Re: Is AI all about time?

casey <jgkjcasey@xxxxxxxxxxxx> wrote:
On Jul 9, 9:56=A0pm, c...@xxxxxxxx (Curt Welch) wrote:
Skinner declared that human behavior (including our
language behavior) is a product of operant conditioning
before I was even born. Read his work John - don't take
my word for it.

Skinner could have declared the moon was made of green
cheese without making it true. I read Skinner's book,
Beyond Freedom and Dignity, as a teenager.

And you should not be confused when I say "something is ..."
that this is a statement of my belief, not a statement of fact.

But then you respond by calling anyone who doesn't share your
beliefs an idiot.

Because they are! :)

That's just arrogance John. It has nothing to do with what we are
discussing or with wither I'm right or wrong. It's just my personality.
You should well understand that by now. (I was conditioned to talk like
that) :)

Why should they be called idiots if your
statements are beliefs not facts? Is it a conditioned reflex?

It's a fact to me John. Just because the rest of the world doesn't
understand the truth yet doesn't mean it's not a truth.

In general, in life, I tend to be understated. I don't tend to suspect I
know the truth most the time. But when it comes to AI, I happen to have
very strong beliefs about what is true - and I share them here with vigor.

When I wrote above "it's just by belief" I didn't mean "it's my guess" or
"it's my best guess", or "it's my current best speculation", I meant what I
wrote. It's my BELIEF - I believe it to be true.

Because much of what I believe in this area is not a widely shared belief,
anyone that didn't have their own beliefs would be wise to heavily discount
what I say - simply because I am standing (mostly) alone in these beliefs.
Odds are, anyone that is out in left field by themselves, are out there for
the reason they are idiots. But most great thinkers were out in left field
by themselves at some point as well, so I'm a great thinking on this
subject or an idiot? You can't tell by reading what I write. Only time
will tell.

Obviously however, I don't think I'm an idiot in these areas. But neither
do any of the idiots. :)

Our high level reasoning is probably an example of
exaptation where some structure adapted for one use
turns out to be suited for other uses if elaborated.


That's absurd.

Why? Because you think it can't happen or because you
think it didn't happen?

Beats me. I'm not sure why I said that there. I think maybe you cut out
too much context in your quote? Or maybe I was thinking you said something
different then you did? I'm not sure what I was thinking when I wrote
that.

However, I'll add another comment about what you wrote (which might have
been connected to what I was thinking when I write it was absurd).

You seem to believe "high level thinking" is yet another type of behavior
behavior, created by yet another different module. I don't. I think what
you call "high level thinking" is just more of the same. It's just yet
another learned behavior that has been shown to be useful at producing
rewards for us.

I don't think "high level thinking" is any different than eating a
sandwich. It's just a sequence of body behaviors that tends to produce
higher rewards than other sequences of body behaviors at that point in our
life so that's why we do that behavior, at that time, instead of some other
behavior.

Calling something absurd doesn't make it so.

It does when I write it. :)

Did the first neural network perform high level reasoning?
If not where did this high level reasoning network come
from except as a mutant form of a low level reasoning net?

I reject your reality and substitute my own at this point.

The "high level thinking" network is NOT A DIFFERENT TYPE OF NETWORK from
the "low level reasoning net". Saying it is doesn't make it so. :)

We could say the low level network did RL and from this
evolved the high level RL network.

http://en.wikipedia.org/wiki/Exaptation

But I understand you like to think that way. Every
behavior has a module to explain it - that's just how
you like to think about the brain.

Every behavior *doesn't* have a physical module to explain it.

In a post to Tim Tyler you wrote:

"I'm not the one here that thinks we have to develop a new
module for every behavior."

Well that is not exactly what I was suggesting.

Yeah, of course I know you don't think that EVERY behavior is a different
module.

Most if not all
high level behaviors, such as language, I would see as the action
of many modules not one single language module. The "modules"
are functional in nature and there is not a single module for
each and every problem. You seem to be suggesting I am saying
that 1+1=3D2 and 1+2=3D3 and 1+3=3D4 exist as separate modules and
I am not saying that at all. There is ONE arithmetic module and
that is a functional module where its parts are used for doing
more than just arithmetic.

I'm just saying you like to believe there are lots of different types of
modules at work creating our intelligence (and that any AI close to human
performance must also use many different types of modules) which is in
stark contrast to my belief that intelligence can be created by only one
type of module.

I am fully aware of the fact we tend to look for preconceived
black boxes performing the functions. But these functions are
derived from our subjective or objective observations and in
fact probably involve shared hardware which is distributed all
over the brain and whose real functions may be transparent to
our observations.

You also wrote:

"I think it will be possible to build strong generic learning
systems, which works well on a very wide range of behavior
learning tasks - so all you have to do is give it one hard
problem to learn, and if it can do that, then it can learn
many hard problems without creating simulations for each."

It is clearly possible for a system to solve a wide range of
problems, the brain being the best example so far. But problems
fall into categories, which the brain can recognize, and thus,
if faced with a "new" problem, it can bring solutions from
similar problems to solve it.

You put emphasis on the temporal part of a problem such as
catching a ball but the higher level problem solving in a
human is not the catching of a ball, which a dog can do, but
rather understanding the physics of a thrown ball.

We don't simply see the moving ball we see the STATIC path
the ball takes and describe its STATIC shape. When we look
at things we look for what doesn't change. Velocity is an
UNCHANGING distance travelled per unit of time. Acceleration
is an UNCHANGING velocity.

Do you actually take the time to think when you write this stuff?

You don understand that a non zero acceleration means the velocity is
CONSTANTLY CHANGING right?

Even time itself we freeze into
a static spatial time line.

Again, you are so damn focused on what happens with static written
communications you can't seem to think any other way.

The only time any of those things "freeze" is WHEN WE WRITE THEM DOWN ON
PAPER and then "THINK ABOUT" what we have written. None of the things you
mention actually ever "freeze" in the universe. It's impossible to keep
the position of anything constant. It's impossible to keep the velocity of
anything constant. It's impossible to keep the acceleration of anything
constant.

Reinforcement is the mechanism of selection not the
mechanism that generates things to select.

ALL REINFORCEMENT LEARNING PROCESSES INCLUDE BEHAVIOR
GENERATION ALGORITHMS.

To suggest they don't shows, AGAIN, you don't have a
clue what you are talking about. Show me one, any one,
reinforcement learning algorithm that doesn't also
include the behavior generation algorithm to back up
your suggestion above.

Read it again. I wrote *reinforcement* is the mechanism
of selection. For example it refers to the increase or
decrease in your gap values (or ratios). I did not write
*reinforcement learning* which of course involves a
mechanism to produce things to reinforce (select).

You can't have one without the other John. Again, you are always so
focused on words, and so seldom focused on physical reality. If the words
make sense to you, you think you have said something valid, even if it
makes no sense in our physical reality. In langauge, you can say
"reinforcement" and pretend it has nothing to do with "reinforcement
learning". They are after all, different words. But in physical reality,
you can't have reinforcement separate from reinforcement learning. They
are one and the same machine you are talking about.

Because evolution lacked a designer to build neural
networks it had to make random networks.


That really makes no sense John. If it lacked a designer,
how the hell did it design a network (random or otherwise)?
Are you suggesting the neurons weren't designed?


Evolution is a process of reinforcement learning and as
such, contains the exact same type of process that makes
us intelligent - the same underlying process that gives
us the ability to design.


Evolution did not "lack a designer" and so created "random
nets". It very much designed the networks it designed
because they worked well at the task at hand.

What planet are you from Curt? Is it really that hard to
understand what I have written? Are you deliberately trying
to make what I write sound silly by a silly interpretation?

Does Dawkins make no sense when he talks about the illusion
of design? Is it so hard for you to see I am saying there
was no one deciding that this net would do such and such
and deliberately building it for that purpose?

If you point was there was no HUMAN designing the net, then you point is
just idiotic instead of stupid.

Evolution did not create "random nets" BECAUSE THERE WAS NO HUMAN HELPING
IT.

It just so happens that such networks can perform useful
actions. Thus those dna sequences that coded for the nets
that enhanced the reproductive success of the individual
were selected (reinforced) resulting in the neural
circuitry we see in today's brains.


Right, it was designed by an act of intelligence - an act
of reinforcement learning.

So you are seeing RL as the designer?

Of course.

You can label it that
way if you like but it doesn't justify your attack on those
that don't talk that way. You seem to want to deliberately
misunderstand or misrepresent what I mean.

I still don't see what you wrote as being intelligence no matter how I
interpret it. I still don't get the point you were trying to communicate
by saying:

>> Because evolution lacked a designer to build neural
>> networks it had to make random networks.

Evolution didn't make "random nets" so I don't get your point as to why you
think it "had to" do something it didn't even do.

An innate prediction system based on lessons, learned
by evolution, about how the real world operates.


If A tends to show up at the same time as B, and B seems
to show up at the same time as C, then we can assume,
by association, that A is likely to show up at the same
time as C, even when we have never seen it happen.

The point I was making of course was it is an innate skill
by a network not something it had to learn in real time.

Yes, and the counter point I was making is there is no evidence it is (or
needs to be) innate.

In the visual networks two inputs say x and X (small
and large) might be categorized as the same. However
it is important that the difference is also kept as
if the x was a predator a small version of the
pattern would mean predator far away and a bigger
version would mean predator too close! In other words
it would be most likely these evolved networks would
incorporate the symmetry of the real world so they
had this knowledge built in for immediate use rather
than every individual having to learn it all again.


Oh, so you are now going back to the "learning sucks,
so obviously evolution wouldn't overuse it theory".

I agreed with you that what we call high level human
behavior is learned not innate. Where we differ is I
see it being learned within an innate framework using
innate skills we share with other animals. I have never
said learning sucks. Unlike most animals with a cortex
we rely on the cortex for all our learning to survive.

When I read my first AI books I remember thinking how
interesting a learning system was compared with one
with fixed behaviors. I even remember thinking along
similar lines to you only my analogy was water shaping
and being shaped by the landscape. But I don't have a
narrow view on the subject.

Well, I've told you exactly how the generic network
will work (what it needs to do),

How it works and what it needs to do are not the same thing.

Sure, and if I solve AI following my ideas, I could then show it to you.
But you know we are not there, and you know nothing I'm saying is based on
the belief we have solved it. It's all my interpretation of the
circumstantial evidence vs yours.

... and I can show you by example of how it works for
everything you bring up -

What did I bring up that you have shown it works for?

That should better have been worded "show how it could work in theory".
And I've done that for probably everything you have brought up as "things
that needs to be done by a special type of brain module.

... but yet that's not good enough for you because you
don't really understand it -

Or maybe they just don't agree with it?

... so you are just waiting for me to figure it out, and
then show it to you.

If I have to figure it out alone, I'm not likely going to
show it to you John. I'm going to take it to the bank instead.

I actually recommended, in another post, you do that .

I've been here for many years sharing my every view and
idea hoping to find people that can understand what I'm
talking about

Me too :)

... and provide some help, and so far, not only have I gotten
almost no help,

Mmmm. I spent a lot of time playing with the pulse sorting net
you suggested but couldn't see how to make it do anything that
couldn't be done simpler with neural-like units. How much time
have you spent helping me with my neural net projects? Waste
of your time no doubt? Perhaps that is how others feel about
your pulse sorting networks?

Of course it is. That's exactly why most people don't do much more than
spend 10 seconds thinking about it. You clearly have done far more, but as
far as "help" with the design of the system, you have provided none. You
have have provided almost infinite help as a sounding board, which is of
value, and pointed me to many references or facts related to AI or brains I
did not know about - which is of value - but in terms of helping me design
a better learning algorithm - no help at all. Almost all my time working
with you on these learning algorithms are attempts to get you up to speed
so you can one day get to the point of actually helping. I doubt that will
happen however - not because you don't have the skill - but because you
don't have the "faith". That is, you don't believe what I believe, and as
such, it keeps you from being able to understand, and work on, the problem
I'm working on.

You wonder how to make high power reinforcement learning.
Well maybe that comes from a neural network capable of forming
a hierarchy of modules with neuron like units rather than a
spaghetti load of "temporal" units.

Standard human neurons are VERY MUCH temporal units so I'm not sure what
you suggesting there.

I've not even found people that can understand what I'm
talking about for the most part.

Didn't understand or didn't agree with?

Didn't understand because you didn't agree with the principles. That is,
your beliefs seem to actually keep you from unstacking much of what I say
when it comes down to the finer points of the theory behind solving AI by
building a strong generic learning algorithm.

That's fine, I've had lots of good conversations and debates
and expect to have lots more - and I've learned a lot about
the philosophy of AI - but have learned basically nothing
here to help me do what I came here to do - build better
learning networks.

Well I have on occasion thought about and programmed ideas
for learning networks using neuron-like units

You used the phrase "neuron-like units" multiple times now. Do you mean
"multiple inputs one out spatial function artificial neurons" we find in
tyical computer sci back prop trained neural networks, or you saying
"neuron-like" simulation models that are actual temporal functions?

but you are
only interested in people who can help you get your pulse
sorting "temporal" unit idea to work.

If you showed me a reinforcement trained neural network that solved the
temporal pattern matching problem I wouldn't care whether it was pulse
sorting or not.

The pulse sorting nets I'm playing with are like a 9 on a scale of 1 to 10
where 10 is the solution to AI. The things I've seen you talk about, or
play with, are about a 2 on that scale.

If you showed me a 9, or a 9.5, or maybe even an 8, that didn't use pulse
sorting, I'd be interested in it. But you have shown me nothing higher
than a 2 yet. I've not been interested in it because it was a 2 - not
because it was not a pulse sorting net.

We also have a different view on what causes intelligent
behaviors. You see it in the singular I see it in the plural.
I see problems and solutions falling into categories even
if learning is involved in all cases. We can form more of
those categories due to our large cortex and in particular
the selective power of the frontal cortex.

JC

All behavior is a category reaction to a stimulus John. That just means
behavior is context sensitive. The fact that our behavior is context
sensitive does not mean we need different modules to support each class of
behaviors we can produce.

You do understand my design is actually far more "modular" than yours
right? My network, which has one type of module, (the network node), will
have 1 million modules if it's got a million nodes, and every one of the
nodes, will be performing a unique, and different purpose. Every module
will play a unique role in every behavior it is part of. You get that
right?

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

Relevant Pages

  • Re: motor emulation, modularity, feedback, prediction, and dreams
    ... > You say it is "not at all hard, for a learning system, to ... The issue is that when you train a network to respond correctly - such as ... this new design does with ease. ... one output using only reinforcement learning. ...
    (comp.ai.philosophy)
  • Re: Is the Curt net a kind of decision tree?
    ... are classified the same way without further learning. ... These systems create invariant detectors based on the invariant effects it ... The network won't just naturally do ... of trying to find a workable network design. ...
    (comp.ai.philosophy)
  • Re: Is AI all about time?
    ... Did the first neural network perform high level reasoning? ... "I think it will be possible to build strong generic learning ... ALL REINFORCEMENT LEARNING PROCESSES INCLUDE BEHAVIOR ... Evolution did not "lack a designer" and so created "random ...
    (comp.ai.philosophy)
  • Re: What is innate and how can we determine it?
    ... learning", and "Learning is all that is important", ... It doesn't confuse me, now, but why would you want to confuse ... of reinforcement learning. ... network already answered all those questions. ...
    (comp.ai.philosophy)
  • Re: Cool visual illusion
    ... >> performing a different function in the network. ... my current network design is not ... >> is a real time, temporal, high dimension, reinforcement ... Everyone would agree that learning is a big part of human behavior. ...
    (comp.ai.philosophy)
Loading