Jan Vorbrueggen <jan at mailhost.neuroinformatik.ruhr-uni-bochum.de> wrote:
>Robert Herndon <rmhj at rmi.net> writes:
>>> To paraphrase a joke among doctors: Yes, the brain has about 10^10 neurons.
>> And 10^11 of them are in the cerebellum.
>>Cortex has about those 10^10 neurons. Cerebellum, for all its neurons, seems
>a lot more like a specialised piece of hardware doing control of your limbs.
>>> I would presume that Vannevar Bush was familiar with the idea that
>> the brain had about 10^10 neurons, and was presuming equivalence
>> between neurons and tubes/gates. This isn't a fair comparison,
>> however, as neurons often have up to ~1000 connections, and
>> these connections are neither simple, symmetric or equivalent to
>> each other, nor immutably fixed in character.
>>Quite so. I'd say the three-dimensional character of the wiring and its
>plasticity is what makes otherwise inferior hardware that performant. And
>our estimate of the complexity of a single cortical neuron is still growing.
My AI textbook lists a factoid saying that a typical desktop computer could
do "10^5 neuron updates per second". It was written in 1995, and listing
figures for a processor with ~100,000 gates. So evidently, a
nearly-obsolete PC CPU running neural net software could update about 10^5
neurodes per second.
The number of "neuron updates per second" for the biological brain is
listed as 10^14. So there is a 9 order of magnitude difference. Based on
the general speed increase over the last 4-5 years, that should be about 8
orders of magnitude by the present day.
If we assumed Moore's Law would continue to hold, this gives us an estimate
more optimistic than the "45 years for brain power on the desktop" estimate
posted earlier. By 2030, the heuristic gives a desktop PC with 10^12
neuron updates per second, 10^14 by 2040.
Note that these figures are for running some kind of (unstated) typical
neural net software on a typical desktop CPU. With some kind of
purpose-optimized system, the job could likely be done with significantly
less computing complexity.
