Re: Curtnetrons Don't Do Parity

Curt Welch wrote:

Michael Olea <oleaj@xxxxxxxxxxxxx> wrote:

Curt's claim: one of his nets, not a collection of such nets, not some
configurtaion of such nets working in concert, but ONE OF HIS NETS, is
capable of learning, via RL, anything that needs to be learned to support
fully human performance.

Yeah, that's mostly the goal. But there's a few points you seem to keep
leaving out.

I've always made it very clear that the feed-forward learning net would be
the core learning module. It would be where the learning was applied.
Around it, would be needed additional support systems. The most important
being feedback paths, and the second most important being various types of
state storage. The most simple of these I've suggested would be pulse
generators which had their internal state adjusted by the outputs of the
learning network. The pulse generators would both act as the output of
the full system, and created feedback by sending their outputs back to
other inputs on the network.

I understand that. I keep leaving those points out, because the question I
am asking is what is it possible for the nets to learn, and what is not
possible for the nets to learn - however they are trained. And for that
question it does not matter that some of its outputs might get fed back
into the net. It is still a feedforward device - a black box with a set of
inputs and outputs. So we can analyze the box in isolation - what
transformations from input to output can it effect.

You can't create any type of central pattern generator without feedback
paths like this. Some type of feedback paths are required to create
anything near human behavior. This is why that extra hardware has always
been assumed by me to be part of the full system required to create human
behavior with this technology. I've never claimed the feed-forward pulse
sorting net could do it alone.

So, if you want to prove something about the entire system, you should be
attempting to test the limits of the entire system and not just the core
learning module which is nothing but a classifier trained by reinforcement
learning.

But unless the rest of the system has tunable parameters that change in
reponse to experince, i.e. the rest of the system "learns", then we can
restrict the analysis of what the whole system can learn to an analysis of
what the feedforward net can learn.

Second, I've also never claimed one of my networks would be able act
human.

Sure, you have - you're just getting more careful with your claims these
days. ;-0

I've always said that real humans have a large set of optimizations
to give
us the exact set of powers we have. The point of this learning technology
is to match the general power we have which I believe most people are
talking about when they talk about human intelligence. I believe, as I've
said a thousand times here, that a strong real time, reinforcement
learning system is the core technology missing from our understanding of
what the
brain is doing to create our intelligent behavior.

But, as you stated in the post a few minutes before this one, there are two
issues:

"Of course, the net is intended to be able to learn different
classifications though reinforcement training, so we really have two
questions to answer in the long term. The first is what type of
classifications can it perform, and what type can it not, not matter how
it's configured. And the second is, of those, what is it able to learn,
and not learn."

So RL is not the only issue. It is the limitation to feedforward nets, and
to conservation of pulses, that give me strong doubts that the nets can
form a sufficintly rich "hypothesis space" to match human learning
abilities - or if it is in theory possible you would need a "brain the size
of a planet" to do it. That's my hunch. But it is still interesting to try
and prove what its capabilities and limits actually are.

How much that core
technology would need to be optimized, and split into sub-units, in order
to duplicate the exact powers, and weaknesses, of a human, I've never
claimed to know. That would be something we easily should be able to
figure out through trial and error testing once we have a strong core
learning technology to work with.

That changes things some, but each sub-unit would have the same limits. Now,
if the subuints are connected in loops, it becomes a different story.

What I believe you have correct about what I've claimed, is that a single
network, configured into it's required supporting external frame work to
support some external state and feedback paths, is the only type of
technology needed to explain all intelligent human behavior. By that,
I've claimed that any single thing humans do which is seen as intelligent,
I expect to be able to duplicate with this type of technology.

Yep. That is the question at issue. And that is the point Dan is missing.

I of course have no clue if my claim is correct. It's just my current
working assumption. I may talk about it as being a fact (that's just my
own arrogance talking), but push comes to shove, I've never denied that it
was anything other than my best guess. But, it's still my current best
guess because what you have shown so far hasn't told me anything I didn't
already know years ago. But if you continue along this path, there's a
good chance you will bring some important and useful insight to me about
what this type of design can do and can't do.

You have said a couple of times that all I have done is prove that the nets
cannot do what they were not designed to do, and we already knew that. But
that's not quite correct:

/*
So far however, all you have done is prove the nets can't do things that
they were never designed to do, and which I never implied they could do.
But you might be close to actually proving something useful about their
limitations.

The feed-forward network was not designed to produce any pulse train for
output. It's pulse conserving design is there for a reason, and by nature
of that design, it obviously can't do things that requires it not to
conserve pulses. So far, all you have proven, is that the network can't
produce two pulses out for one pulse in. Duh. It also can't fly.
*/

I'm not trying to prove that. I am taking that as a premise, and exploring
the consequences - what are the limits given the premise.

If I can get a clear understanding of something it can't do, and a clear
understanding of why it is important to do that in order to create human
like intelligent behavior, then I will have something clear to fix by
changing the design. But without a clear indication that the current
design is wrong, I don't have any reason to believe it's the wrong
approach.

That's the goal. One other point, looking at the nets as mapping bit-strings
to bit-strings is just one way of characterizing what they do. It does not
matter that you see them as sorting pulses into buckets. And, by the way,
I'm assuming there can be multiple pulses traversing the net at any time -
each input line generating pulses trains independendly of the others - if
you want to restrict it to a one pulse at a time machine, then I will
continue to analyze the multiple streams at once case, and call it a
generalized curtnetron.

Anyhow the string-to-string mapping brings out some features. This allows
some machinary to be brought to bear on possible pattern descriminations.
But another characterization I think is worthwhile is the "wave-mixer" view
I mentioned in passing.

-- Michael

.

Relevant Pages

  • Re: learning ethical hacking
    ... there is no end of learning, and frankly, that's why i sometimes get ... >> learning the technology behind it all. ... > thoroughly tested by folks who DO understand the design, ... > a certified computer examiner, learn to recover trace data left behind by ...
    (Security-Basics)
  • RE: learning ethical hacking
    ... > learning the technology behind it all. ... thoroughly tested by folks who DO understand the design, ... achieved that level of understanding. ...
    (Security-Basics)
  • Re: Why dont we have a stong AI by now? (To Curt Welch and other)
    ... a matter of good intuition. ... It amounted to an attempt to do a pulse version ... That was off track only with regards to learning new ways of doing ... Brain Damage and the Mind, ...
    (comp.ai.philosophy)
  • Re: Is AI all about time?
    ... Using your pulse network of pulse sorting units based on the ... in two steps with two different learning systems. ... the higher resolution break down of features you end up with. ...
    (comp.ai.philosophy)
  • Re: What is missing from AI ?
    ... I focus only of principles of temporal learning. ... and that all you need to know is how your logic gates ... orientation sensors that input pulse coded data and output ... Or some "neural" network equivalent. ...
    (comp.ai.philosophy)