Re: How much intelligence?

"JGCASEY" <jgkjcasey@xxxxxxxxxxxx> wrote:
Curt Welch wrote:

It's true that people can do this. But it's also
true that behavior like that is easilly explained
in terms of reinforcement learning. All langauge
behavior is easilly explained in terms of
reinforcement learning.

It depends what counts as an "explanation".

Sure. :)

If you
think RL easily explains language acquisition you
need to give a demonstration, you need to spell it
out, not just imagine it happening.

Have you read Skinner's book? I suspect he did a fine job spelling it out
50 years ago. That's why he wrote that book. I've not read it because I
don't need to be convinced it's possible. I already believe it.

I can believe it simply because a context sensitive reinforced trained
reaction machine is general enough to explain any behavior of any machine.
Anything, and everything we do (by a simple fact of the laws of physics),
MUST be a physical reaction - it must be the result of energy flows which
are seen as a chain of reactions (it must be a chain because energy is
conserved in this universe). Everything that every machine does is a
physical reaction to previous energy flows.

All machines can be translated into a generic machine which responds to
every stimulus by changing internal state and then producing outputs as a
function of internal state. Turning machines do this with only discrete
internal states, but real machines can use analog internal states. If you
start with a general purpose machine like this and change it's behavior
over time with reinforcement, there's no reason to believe that the machine
can't be shaped to produce any possible behavior of any machine.

So simply showing it's possible, is trivial, but not very informative.
What's non trivial is showing what type of reaction machine is needed to
produce all human language behavior, and to how it's shaped by
reinforcement learning to get it into the correct configuration to be a
English speaking machine.

If we can't find the machine, and the learning algorithm to shape it, then
we can't do it. But that's not why people reject the idea that Language
can be learned by reinforcement. They reject it because they simply don't
understand the true power of reinforcement learning to shape complex
behavior. The just don't understand reinforcement learning.

Another example Pinker makes reference to is the
imitative behavior of chimpanzee. One example was
a chimp that seemed to imitate the washing of
dishes but the dish wasn't necessarily any cleaner
after imitating the rubbing motion and this is
apparently typical. The chimp did not twig to the
purpose or goal of rubbing the dish was to clean it.

Autistic people apparently have a similar problem.

Maybe this was because they didn't have enough short term memory to
recognize the fact that the long procedure caused the state of the plate to
change from dirty to clean? Maybe autistic people have a brain difference
that causes their brain to allocate more neurons to short term effects so
they get a far higher resolution memory of short term events but a far
lower resolution of longer term events (many seconds out). In other words
their resolution vs memory curve drops off much faster than normal humans.
This is the type of thing that can be controlled in my type of network
simply with the topology. If you use a long and narrow topology, it will
have very short term memory, but if you create a large fan-out topology,
the same number of neurons will have a much longer term memory, but less
resolution of information over the period. I could easily see how some
simple gene change might slow up or speed up the formation of the network
in the brain which could lead to it building a network with a very
different short term memory profile.

This in fact might be the type of thing that makes it harder for apes to
learn language. Their brain might be more optimized to short term
reactions (maybe more useful for climbing trees and jumping limb to limb)
which prevents them from responding correctly to the long term context set
up by a string of English words. And of course, simple brain size could be
a big factor in the amount of context we can hold and respond to in our
short term memory.

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

Relevant Pages

  • Re: Syntax and robot behavior
    ... now because what I have to say below about reinforcement learning touches ... of the brain accessible only indirectly, ... entire neocortex is made up of the same fundamental hardware structure ... Yes, but a computer memory can, though structurally amorphous, ...
    (comp.robotics.misc)
  • Re: Reinforce learn this
    ... real advantage to calling anything in the brain software. ... each module has a different design specified in DNA, ... Well, externally, we see classical conditioning which the Behaviorists ... The process of evolution itself is a reinforcement learning ...
    (comp.ai.philosophy)
  • Re: Reinforce learn this
    ... Reinforcement learning can be reduced to explaining ... is the brain that is actually ... speaking English - not the brain were were born with which had no ... As you said with the sun example, ...
    (comp.ai.philosophy)
  • Re: Towards a theory of the semiotic mind-body link
    ... by the complex network in our brain which defines how we ... - aka it would be a brain wiring dump. ... decent with modification - by a reinforcement learning process. ... Humans design large complex systems by putting together small modules, ...
    (comp.ai.philosophy)