Re: What is innate and how can we determine it?

casey <jgkjcasey@xxxxxxxxxxxx> wrote:
On Jun 24, 1:10=A0pm, c...@xxxxxxxx (Curt Welch) wrote:
casey <jgkjca...@xxxxxxxxxxxx> wrote:
On Jun 23, 1:04=3DA0pm, c...@xxxxxxxx (Curt Welch) wrote:

... when I do say things like "it's just reinforcement
learning", and "Learning is all that is important",
I'm not talking about all human behavior. I'm talking
about _intelligence_.

No you are talking about *your* definition of intelligence!
And that has been a major confusion in all your past posts.

Yes, I could see it causing confusion for others, but
after 5 years of you knowing what my definition of
intelligence is, why would it confuse you?

It doesn't confuse me, now, but why would you want to confuse
others by persisting in using your own private definitions?

When you say things like "people don't understand that
real intelligence is reinforcement learning" there is
nothing to understand except your eccentric use of the
word intelligent.

What you mean is that human intelligence is the result
of reinforcement learning. That is all you have to say.


No I don't mean that John. I mean reinforcement learning
IS intelligence.

Again you persist in confusing the issue with your private
use of the word "intelligence". I use the word in a way
others will understand it. After 5 years of me using it
that way why would it confuse you!

I was not confused. :) You told me that I should put your word in my mouth
and I simply response that I would not do that. Why would that confuse you
into thinking I was confused? :)

Again, you need to stop seeing the behavior as being the
intelligence, and learn to see the behavior of "learning
behavior" as being our real intelligence.

We label behaviors and one of those labels is the word
"intelligent" as the opposite to "stupid".

Right, but as I go on to explain, you stop short and fail to correctly
label the _entire_ behavior in your examples becuase you don't also include
the most important of the behavior - which are the preceding events that
caused us to acquire the part of the behavior you do choose to acknowledge
or the fact that the consequences of our actions will always shape our
beahvior.

I'm labeling the same beahvior you are, I'm just include the parts you
choose to pretend aren't there as well and then abstracting out what's
really common about all of them.

I don't talk like this for the fun of it or to try to
irritate or confuse you or others. I do it because I'm
searching for the truth and I think what I say here is
the truth.

No you are using your own definitions and being confusing.

When the world is being stupid, it's not up to me be act dumb so they don't
get confused.

When we observe a behaviour we do not need to know if
that behaviour was the result of reinforcement learning
or if it was innate in order to label that behaviour
as intelligent or stupid which are only two of many
possible behavioural labels.

But you need to stop looking at the behavior and labelling
it as "intelligent" because it's not our behavior that
makes us intelligent.

More confusion you need to stop using the word intelligent
in that way. We are NOT saying it is our behavior that MAKES
us intelligent. We do not make any reference to the cause
of intelligent behavior. Or stupid behavior. Or aggressive
behavior. You say the CAUSE is reinforcement learning. OK
fine. Intelligence is the result of reinforcement learning,
that is all you have to say.

It's our ability to learn behavior that makes us intelligent.

I know you can't tell the difference between a noun and a
verb as you see them as the same thing but others do not
define dance as a thing but rather something a thing does
regardless of the cause.

My definition of intelligence and your definition of intelligence are both
verbs john. They are both labels of behavior. I'm just abstracting out a
higher level feature of the behavior than you are.

Do you realize the difference between the behavior of
playing chess and the behavior of being trained to play
chess?

What do you think?

I think I'm hungry. :)

None. Except one has a longer scope and includes the
events that caused the behavior to show up in us where
as the other only shows the final behavior.

First you say there is no difference and then you give
the difference, one has a longer scope and includes the
events that caused the behavior to show up. What is
the difference between a built house and a house that
is being built? This is really silly.

No john, you example is invalid. Neither of us are talking about the state
of the human at a single point in time. Or are you? We are talking about
the actions a human produces over a period of time. You are just focusing
on a shorter period of time.

If we look at the entire sequence, we see "real"
intelligence,

Yes I know your idea of "real" intelligence. It is
reinforcement learning.

When we make the mistake of thinking chess playing
is an example of human intelligence, and then try
to build that into a machine, we end up with a
machine that plays chess better than any human, but
yet no one besides some people in AI think the chess
program is a good example of human intelligence.

It is not an example of reinforcement learning if that
is your point.

It's because operant conditioning is the only behavior
people look at and think of as intelligence even if
they don't understand well enough to explain it like
that.

Yes I get your viewpoint. However the meaning of the word
intelligence depends on context and if you want to be an
effective communicator you will take that into consideration
before berating people for not "understanding".

The context here in c.a.p. is AI. When I chat informally with people I use
the word intelligence and understand their use of the word intelligence in
the fairly default ways we have all been trained to understand it.

This however, is a philosophy group ABOUT INTELLIGENCE. It's the single
most important topic debated here. The fact that no one has solved AI, is
a simple proof that no one understand what intelligence IS. AI is in fact,
a search for the CORRECT definition of the word intelligence. Once AI is
solved, we will know what the correct definition is. Until then, becuase no
one has solved it, the one thing we do know for sure, is that NO COMMON
DEFINITION IF INTELLIGENCE IS FULLY CORRECT.

As such, one of the single most import things to do is, is debate, and
attempt to find, the CORRECT definition of intelligence.

My position on the correct answer to this all important question is clear.
To tell me I'm "confusing people" (here in c.a.p.) because I'm not using
the socially accepted definition of the word, is to show you don't
understand the point of what needs to be done here.

We don't solve these sort of problems by doing what everyone else that came
before us has done - becuase every one of them FAILED to solve this
problem.

The simple line following network is no longer sufficient.
How can it modify itself to become even more complex? Does
it simply have to become bigger because it is "conditioning
all the way up"? Or does it make duplicates of itself in
which case how do the resulting networks integrate with
each other to enable them to produce behaviors that can be
associated with running over more discs?

Well, those are the questions to answer aren't they!

What you probably have no clue of, is that my chicken wire
network already answered all those questions. It already
addresses exactly how those issues are resolved in a single
learning system.

So you can use your chicken wire network in such a robot and
train it to follow lines and/or maximize running over discs?

Of course. It doesn't do a very good job of it (meaning the solution it
will find sucks), but it can do it. More important however is that it's a
working demonstration of a viable _approach_ to how to solve AI.

That is great. Why don't you publish your work or patent it?

I published it here. No one seems able to understand it.

Our motivation system produced human level intelligence
and I thought that was one of your goals. Learning to
survive in this environment I would suggest would result
in learning 3D navigation.

Our motivation didn't produce our intelligence. You just
don't understand what intelligence is. It's not "paw shaking",
it's "the ability to learn to shake a paw to get a treat".
Your view of intelligence is far too near sighted.

What I meant was our motivation, the rewards to be maximized,
resulted in the kind of intelligent behaviors we see in humans.

Yes, like war, and murder, and sex drugs and R&R.

If those rewards were pleasure in hitting yourself on the head
we would not be here.

That's right.

In other words those rewards that resulted
in our survival, behaviors that don't interest you, nevertheless
resulted in our human level intelligence when we acted to
maximize those rewards.

Yeah, except with my definition of intelligence, what you just wrote is
nonsense. And since I believe my definition of intelligence is in fact the
CORRECT definition of intelligence (the one which all of society will agree
on once AI is solved and once society as a whole understands how it was
solved and what it means about humans), I just see what you wrote as more
continued nonsense. It's nonsense that results from you not understanding
what intelligence is.

Our ability to CREATE NEW BEHAVIOR is the true measure of our
intelligence.

And if that new behavior is hitting yourself on the head because
it maximizes a reward signal you call the ability to create that
behaviour as intelligence. Yes I get your use of the word. But
if our behavior was hitting ourselves on the head, a behavior the
rest of us would call stupid behavior even if you call it an
example of intelligence because it was the result of maximizing
some reward, we would not be here. Maximizing that reward did
not result in human level intelligence even if you want to say
it was an example of intelligence.

Your argument is making the assumption that a beahvior is not intelligent
if it's goal is not survival. That's just stupid john. You are again
using folk psychology to argue a technical issue.

An AI can do science and math and philosophy and make great new discoveries
and solve problems no human has never been able to solve - all of it
motivated by goals that are NOT survival - but yet according to you, none
of it would be intelligent behavior because it was not motivated by
survival. This is inconsistent with the folk psychology definition of
intelligence because the folk psychology definition of intelligence simply
is stupid and inconsistent with itself and inconsistent with known facts of
reality - which is the norm for folk psychology concepts. But yet, here
you are again, playing another card from the folk psychology handbook in an
argument about AI.

Their intelligence is measured not by the problem they were
given to solve, but by the solution they created.

Which I understand you to mean, by how well that solution
maximizes a reward signal even if it is banging their heads
on a brick wall.

Our motivation system simply defines the problem we have in
front of us to solve ...

Which I understand you to mean is the problem of how to
maximize some reward signal even if that is how effectively
to bang our heads on a brick wall.

That's what makes it intelligent - the power to design
its own circuits.

Well I have no issue with that, taking into account what
you mean by intelligent. We mainly disagree on how it will
be done. For you it is a bag of units, for me it will involve
more innate structure (even if evolved or learned) as I
believe there are control issues not solvable by a bag of
units unless you use an evolutionary approach which will,
I believe, result in a structured network, maybe with some
physical modules such as found in the human brain.


In other words, you don't believe a Turing machine can be
programmed to solve all these different tasks because it
doesn't have enough "innate structure" correctly evolved
to solve the "word processor" problem, and the "display a
video" problem, and the "surf the internet" problem, and
the "play chess" problem?

You start of with "programmed to solve" and then you talk
about "evolved"? We don't evolve these modules in a word
processor program or a paint program or a spread***
program we program the modules as you well know.

I used the word "evolve" in that way to mimic you. You have talked about
how our human designs are "evolved". Maybe I misunderstood your point in
the past but I thought that's how you were using it (and I happen to agree
with that usage).

In a chess program there is that option to evolve modules.

So before such a simple machine can do all these very
different functions, it has to be given innate modules
to support each function?

How can you program a machine with nothing to program
it with? We program machines with modules. PRINT is a
module. SIN is a module. LINE is a module. Or if you
like you can go down to the hardware. MOV is a module.
INC is a module. JNC is a module. We use them to make
our own modules which we can then use in other modules.

I did not suggest it had NO modules. I suggested (to reflect your invalid
argument) that each high level function had to be mirrored by a hardware
module (which is just false).

The only thing wrong with connectionist models, is they
aren't good enough yet.

Well Pinker goes through some of the things they are not good
at so maybe you can solve them without "cheats". I suspect
the brain "cheated".

Well, that's a good question yet to be fully answered.

Minsky did the same thing in his famous paper with Papert. He showed what
single level perceptions could not do a simple XOR, and then that led
people to the invalid conclusion that ALL forms of confectionist modules
had such restrictions which led to a decade or more of the AI community
mostly ignoring the connectionist approach. Now, you are making the exact
same totally stupid argument by suggesting that Pinker has shown that
connectionist models can't do things without "cheating".

Pinker can't show the limitations of a network design he's never seen, or
thought about, any more than someone as smart as Minsky can.

If you have a bad network design, then you have to "cheat" by using
external help to make up for what what your bad network couldn't do.
That's true - and it's easy to see how that's been done with lots of
working examples of how neural networks have found applications (TD-Gammon
is even one example).

But the fact that current networks are limited, is not proof that there are
no network designs that might be strong.

It's fine for you to believe that the brain "cheats" and that any AI we
build will also have to "cheat" to make it work. That's just speculation
following your instincts. It's wrong to suggest that a better network
design doesn't exist simply becuase someone like Pinker has shown how
people have had to "cheat" to make the current connectionist systems do
something useful.

I'll also point out something I think is even more significant. No one has
solved AI using any of these "bad" network designs, with or without
"cheating". To suggest we can solve it by adding better "cheats" is
illogical to me. You solve it by throwing out the bad design, and creating
a better one, not by wasting time trying to add more duct tape to the bad
design.

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

Quantcast