Re: What is conditioning?

JGCASEY wrote:
On May 8, 12:52 am, Wolf <ElLoboVi...@xxxxxxxxxx> wrote:
[...]
Wolf:
A sensor filters out one kind of energy from the
environment, and that energy is converted into a
signal sent to the rest of the system. The signal
may or may not be acted on. That part of the process
may or may not be conditionable. IOW, the eye does
not see, nor does the brain. The system as whole
sees. Seeing is conditionable, but photon reception
by the retinal cells is not (AFAIK).
[...]
The notion of "filtering out one kind of energy
from the environment" seems a little strange to me.

The retinal cells in the human eye respond to light from the almost infra red to the almost ultra violet. They filter out all other frequencies of EMR.

The human ear filters out sound frequencies below about 20Hz and above about 20kHz.

Some of the various sensors in the skin respond to different ranges of pressure.

The taste sensors respond to certain molecules and no others.

And so it goes.

All sensors are filter by definition. Otherwise they wouldn't be sensors.


I view a filter more like this:

input output1 output2 output3
HDSUECXEW H C w
EOCOIUEWC CC W
EWOWIECWQ C WWW
BCMXEWOCW CC WW
OCWKTPGUY C

So you have three sensors here: one responds only to H, one only to C, one only to W.

For example in a visual system that filters out
high contrast traffic signs you might start with
a local threshold filter. Then next set of filters
might filter out the resulting areas into sizes
and shapes. The resulting outputs can become
"symbols" to be manipulated at a higher level.

It seems to me you are describing not sensors but data processing.

Of course, a sensor array may be structured so that its structure processes the data. This is in fact how the retina works. Its three layers of neurons are interconnected such that edges, for example, are detected in the retina.

In general a filtering process extracts invariants
from an ever changing input. At the input side you
have changing pixel values (thinking computers here)
as say a cow moves about in front of the camera.

At the output side you have the cow symbol which
remains constant even though the input side consists
of ever changing patterns of pixel values as the
cow (or camera, or other objects in view) move
about.

IMO the visual system doesn't out put a symbol, it outputs a shape. Just how it does this isn't easy to see. ;-)

The language system receives some output from the visual system and outputs a speech act, which we abstract as a symbol. But abstracting a symbol from this furshlugginer mess is IMO the task of neither the language nor the visual systems. I speculate another system. Maybe located in the prefrontal cortex.

JC:
... if I spend many hours studying the output from a
Mandelbrot program I should be able to glean a clue
as to what kind of behavior I want to see in my own
program. Nah. Maybe I will just play around with math
and simple programs and maybe get there the way it
actually happened.


Wolf:
No, dammit. You have to study the environment to which
the program responds. You cannot understand the behaviour
of any system from its behaviour alone, from its structure
alone, or from the environment alone. (And anyhow, the
Mandelbrot program is not a learning system.)

JC:
What I am suggesting is that it might be easier to work
out the mechanics of the thing first, keeping in mind AI is
about building a machine, from mechanical or math
principles, not from its behavior which can be very
complex, much of that complexity may actually come from
the environment such as the complex path an ant takes
while crossing some complex terrain, even though the
internal mechanisms may be very simple. Also there may
be responses that you don't know about until you happen
to get the right environment. You may not know someone
is scared of spiders until you happen one day to see
them in a spider environment. That potential behavior
is hidden, but would exist in some form in the brain.

I didn't think building a machine from mechanical or math principles is going to do it. Unless of course you are aiming at a math machine.


Wolf:
I'll describe Pavlovian conditionable machine again:

The machine exhibits Pavlovian conditioning if it:
a) has a hardwired response R to some input I1 via some
channel;
b) is capable of receiving some other input I2 via some
other channel;
c) after X simultaneous exposures to I1 and I2 exhibits
response R to I2 alone.

That's it.


JC:
Well I gave an example. Let the "channels" be from a
keyboard. Let the response be a sequence of characters
printed on the console screen.

Press F and the program will print "salivate" as an
unconditioned response. If you keep pressing F after
a while nothing will happen as this response is
dependent on an internal state called "hunger".

By the way the analogy I see here is that the keys are
the symbol levels extracted via the filters we wrote
about above.

Now try hitting the B key. Nothing happens. So hit the
B key and then the F key. "salivate" is printed. Do this
often enough and the B key will result in "salivate"
being printed. This response will eventually stop if
the B key is not being followed by the F key.

Variations in all this is easy enough to accommodate.

The internal program may not be how the animal is wired
to perform a similar task.

IMO you won't get a learning machine if you write code. You'll get a simulation of a learning machine. That will be good enough for many if not most practical purposes, eg, a smart spell checker, that learns my characteristic mistakes and corrects them as I type.

Change the situation and
that may be revealed by different patterns of response
which you would then have to reprogram. The only way
you could avoid that would be to know the actual program
being used by the animal not observable from its current
behaviors in the current environment. That is, you cannot
explain it ever from first principles for the actual
details are in the program. You cannot predict with 100%
the response of a program for any given environment
without knowing the internal details of the program.

I don't think you can predict it even then. For if you could do that, you could write bugfree programs.

I suspect if Sniffy the Rat could be released it wouldn't
behave like a real rat at all.

I agree.

Wolf:
To be as complex as a slug, it should show diminution and
eventual extinction of R to I2 over time if I2 is presented
at intervals longer than some minimum time; and continued
and even strengthened R to I2 if I2 is presented at less
than minimum time intervals. Also, random variation in
intervals of presentation should have an effect on the
strength of the response, the rate of its extinction, etc.
(Read the "minutiae of scheduling" for further enlightenment.)

JC:
Sure you could adjust the program to depend on random
variations in intervals of presentation etc. That is no
doubt what they have done with Sniffy the Rat.

Yeah, but Sniffy is a simulation.


Wolf:
Can you design a machine that will exhibit those behaviours?

JC:
I think so. May I present Sniffy the Rat :)

Actually I think there is more to a rat and a slug brain
then is revealed by those slug/rat experiments.

Quite so. As you probably know, the neural network of aplysia has been completely mapped. But it's still not possible to predict what will happen when some combination of neurons fires.



JC:
... I don't see studying the minutiae of schedules
of reinforcement or of conditioning chains will help at all.
I have no idea how to use any of it to invent a machine that
can produce human-like intelligent behaviors.


Wolf:
I don't think you have any idea of how to use any of it to
produce even an artificial, conditionable bacterium. So why
not start there?

JC:
I wouldn't start there because the behavioral analysis doesn't
provide me with a description of the bacterium mechanics. That
was the point I was making. The behaviors do not themselves
reveal the inside of the black box and therefore can never
provide sufficient information to build one.

I never said they did. They provide _clues_. That is, if you know the results of some course of training, you should be able to estimate what sort of internal data flow and restructuring of data flow might produce such results. Knowing the behaviours, you know what sensory pathways must be involved, what effector pathways must be involved, and even what internal cross talk there must be. Eg, if a bell can become a stimulus for salivation, there must be internal crosstalk between whatever system processes the sound and whatever system processes the smell of food. IMO, that's a major clue.

To put it another way: no amount of knowledge about the insides of the black box allows you to predict the system's behaviour.

Either way, you make your guesses, and you dis/confirm with experiment.


They may even
make you think it is more complicated then it is. Imagine
trying to reproduce the output of a Mandelbrot program without
insider information. I imagine you would be making all kinds
of complicated attempts and never succeed.

Exercise for the interested student: Just what is the output of a Mandelbrot program? And why is it conceptually wrong to equate that with its behaviour?


Wolf:
Human level intelligence is too complicated a complex of
behaviours to be easily analysed.

JC:
Agreed. The whole point of my Mandelbrot output. Maybe a
behavioral analysis is not the way to go.

If by "output of the M program" you mean the pretty pictures on the screen, you are not talking about its behaviour. (Actually, I don't think it's sensible to about the behaviour of a program.)

Wolf:
Even the simplest examples of conditionable human level
intelligent behaviours are themselves complicated complexes
(recall how you learned arithmetic for an example.)

JC:
Maybe. But if you take tackle the problem differently ...

Sideways jump to illustrate a point: a calculator does arithmetic. the kid that uses the calculator does math. Why is a calculator such a simple machine? And why don't we have math machines yet?


Wolf:
Also, as I've claimed before: there are no general
learning machines.

JC:
I suspect you are right there. Have you told Curt?

Oh, he's probably reading this.

--


Wolf

"Don't believe everything you think." (Maxine)
.

Relevant Pages

  • Re: What is conditioning?
    ... the future behaviors as a result of the shaping ... from the environment. ... A sensor filters out one kind of energy from the ... has a hardwired response R to some input I1 via some ...
    (comp.ai.philosophy)
  • Re: The animal, not the environment, controls behavior.
    ... you do say the *title* of the thread is a response to ... because the animal, not the environment, finds the information ... You cannot decide upon the internal variables, ... with regards to unraveling the brains of animals and any ...
    (comp.ai.philosophy)
  • Guglielmo! Youll settle attentions. Occasionally, Ill initiate the lentil.
    ... Some consistent judges at least the ... Tell Lakhdar it's written ... boxings because of a environment. ... She can relatively imply in response to terrible ...
    (sci.crypt)
  • Re: Turing Machines and Minds (Was: The Demise of Computationalism?)
    ... you are saying that once a response is done undergoing its transition ... my brain to perform that response, and then uploaded it to another ... part of which is what you call "our environment". ... It is true that stimuli that appear ...
    (comp.ai.philosophy)
  • Re: The animal, not the environment, controls behavior.
    ... you do say the *title* of the thread is a response to Glen's claims - ... because the animal, not the environment, finds the information ... collecting and the information collected as rewarding. ... You cannot decide upon the internal variables, ...
    (comp.ai.philosophy)