Re: Finding useful functions- part 1

From: Wolf Kirchmeir (wwolfkir_at_sympatico.ca)
Date: 10/26/04 

Date: Tue, 26 Oct 2004 12:45:52 -0400

Bill Modlin wrote:
[...]
> Glen on the other hand actually disagrees with me.
>
> He claims that learning occurs only by adjusting a behavior
> mediating function in response to consequences of its
> behavioral output. An organism does something, that has
> contingent results, and it is those contingencies that
> control learning. There is no place in his view for
> learning from mere correlations among environmental inputs
> independent of behavior, there must be some sense in which
> the controlling inputs are viewable as contingent on a
> behavior that is to be modified.
>
> I believe that he is mistaken, that adjustments in
> function induced by correlations among signals not
> contingent on behavior play an important role in
> behavioral adaptation. My posting was intended to be
> a counter to his position, a listing of several places
> where functions are developed without reliance on
> behavioral contingencies, which functions nevertheless
> have critical bearing on subsequent behavior.
>
> You seem to be taking the same stance as I am. You
> speak of developmental processes that must occur for
> a child to be ready to learn any particular lesson just
> as I speak of developmental processes which must occur
> before a system is able to detect and learn from
> behavioral contingencies.
>
> You seem to agree that these developments include the
> same ones I cited as examples. Yet you frame your post
> as one of disagreement. Can you explain this?
>
> Bill

Well, actually I do disagree with you. You refer to signals not
contingent on behaviour. I see nothing but signals contingent on
behaviour. That's the point of my reference to the kitten experiment.
The VC of the kitten can learn only from the signals delivered by the
eyes - and if none of those signals correlate with vertical edges, the
kitten will not learn to perceive vertical edges. If "local
correlations" were sufficient, the kitten would perceive vertical edges
no matter what it's early visual experience. At any rate, that's how I
interpret "local correlations": you invoke them to explain learning in
the absence of external contingencies. But in so doing you assume that
such learning actually happens. AFAIK, it doesn't. "Development" is as
dependent on external contingencies as "learning" is - that's why I'm
inclined to think it's pointless to make a distinction.

It's apparent that "behaviour" to you means only "gross behaviour", that
of the organism as a whole. I include anything outside the neural net as
behaviour. Eg, your eye moves, the image on its retina changes, the
signals from the retina arrive at the VC, and the VC "develops" -- that
is, it learns to distinguish between vertical and horizontal edges, for
example. (NB that the retinal cells fire only if the light falling on
them changes -0 that's why spontaneous eye movement is necessary.)

It occurs to me that by "local correlations" you may mean something like
the bundling of retinal receptors, so that their signals "mean" and edge
moving vertically, diagonally, etc as the eye moves. If so, that's still
signals contingent on behaviour - the behaviour of the eye, in this
case. It's also a local correlation removed from the VC.

However, when it comes to engineering a learning machine, the
distinction between "local correlations" and "feedback" may matter - not
in the design of the hardware, but in the design of the software, which
is your focus. OTOH, any algorithm that produces learning in a neural
net must include those local correlation algorithms. Whether the local
correlation algorithms have any use on their own is another question.
You talk of pruning synaptic connections, but whatever criteria you
build into the algorithm must be external to the LC's. In nature, AFAIK
the spontaneously generated synaptic connections are in fact pruned by
feedback from precisely those external contingencies that you seem to
wish to eliminate from the development of the organism.