pm | Permalink

   One more note on dynamic power; it scales as the square of voltage, ie:

   Power = C * V^2 * f

   Modern CMOS voltage is around 1V, whereas the brain is around 1 mV. When you square that, you end up with a million-fold power advantage of neurons over CMOS just based on the voltage alone. One single fact explains a lot.

6. Forrest
   Posted December 9, 2010 at 2:19 pm | Permalink

   Yet more on power, and a pointer to an organization/learning model.

   The DARPA SyNAPSE project (2009-2016) has the following hardware goals:

   16^6 “neurons”/cm^2 ( = brain density)
   10^10 “synapses”/cm^2 (= brain density)
   100 mW/cm^2 (50x brain)
   total power = 1 kW (50x brain)

   One would like to hope that DARPA isn’t setting goals completely outside of the reality zone … but one might also not want to bet on it :-).

   I have posted these questions to the authors of a recent IEEE article on this,

   1. How far have we advanced on these goals?
   2. It it likely that all or some of them will be achieved by 2016?

   And am awaiting a response here:

   www.neurdon.com/2010/12/07/why-is-neuromorphic-computing-important/comment-page-1/

   The original article is here, where they also discuss a project to handle the organization and learning for this device:

   spectrum.ieee.org/robotics/artificial-intelligence/moneta-a-mind-made-from-memristors/1

Post a Comment Cancel reply

Your email is never published nor shared. Required fields are marked *