Re: Finding useful functions- part 1

From: Wolf Kirchmeir (wwolfkir_at_sympatico.ca)
Date: 10/29/04 

Date: Fri, 29 Oct 2004 11:12:22 -0400

Bill Modlin wrote:
[...] snip quibbles about alleged quibbles, which were rather attempts
at clarifying the issues/processes]

> Was your substitution of "external" for "behavioral"
> significant?. "External contingencies", in the sense of
> statistical regularities suggesting the presence of a
> coherent object, are indeed initiators for most of the
> changes of interest.

At what level? Why statistical regularities? AFAIK, it's the
organisation (structure, topology) of the network of NNs that "indicates
the presence of a coherent object." "Statistical regularities" are
relevant when computing "coherent objects" in a bit-mapped image, for
example, but the retina + visual cortex appears to do it differently. Or
so I understand from my reading on the subject.

> But the behavior of the organism
> enters only incidentally into the causal chain leading to
> the signalling of those regularities to the neurons and
> their adaptation to them.

So do all inputs, at all levels, so your use of "incidentally" seems to
me an error. IOW, you might just as well characterise the inputs from a
neighbouring NN as "incidental."

> The same changes would ensue if
> a similar series of stimulii were observed passively in a
> moving scene rather than scanned behaviorally from a static
> scene. Behavior itself does not enter into the relevant
> contingencies.

If you think "behaviour" is limited to skeletal-muscular events, think
again. IMO this misconception on your part is one of the reasons for
some of your errors. Also, there is no such thing as "passive
observation." Observation (perception) is a behaviour, and a very
complex one. So complex, it's not fully understood yet. (At one end, the
"conscious awareness" end, it may never be fully understood, only
described, much as quantum mechanics is not fully understood, only
described. This analogy may account for Penrose's speculations.)

>>>Let me try to pose an unambiguous example of the
>>>conflict.
>>>
>>>A pigeon can be trained to discriminate pictures
>>>containing trucks from other pictures lacking trucks.
[snip for brevity]
>>I see no reason to talk about "supervised" learning
>>processes, since that word smuggles in the experimenter's
>>intentions. The pigeon will leran in exactly the same in
>>naturem, the only difference being that random behaviours
>>will be reinforced rather trhan pre-selected ones. So
>>what?
>
>
> Obviously you are not aware of the ordinary technical usage
> of the word "supervised" in this context. Go read a book
> or something, I'm getting tired of trying to educate you.
> Hint: it is still just as "supervised" when natural forces
> do the supervision.

Firstly, that's not "ordinary technical usage", it's AI usage,
apparently. So why use the word "supervised"? I've been enlightened
about the meaning of "supervised" by Stephen Harris - I think it's a
stupid term. It assumes that there is some other non-supervised form of
learning, which in turn assumes that learning is something other than a
change in behaviour. If so, just what is it? Are you claiming that
learning "really" is something other than a change in behaviour, and
that changes in behaviour are only "incidentally" signs that learning
has taken place? If so, you have relocated learning somewhere else, and
should use a different term. I vote for "reconfiguration of neural
networks" I don't have enough Latin or Greek to make up a nice technical
word to replace that phrase, though - sorry about that.

>>The mechnaisms that "adjust the pigeon's behaviour"
>>include the cellular changes that you seem to think
>>exem[pligy some other kind of learning.
>>
>>
>>>That's fine, so far as it goes.
>>>
>>>But when I look at discriminating that class of pictures
>>>so that it can be recognized as a condition for the
>>>rewarded behavior, I see a pretty complicated process.
>>>There are billions of cells computing functions of
>>>whatever inputs they have access to, responding to
>>>all sorts of "features" at dozens of levels, bringing
>>>together information from many areas of the picture, to
>>>eventually reach a level at which there is a signal of
>>>some sort that indicates whether or not there is a truck
>>>somewhere in the picture.
>>
>>
>>So the process is complicated. So what? When I watcvh a
>>rainstorm, I see billions of raindrops, millions of
>>turbulence scells, etc. The proces seems pretty
>>complicated. The net result is still that tings get
>>very wet.
>
>
> An analogous situation would be if you claimed that "things
> get very wet" is relevant to the path of a raindrop. The
> path is part of a big picture in which all those paths
> add up to a net result of things getting wet. But you can
> hardly explain the details of a particular path by saying
> that a drop moved a particular way just to get things wet.

Sorry, I was refuting your assumption that that is what I was claiming.
I don't and didn't.

You're also assuming intentions and purposes. I guess you are assuming
that learning must be for some purpose. It isn't. Learning is just a
change in behaviour.

[snip brief summary pf pigeon pecking trials]
>>>That truck-signal is correlated with the rewards and the
>>>behavior, so it makes sense at least at a handwaving
>>>level that a supervised learning process could
>>>incorporate it, and produce the behavioral
>
>>>But most of those intermediate signals in the long path
>>>from retina to truck-signal are not correllated with
>>>anything in the high level description of the
>>>experiment. They aren't correlated with trucks,
>>>or rewards, or pecking, and therefore could not have
>>>been shaped by any of those things.
>>
>>Yes, that's true, but why should they be?
>
>
> They should not and are not. But the behaviorist viewpoint
> Glen espouses entails that contingencies among such
> externally describable and observable things are all we are
> allowed to invoke in explanation of the shaping of behavior.

You are either misunderstanding or misrepresenting Glen's p.o.v., which
is that "Physiology mediates behaviour", and "Physiology is not needed
to account for relationships between contingencies and behaviour." I
would go further: adding physiology to account for those relationships
is likely to obfuscate rather than enlighten. The question of what
physiological changes are necessary for changes in behaviour to occur is
at a different level of explanation than the relationships between
contingencies and behaviour. Some generalisations may be possible along
the lines of "When a human reaches puberty, a number of synapses are
destroyed and new ones formed, hence the addition of courtship
behaviours to the human repertoire." But one would still have to account
for actual courtship behaviour in terms of the contingencies encountered
by Jim and Joan (which include the culture in which they are raised, etc
etc etc etc). There's no other way. As you yourself complain,
information is changed so much at the NN level that we can't correlate
those functions with a glimpse of Joan's or Jim's face on the other side
of the room.

To put it another way: if you want to know why some specific behaviour
occurs, or why it is shaped in some specific way, all that's available
to explain that are the contingencies (and the histories of the
individuals, but let's not complicate matters further). If you want to
know why some _class_ of behaviours is possible at all (and hence it is
possible to shape them in some way), then physiology (including a lot
more than NNs) is necessary. If you want to know how to make a machine
that will learn with experimental "supervision" or without, then you
must study both levels of organisation. At least.

As for "explanation" - it's not clear to me what you understand by that
term.

>>ASn analogous problem: how do the hundreds or thousands
>>of fish in a school of fish all "know how to change
>>direction? They don't. Each fish knows that the immediately
>>surrounding [snip for brevity]
>
> Ah! Finally a clue as to how you can believe things that
> appear so silly to me. Perhaps this explains Glen's
> misperception as well. Thank you, thank you, thank you!
>
> The problem with this analogy is that the coupling functions
> between neighboring fish in a school are linear and thus
> correlation-preserving, while the coupling between
> neighboring neurons in a network is nonlinear and
> correlation-changing.

So? My comments allow for that. The principle is the same whether the
system's (school of fish, NN) behaviour changes (is different the next
time it encounters a shark) or not. It would hold no matter what the
form of the correlations: local action results in global behaviour.
Granted, non-linera functions tend to be counter-intuitive - that;s why
I used school of fish. It's easier to see the principle operating in
such a system. But surely you should have no trouble scaling it up
another level of complexity.

[snip explanation of linear and non-linear functions, and how systems
instantiate them]
> In general, there is no correlation between inputs and
> outputs of a multi-stage nonlinear multivariate transform.

Quite so, if you mean it's impossible to predict the output as a
function of the input alone, but that it is necessary to compute all the
intermediate steps to determine the output. I understand all that. What
I don't understand is why you think this refutes the behaviorist claim.
I also don't understand why the output of a non-linear function is not a
correlation, while a linear one is, but I'm probably just using
"correlation" to mean " computed/able relationship between two or more
variables" or some such - not ordinary technical usage, I guess.

> For a deep multilayer network, interior variables will
> generally have no significant correllation with either
> inputs or outputs of the network as a whole, but only with a
> subset of the interior hidden variables generated at similar
> levels of indirectness or abstraction.
>
> In such networks there are no spreading ripples conveying
> the same message throughout the network. At each node there
> is a new message, conveying new and different information.
> Some of these may be correlated, most are not. Only
> correlated signals are useful for the directed modification
> of the function to be performed by any particular node.

Quite so. I understand all that, perhaps even more intuitively than you
do. But I still don't understand why you think it refutes or sidelines
the behaviorist claim.

> -----
>
> That should be enough to make the original point that
> externally identifiable factors such as behavior or
> contingent results of that behavior are generally
> uncorrelated with the proper function of interior nodes of a
> complex nonlinear network, even though those interior nodes
> are ultimately part of the causal chain leading to the
> behavior.

Again, your usage of un/correlated is confusing to me. I think you just
mean "there is no simple correlation between behaviours and
contingencies on the one hand and the functions of neural networks on
the other." True, true.

> Which means that behavior-based explanations are irrelevant
> to the appropriate adaptation of those interior functions.

I thought that's what I said, and have said several times. I'll say it
again: The motor cortex NNs that initiate and control the pigeon's
pecking at a red button "know" nothing at all about the colour red, nor
do they know anything about a truck. Neither do they "know" about the
feedback between the visual cortex and the motor cortex that guides the
peck to nearly the centre of the button every time. Etc.

So what? Just what is your point? Why are you so het up about
eliminating behaviour from an attempt to figure out how to make a
learning machine? Granted, you don't need to know what specific
behaviour produces/triggers the changes in the NN, but you sure as hell
need to know about it when you test your machine.

> The next step is to introduce relevant explanations of how
> interior cells come to perform useful functions, which is
> what I intended for "part 2" of this topic. It may be
> obvious by now that first we need ways to collect correlated
> signals together, and then we need ways to generate useful
> new outputs from those correlated signals.
>
> Part 2, a sketch of possible approaches to finding and
> exploiting correlations among interior hidden variables,
> should be coming soon.
>
> Right now it's hard to see for sure what will be in part 3.
> I'd guess that the next topic should probably be about how
> we bridge from unsupervised data-directed self organization
> to goal directed behavior...

Hey, what's with this "unsupervised" here? Are you referring to
spontaneous changes within the NN? If so, what do does "data-directed"
mean? Aren't inputs that originate in, uh, behaviours a kind of data?
Aren't the behaviours themselves a kind of data? Or are you talking
about data structures/arrays/whatever, stored in some device outside the
NN, that the NN "learns"? If so, that would look like "supervised"
learning to me - the "training signal" would be a match between the NNs
output and the data structure, wouldn't it?

Once again, terminological fuzz. Sigh.

>how we actually implement
> reinforcement learning or operant conditioning on top of this
> substructure. But I'll have to see how the conversation
> goes...

IMO, the "useful functions" that you refer to are probably those
"reinforcements" that occur in operant conditioning. But it's too early
to tell, since I'm not sure what you would class as useful functions.

I also suspect that behavioral neurologists (you know, the guys and gals
that try to correlate changes in neural networks in the brain with
behavioral changes of the organism) will arrive at some useful stage
sooner than the effort you're exemplifying. But what do I know. It's
been years since I did any coding, let alone any program design, and
that was pretty simple stuff, just about using data to learn things I
needed to know about a student.

Relevant Pages

  • Re: Sarah Palin- creationist VP candidate?
    ... Correlation does not imply causality. ... that makes learning real mathematics impossible. ... I would argue that the responses to your suggestion of change are only ...
    (talk.origins)
  • Re: Sarah Palin- creationist VP candidate?
    ... you can improve understanding and learning for any reasonable portion ... Correlation does not imply causality. ... You have yet to provide any solid evidence at all that your approach ... that makes learning real mathematics impossible. ...
    (talk.origins)
  • Re: Neural Network (choose of the target ?)
    ... Then I create a new network, I clcik on train, I have to select my inputs and targets but I can just select one for each, finally I click on "Train network" and I have this: ... I recommend first checking the all variable correlation ... performance of the constant model ...
    (comp.soft-sys.matlab)
  • Re: Can Excel reveal factors for success?
    ... statistics will help you. ... > on 6 of the 7 days that I both ate bananas AND got eight hours sleep, ... Thus the highest percentage or correlation for successful ... > Always Learning ...
    (microsoft.public.excel.misc)
  • RE: amount of alarms generated by IDS
    ... there is no "magic product" that will "do it all", most in the IDS ... worrying about Joe Blow scripter scanning the outside of your network, ... automatically log him has a future correlation event, ... Just given a set of FP on one port, ...
    (Focus-IDS)