Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.ai.philosophy > #2560

Re: John McCarthy R.I.P.

Cross-posted to 2 groups.

Show all headers | View raw

"Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> wrote:
> On 10 Nov 2011 21:07:03 GMT, Curt Welch wrote:
>
> > "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> wrote:
> >> On 10 Nov 2011 16:03:00 GMT, Curt Welch wrote:
> >>
> >>> But we do know some other important things, like approximate
> >>> information bandwidth of the eyes, and touch, and of the control
> >>> signals needed to move our arms and legs.  And these bandwidth based
> >>> numbers show that the amount of information the flows thought the
> >>> brain, is trivial compared to our modern data processing munchies.
> >>
> >> That is not the bandwidth of "information", it is one of the
> >> representation of that information. The information space encoded by N
> >> bits has the cardinality of 2**N. This is far beyond any possible
> >> computational power when approached by brute force. Which means that
> >> there is more than enough room for David to beat Goliath.
> >
> > Not sure what point you think you are making there.
> >
> > Are you saying that we can't push a MBit/sec through a computer because
> > it's information space is 2^^(MBit)??
>
> Pushing implies data, not information. If you are talking about
> information that would push + interpret. The complexity of interpretation
> depends on the number of distinct states in the 2**N space.

Well, it does and it doesn't.  The potential MAX amount of computation
could be said to expand at that rate, but it's just silly to talk about it.

> > Bandwidth doesn't tell us how much computation has to be performed on
> > the data as it passes through, but it gives us a starting point.
>
> It gives the lower bound. 2**N is the upper bound.

Yeah, well, at 1 Mbit per second, the 2**N state space quickly expands to
the point of being too large to explore even if all the matter in the
universe were turned into a giant computer.   So that's a stupid upper
limit to even bring up.  It's by definition not the "real" upper limit
needed for anything the brain is doing.

But your point is certainly valid that if we don't know how much
computation we need to do, then just talking about the bandwidth doesn't
tell us much of anything. The real question hinges on how much computation
is needed per bit "pushed" and since that is ultimately still unknown, we
really just don't have the facts we need to make a reasonable estimate.

> >>> I think neurons suck at information processing, and the great
> >>> complexity we find in the brain, is there MOSTLY not due to how
> >>> "smart" humans are, but to how much evolution had to bend over
> >>> backwards with massive parallelism to get useful function out of one
> >>> of the worse information processing technologies on the planet.
> >>
> >> Possible yes, but again it could be apples and oranges. We don't know
> >> how much the computational substrate of neurons of brain vs.
> >> electronic switches of CPU influences the complexity of the task. Your
> >> marble machine is an example. Or consider a pendulum, as a machine.
> >> That machine solves differential equation of pendulum. It does this
> >> quite good. How good is iPhone in solving this equation? I bet that
> >> iPhone is much worse. Is iPhone simpler or more complex than pendulum?
> >
> > That's mostly nonsense.  The pendulum doesn't "solve the differential
> > equations" so the point is nonsense.
>
> There is no magic difference between the CPU and pendulum, both are mere
> physical systems. Any semantics of what a system "does" is our
> interpretation.

Sure they are just physical systems.  But we have words that talk about
characteristics of physical systems, such as "wheel", and, "clock", and
"digital computer".  You can't call a lake a "wheel" just because both are
physical system.

You used the word "solve the differential equations", which like "wheel" is
a very specific type of behavior we find in physical systems, and one which
I'm sorry, but a pendulum doesn't do.

If you want to predict the motion of the pendulum using the LANGUAGE OF
MATHEMATICS, you would need to solve the differential equations to make
such a prediction.  But the pendulum is not a language processing machine
which is busy at work solving differential equations in order to predict
the motion of some other pendulum (or of itself).  To suggest it is is as
ABSURD as trying to argue that a lake is the same thing as a planet.

Yes, it is very much our interpretation of what the words "solve the
differential equations" means about physical systems, but by social
conventions, we do have a fairly clear definition of what that is and it's
not what a pendulum is able to do.

> >>> I strongly believe, that most people believe AI is far harder, and
> >>> far more complex than it really is, and that belief causes them to
> >>> make these computation estimates WAY too high - they don't want to
> >>> think of it as easy, because 1) that shows how great their failures
> >>> have been at trying to understand AI
> >>
> >> BTW, there is a question if the power of general intelligence were
> >> sufficient for understanding intelligence. This is not same as being
> >> intelligent. In which relation both problems are is unknown.
> >
> > Yeah, it's a valid question, but I'm fairly sure I know what
> > intelligence is, and that question doesn't get in the way of the
> > answer.  It's because intelligence is the emergent property of an
> > optimization process - a learning process.  What we can "understand" is
> > limited to what we can learn, and the brain is too complex to
> > understand itself in that way.  But the underlying process that creates
> > the complexity is not too complex to understand - which is how we can
> > expect to get around the problem of intelligence trying to understand
> > itself.  The answer is, we don't need to understand ourselves, we only
> > need to understand the learning process that created our (adult) human
> > intelligence.
>
> It does not simplify the problem. I mean it could simplify it if the
> system being taught could be considered as a black box, i.e. as a
> "hardware".

It must be hardware.  There is nothing else that exists in this universe.
Not sure what you mean by "is a hardware".

> It is indeed so with human pupils. But the idea of AI is not
> only about the process of learning (though 16 years of learning is a not
> what we would expect from an industrial AI system), but it is also about
> the "hardware". This hardware need to be built and has to be understood.

No it doesn't.  That's why you are failing to grasp how learning algorithms
work.

TD-Gammon used neural networks combined with reinforcement learning to
learn how to play the game of backgammon.  The guy that wrote it,
understood exactly how and why such a system would work.  But after he let
it play itself a few million games, it was able to play as well as the best
human players.

What it ended up doing, in effect, was evolving a function that could
produce an estimated value for any game board position.  It was in effect,
the learning machines estimation of the probability of winning the game
from any game board position.

The function created by the learning system, was a neural network which
used a few hundred weights (each weight was just a floating point number).

This program calculated optimal values for those weights - which created
the definition of the function that mapped a game board to a 0 to 1
probability.

Using that evaluation function to guide the selection of moves, the program
plays as well as the best humans.  SO the function the learning algorithm
created, proved to be roughly equal to the function the brain creates, when
a human goes about learning how to play backgammon.

Now, instead of using a learning algorithm to create the function, the
author could have attempted to set the values of the weights by hand - to
hand-program the evaluation function.

He in fact, had tried to do just that, in past Backgammon programs he had
written.  But the function created by the learning algorithm was better
than anything he had ever created.  And not only was it better, the author
had no clue how it worked - or why those values, made it a "better
function" than anything he had tried to created by hand.

Setting those weights, is how the program was "programmed" to play
backgammon.  But yet, the setting of those weights, is beyond the
comprehension of any human. There is NO HUMAN that could hand-program that
function.  The game of backgammon is too complex for a human to understand
at the level needed to hand program a solution like that by setting the
values of a few hundred weights of a neural network.

The author of the program actually patented the weights - aka the "program"
created by his learning algorithm.

Learning algorithms program computers for us, so we don't have to.  And
they do it by nothing more than calculating statistics for us, which is
beyond our ability to calculate (too many calculations for us to do by hand
in many life times).

By analyzing all the numbers of a million games of backgammon, they create
an "understanding" that goes beyond what any human can understand.  But
yet, we, as humans, can understand why the learning algorithm works.

There are learning algorithms that are in our reach of "understanding"
which can, create solutions, that our beyond our reach of understanding.

And that's one of the beauties of learning algorithms.  They create a level
of complexity, from their own simplicity.  They are something simple, that
gives rise to something more complex than itself.


> It is a big question if the architecture:
>
>    PC hardware (low-level)
>        |
>    Software system with an ability to learn (higher-level)
>        |
>    Training process
>
> would simplify things. Certainly it would not understanding.

It doesn't "simply things".  When it "learns" it is actually building a
machine - a machine that is normally more complex than the machine that is
doing the "learning".

And I think these sorts of systems certainly do "understand".  But we would
have to get into a debate about what "understanding" is.  Which I'm willing
to do if you want to.

> It is hoped that training magically produces intelligence. But again, in
> order to *know* this a few things must be shown:

Well, we have to define what intelligence is.  I define it in a way that
removes all the "magic".  So I don't have to "hope" that it "might" produce
"intelligence".  I know for fact it DOES produce intelligence per my
definition of it.

The only think I have to hope for is that my definition of intelligence is
correct.  I do hope for that, but that is being answered by my attempts to
create better learning systems.  It will be answered IFF this path yields
machines that people generally agree are intelligent  That is just a wait
and see problem.

> 1. the completeness of the software-hardware system, i.e. that there are
> states of the system corresponding to "intelligence".

Yes.

> 2. That the training process converges to one of these states.

Yes.

> The scenario described by Asimov in his novels: systems exposing
> intelligence, while nobody actually understand how do they work, is what
> many are putting their hopes into. Does not this vividly resemble
> searching for the philosopher's stone?

There really isn't anything magical about human intelligence. Skinner and
the other behaviorists figured all that out 70 years ago.  Anyone that
doesn't get that is just confused (and there are a LOT of people still very
confused about that).  The only thing waiting to be resolved, is solving
the engineering problem of how to build a practical reinforcement learning
algorithm that operates in the high dimension domain that humans and
animals operate in.  That engineering problem has proved to be a very
tricky one (no one has figured it out even tough lots of people have been
trying over the past 70 years), but good progress has been made, and we are
getting much closer to the solution every day.

-- 
Curt Welch                                            http://CurtWelch.Com/
curt@kcwc.com                                        http://NewsReader.Com/

Back to comp.ai.philosophy | Previous | Next — Previous in thread | Next in thread | Find similar

Thread

John McCarthy R.I.P. RichD <r_delaney2001@yahoo.com> - 2011-11-04 13:17 -0700
  Re: John McCarthy R.I.P. Don Geddis <don@geddis.org> - 2011-11-04 16:47 -0700
    Re: John McCarthy R.I.P. RichD <r_delaney2001@yahoo.com> - 2011-11-09 13:17 -0800
      Re: John McCarthy R.I.P. casey <jgkjcasey@yahoo.com.au> - 2011-11-09 14:04 -0800
        Re: John McCarthy R.I.P. "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-11-10 09:45 +0100
          Re: John McCarthy R.I.P. casey <jgkjcasey@yahoo.com.au> - 2011-11-10 01:24 -0800
            Re: John McCarthy R.I.P. "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-11-10 12:14 +0100
          Re: John McCarthy R.I.P. Antti J Ylikoski <antti.ylikoski@aalto.fi> - 2011-11-10 11:38 +0200
            Re: John McCarthy R.I.P. "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-11-10 11:54 +0100
              Re: John McCarthy R.I.P. Antti J Ylikoski <antti.ylikoski@aalto.fi> - 2011-11-10 14:09 +0200
                Re: John McCarthy R.I.P. "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-11-10 14:46 +0100
                Re: John McCarthy R.I.P. seeWebInstead@rem.intarweb.org (Robert Maas, http://tinyurl.com/uh3t) - 2011-11-13 18:00 -0800
                Re: John McCarthy R.I.P. "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-11-14 12:07 +0100
              Re: John McCarthy R.I.P. curt@kcwc.com (Curt Welch) - 2011-11-10 16:03 +0000
                Re: John McCarthy R.I.P. casey <jgkjcasey@yahoo.com.au> - 2011-11-10 12:16 -0800
                Re: John McCarthy R.I.P. "James" <no@spam.invalid> - 2011-11-10 13:00 -0800
                Re: John McCarthy R.I.P. "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-11-10 21:50 +0100
                Re: John McCarthy R.I.P. curt@kcwc.com (Curt Welch) - 2011-11-10 21:07 +0000
                Re: John McCarthy R.I.P. "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-11-11 11:43 +0100
                Re: John McCarthy R.I.P. curt@kcwc.com (Curt Welch) - 2011-11-12 22:38 +0000
                Re: John McCarthy R.I.P. casey <jgkjcasey@yahoo.com.au> - 2011-11-13 01:32 -0800
                Re: John McCarthy R.I.P. curt@kcwc.com (Curt Welch) - 2011-11-14 15:28 +0000
                Re: John McCarthy R.I.P. "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-11-14 16:57 +0100
                Re: John McCarthy R.I.P. curt@kcwc.com (Curt Welch) - 2011-11-17 22:19 +0000
                Re: John McCarthy R.I.P. "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-11-18 10:51 +0100
                Re: John McCarthy R.I.P. casey <jgkjcasey@yahoo.com.au> - 2011-11-14 11:42 -0800
                Re: John McCarthy R.I.P. "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> - 2011-11-13 12:45 +0100
      Re: John McCarthy R.I.P. RichD <r_delaney2001@yahoo.com> - 2011-12-14 10:28 -0800
  Re: John McCarthy R.I.P. seeWebInstead@rem.intarweb.org (Robert Maas, http://tinyurl.com/uh3t) - 2011-11-13 16:00 -0800

csiph-web