Re: What did that thread indicate?

"JGCASEY" <jgkjcasey@xxxxxxxxxxxx> wrote:
> Curt Welch wrote:
> > "JGCASEY" <jgkjcasey@xxxxxxxxxxxx> wrote:
> "JGCASEY" <jgkjca...@xxxxxxxxxxxx> wrote:
> > > Curt Welch wrote:
> > > > Well, that's exactly what my nodes do.
> >
> >
> > > But to get some cred, Curt, you need to show
> > > what your net can do. And I mean *show* what
> > > it can do not simply say it can do this that
> > > or the other.
> >
> > But it's just so much more fun to talk instead
> > of doing real work. :)
>
> At least you are being honest.
>
> The other reason is, I suspect, that you have no
> idea where to go next with your net.

No, that's not really the problem at all. The problem is I know exactly
where I have to go and it's a lot of hard boring work which I find very
dificult to motivate myself to do on my own when there is such many other
things I am forced to work on to feed myself.

Up until now, my AI work has been a very exciting hunt for an answer. I
was motivated to do it becaue it always felt like the answer might be under
the next stone because I kept finding clues that showed I was getting
closer. I felt there might be a gold mine of fame and forture just wating
to be grabbed by a few more hours of thought.

But then I found the answer I spend all these years looking for. And it
turns out there is no fame or fortune here. It turns out that I'm the only
one that cares this answer was found. :)

> We know what it can do but not what it
> can't do.

This is where I have to go next. I know it does what I was looking for.
But what I don't know is how far that will take us to general AI. So I
have to test and experiment, and evolve the design, to see how far it can
go. I have to figure out what it can't do.

I have to make up extremly simple low-level reinforcement learning
problems, and test the net to see how well it works on that problem. Then
study possible variations of toplogy and learning algorithm adjustments to
see how that effects learning speed and convergence properities. I have to
continue developing stronger understanding of what this class of network
can do for increasinly harder problems until I either reach a dead end
(deciding that this approach can't go any further and a new approach must
be found), or until the approach takes us to full human level intelligence.
It's a long road because so far, I've seen no sign of a dead end. So now,
instead of dreaming up a new approach (which is fun work for me), I have to
simply grind through with testing and development of the current appraoch
to see where it leads. And that's what I'm having problems motivating
myself to do.

> But writing a GOFAI program or artificial neural net
> program of some kind that "learns by reinforcement"
> is a long way from understanding human intelligence.

Yes, very long because a simple reinforcement learning machine is a very
very low level start.

However, I think it's a straight shot path from here to there if you have
the correct reinforcement learning machine. I have my own answers to
explain how all issues of human intellgence can be answerd by a
reinforcement learning machine. But I don't know if my higher level
answers are valid, or just wishful thinking. That can only be answered by
following the path and doing the experimentation and testing.

> Just like Richard Dawkins biomorphs that use
> artificial selection are a long way from knowing
> what mechanisms were required for life to evolve.
>
> There is more to evolution than natural selection

Evolution is only two things. Change between stable states + selection.
The entire maco level (atomic and above) stucture of the universe can be
explained in these terms of a long history of this processes at work.
There's not really much evidence at all to believe otherwise.

The hard part is trying to reverse the history of that processes to explain
how we got to where we are today.

> and for the same reason there is more to human
> intelligence than reinforcement. There is a design
> that is *capable* of learning as a result of
> "reinforcement" in the right environment.

Right, but what I've been looking for all these years is just that. A
design that is capable of learning what it needs to learn. It was damn
hard to find one that actually was cablable of learning all the things it
needed to learn (especial the training of "hidden layers" with
reinforcement learning). But now I have one. It works as expected on
everything I've tested it on. Now I just have to have to do a ton of grunt
work to see how far I can take this design.

> The theory of evolution for example had issues
> back in the time of Darwin, not because the idea
> of natural selection didn't make sense, but rather
> the mechanisms for it to happen could not be
> imagined. The discovery of DNA was the first step
> in understanding these mechanisms.

Well, DNA is not the mechanism of evolution. DNA is one of many creations
of evolution. It's a machine designed and built by the process of
evolution. It is a machine which happens to explain how DNA based life
forms evolve, but it doesn't explain the fundimential force of evolution
which shapes the planets and stars as much as it shapes the animals.

The fundimential force of evolution is easy to explain. Like I said, all
you need is a substrate that can form stable configuations combined with
local forces effecting constant localized changes. That creates the
selection processes which kicks off evolution.

But, filling in the gap between that, and human intelligent life forms, is
a huge gap to fill. And that's what eveyone is busy doing. DNA is just
once piece in that huge puzzle.

Likewise, understanding human intelligence is another important piece in
the complex puzzle of evolution.

Likewise, bridging the gap from a simple low level learning system to full
human intelligence is a large gap to be filled. I've now got a low level
learning system that has all the basic properties I think it needs to
bridge that gap. Now I just have to see how much it can in fact bridge and
that's not a few hours or works, it's months and months of hard work
studing the properites of networks with hundreds and thousands of nodes.

To me, this type of network is the DNA of intelligent beahvior. It's as
important to understanding intelligence, as DNA was to understanding how
life evoloves. But, having found DNA, was just the first step.
Understanding how it actually manages to code all the information of life,
is much harder. Having found my network design is much the same. It seems
to me to have the power needed to explain where human intelligence comes
from just like DNA seems to have the power to explain where human life
comes from. But, going from this simple network design, to a full
functioning system that has powers equal to human intelligence, is a long
research path. There is a huge number of important details still left to
fill in even if the the path is workable.

It is something I will continue to work on it. But when life gets busy (as
it always does), it's easy to find a few minutes escape from my life by
wirting another Usenet post which I know will be done shortly, but very
hard to get started on a project which I know will consume days or months
of my time when I've got so many other things I need to be working on in my
life just so I can pay my rent and put the kids through college. :)

For example, I'd love to do nothing for the next year expect do research on
my network design. But this morning, the new firwall I orderd last week
showed up, and it's waiting for me to study and master, and then redesign
my network of servers to use. And that's of course just 1 of about 10
important projects I need to get done yesterday. But instead of working on
that, I decided to waste another 15 minutes writing this post because
thinking about AI is a hell of a lot more fun to me than installing
firewalls and everything else I must do these days....

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

Relevant Pages

  • Re: brain thoughts
    ... the lessons of evolution don't get applied to the modules ... The modules are not where the learning in evolution ... module representation system called DNA. ...
    (comp.ai.philosophy)
  • Re: No easy road to AI
    ... Yes, and for any circuits created by evolution, we can assume evolution ... And we can assume lots of that happens with humans as well. ... our innate sensory systems to fit what worked well for our survival. ... data into a generic learning system that must find, in that data, what is ...
    (comp.ai.philosophy)
  • Re: No easy road to AI
    ... the difference between building learning machines, ... Evolution didn't do the later - it did the former for us. ... mechanisms in the brain, they exist at every level because they ... In the case of humans all high ...
    (comp.ai.philosophy)
  • Re: The Best Explanation: Creation or Evolution?
    ... The Science Framework for California Public Schools adopted by the ... A scientific explanation does more than provide a plausible account; ... The goal of this study on evolution is to determine to what degree the ... human intelligence, we are constrained to propose that human intelligence ...
    (talk.origins)
  • Re: No easy road to AI
    ... back to "evolution solved that for us". ... evolution millions of years to evolve a human brain. ... But the way you deal with high dimension problems in general, ... learning problem and solutions have been found to make those practical by ...
    (comp.ai.philosophy)