Prev: need help
Next: mainframe material
From: tim Josling on 11 Jan 2008 14:39
On Fri, 11 Jan 2008 10:04:27 +0100, Sebastian Hanigk wrote:

> "Charles Hottel" <chottel(a)earthlink.net> writes:
>
>> When computers and software can simulate a human brain at speeds faster
>> than a human there will be little need for human programmers.
>
> A few problems: we still do not understand the human brain and looking
> at it from a complexity point of view it seems a very daunting task for
> simulating neurons.
>
> Look at it that way: we understand the heart muscle quite well, but are
> not able to construct a replacement that works for decades like the
> original (in fact we do recycle hearts due to their superiority).
>
>
> Sebastian

If you read "On Intelligence" by Palm Pilot-inventor Jeff Hawkins

http://en.wikipedia.org/wiki/On_Intelligence, there is a lot of evidence
of progress in understanding how the brain works.

There is also a lot of interesting material in Ray Kurzwel's book "The
Singularity Is Near".

In some respects technology has not lived up to its promises. Energy is
definitely not "too cheap to meter". No flying cars.

Artificial intelligence has been the biggest disappointment. If you look
into this, one reason is very clear.

Early computers were physically large but, compared to the human brain,
were slow and weak. The brain has about 100,000,000,000 neurons. Each
neuron has about 1,000 connections to other neurons on average. The cycle
time is about 20ms. It seems the connections between neurons are the
active elements. This translates into about 5 X 10**14 computations per
second. Other approaches come up with roughly similar numbers.

Compare that to my desktop system which does about:

2 (instructions per cycle) X 2.4 X 10**9 (2.4 ghz) X 4 (CPUs) = 1.9 X
10**10 operations per second.

We still have a gap which is a factor of 20,000 or so. Imagine trying to
win a car race and the other guy's engine is 20,000 times more powerful
than yours.

The gap in power usage is even more extreme. The brain uses about 40 watts.
My 4 CPUs use about 100 watts. Per computation, silicon is about 100,000
times less power-efficient than your brain.

If you go back to the height of the AI era, the gap was far higher - a
factor of 10,000,000 or so between the processing power then available and
the human brain.

So it is not surprising that the attempts to match the brain were
miserable failures. If you read the literature, again and again the
problem is that the algorithms don't "scale". They handle toy problems but
are impossibly slow on real life problems.

I am not saying it is all about lame hardware. There are gaps in
algorithms too. But it's hard to overcome a factor of 20,000 or more.

---


The good news is that Moore's law continues unabated and seems likely to
continue. Individual CPUs are not getting much faster, but we are getting
more cores on each chip. My 4-way cost about $1200 USD earlier this year.
Within 5 years we are looking at 50 way and in ten years about 1000 CPUs
on a chip. In 20 years, 100,000 CPUs on a chip!

It's characteristic of exponential growth that nothing much seems to
happen and then there appears to be an explosion. Examples include plagues
of animals, disease epidemics.

The best example is the internet. It had steady exponential growth for
decades. For most people it seemed to suddenly come from nowhere.

My expectation is that as processing power approaches that of humans, then
"suddenly" computers will be doing all sorts of things they would not do
before.

The human brain is wired to do linear extrapolation. This does not work
very well in the face of exponential phenomena such as Moore's law.

Tim Josling
From: tim Josling on 12 Jan 2008 03:26
On Fri, 11 Jan 2008 20:26:36 +0100, Sebastian Hanigk wrote:

> "Charles Hottel" <chottel(a)earthlink.net> writes:
> I've brushed up on the author a bit and will get my hands on books, but
> regarding the singularity theory I have my doubts, especially regarding
> the often cited hypothesis of Moore. In my opinion we should expect a
> logistic growth model which will taper of on a certain niveau after a
> rapid growth phase.
> Sebastian

Any exponential growth must eventually taper off. The question is "when".

Kurzweil makes a strong case that the exponential reduction in cost per
calculation has been doing on for at least 130 years, across numerous
technologies: from memory - humans, mechanical, electro-machanical, valves,
transistors, ICs etc.

I would also argue that if the human brain can calculate at X calculations
per second, that suggests that it *is* possible to calculate at that rate.
It does give hope that at least that level of performance is feasible.

Even given current technology, custom circuits can provide a 100X speed-up
for specific tasks, if it comes to that.

Kurzweil's predictions have been quite good over a periof of nearly 20
years, though slightly optimistic on some things. I went over to the
library yesterday to have a look at some of his older books just to check.
The wikipedia article covers this issue to some extent.

On the issue of how to live long enough to live forever, he has change =d
his views on the optimal diet very radically. His initial theory was the
high carb Pritikin/Ornish diet, but lately he has been pushing low-carb
and lots of supplements.

Tim Josling
From: tim Josling on 12 Jan 2008 03:31
On Fri, 11 Jan 2008 21:51:38 -0500, Charles Hottel wrote:

> "Howard Brazee" <howard(a)brazee.net> wrote in message

>> Making a human brain like computer may be like making a human body
>> like truck. While there is some utility in creating a truck that can
>> walk on two legs, it isn't the most useful design.
>
> Scientists want computers that work like human brains because the process of
> constucting them will hopefully give new insights into the nature of
> intelligence.

Also because there are lots of things computers can't do, and as a result
people have to do them. If everyone could have a computer that has the
capabilities of a competent secretary, that would be very useful.

Also if we can make computers that are smarter than us, maybe they can
solve problems we can't solve.

Kurzweil paints a picture of a world where computers are seamless add-ons
to our brains. It is not like having a computer that can translate from
another language, it will be more like a coprocessor that means *you*
understand the language. The possibilities are endless.

Tim Josling
From: tim Josling on 12 Jan 2008 14:45
On Sat, 12 Jan 2008 10:01:22 -0500, Charles Hottel wrote:

> Besides learnig more about intelligence we may also learn more about
> parallel processes and how they can communicate and be synchronized. As
> Howard alluded to via "voting mechanism" the brain seem to have many
> parallel processes and also seems to pre-compute and store patterns instead
> of computing them from scratch all the time. The actual processing speed of
> the neurons seems quite slow compared to computer circuits.

Individual neurons are slow, with a cycle time of about 20ms. That's maybe
a million times slower than silicon. But there are a phenomenal number of
neurons, and each neuron has on average 1,000 active links to other
neurons where processing of some kind occurs.

The implication of this, as you point out, is that there must be a huge
amount of parallel processing going on.

Give me 100 trillion active processing units, even if slow, and with the
right algorithms I can do great things.

Graphics cards currently employ a high amount of parallelism. Individual
units are a lot slower than modern CPUs. But for the right problem they
are far faster than a standard CPU.

Tim Josling
 | 
Pages: 1
Prev: need help
Next: mainframe material