Re: Schyns Comms 10-12 french gauker grossberg

From: HARNAD Stevan (harnad@coglit.soton.ac.uk)
Date: Mon May 04 1998 - 16:21:56 BST


> From: "Valenti Gianni" <gv195@coglit.soton.ac.uk>

gv> This could be possible but for all categories to be learned still some
gv> generalising mechanism must be present as learning the category of
gv> every single objet animal and so forth to even more abstract concepts
gv> such as emotional states would take an incredibly large amount of time
gv> and memory space in the brain.

But the only relevant question here is whether it could be done with a
fixed repertoire of feature-detectors, or whether it requires creating
feature-detectors..

gv> our innate fear of snakes could be a possible
gv> evidence for the opposite: the existence of inborn known feature
gv> detection.
gv> Yet to have all possible features for all instances could be
gv> over-loading the system as well as being under the danger of possible
gv> genetic mutation.

I wouldn't worry about genetic mutation; many other things are largely
inborn, yet not prone to distortion by mutation.

It would be useful, though, to take an analogy from fixed vs. flexible
effects in the genetic coding of both anatomical and behavioural
traits: In general, where there is the option of letting the "world" be
its guide, even for spatial properties evolution "prefers" to code them
flexibly: Instead of coding growth as "grow 5 microns forward, then
turn left," evolution prefers to code "grow till you hit an obstacle,
then follow the easier direction." This will be fine if it is highly
probable that, based on the growth of other structures at the same time,
an obstacle will be reached by the time 5 microns are grown, and
if it is also highly probable (perhaps because, say, the heart is on the
left and there is already a bias in the tissue) that it will be easier
to grow left than right when the obstacle is reached.

This is how genes use the predictable geometry and timing of the
environment to code things flexibly and interactively, instead of
rigidly. It is rigidity that is at risk from mutation, or from a
developmental error, or even from a change in the environment;
flexibility is what protects and buffers us from these.

The same is true for nonspatial genetic traits: Rather than coding a
behaviour pattern rigidly, evolution takes advantage of predictable
features of the environment: The imprinting duck is not programmed
rigidly with the shape and movement of ducks; it is just programmed to
follow and imprint on the first moving thing it sees (which, with high
probability, will be its mother). It is safe then to direct sexual
behaviour at maturity toward whatever looks like that. Of course, if
instead of the predictable mother-duck, the newborn duck instead sees
and follows a human, as they did with Konrad Lorenz, then that
"flexible feature detector" will incline the duck to try to mate with
humans instead of ducks at maturity!

So this is more than just an analogy: One of the traits that evolution
has an option to code rigidly or flexibly is feature detection itself.
Now, most of this kind of "leaning" on the predictable environment to
provide constraints that then need not be rigidly coded genetically is
indeed based on the highly predictable features of the environment. So
in that respect it is not much more flexible than fixed features,
because the kinds of categories that evolution would expect --
encountering the mother and the mammalian breast, the vocalisations of
our species, perhaps also our predators and prey in our original
environment of evolutionary adaptedness (EEA) -- would all be pretty
universal to all of us. In contrast, the infinity of shapes that matter
to us in the modern world go far beyond any of the expectations of our
genes, which are based on our EEA, in which and for which we evolved.

But somehow, Baldwinian Evolution (the evolutionary selection for
faster and better learning of a predictable behaviour, flexibly guided
by the environment, rather than rigidly coded in advance) seems to have
carried us into an extremely general world of all possible categories,
categories for which it is not only hard to imagine that we had
"prepared," fixed feature detectors, but just as hard to imagine that
we had flexible feature-detectors, ready to be "tuned" by the
predictable environment. The reason is that most of are categories are
not predictable at all! They are not part of the structure of the world
that our genes encountered for a long time predictably in the EEA,
allowing them the luxury of constructing detectors for them through
learning rather than coding them them fully in advance. In fact, some
of our categories are so recent that even our grandparents would not
have recognised them, let alone our ancestors in the EEA.

Now it may be that, by some sort of luck, the flexibility that was
shaped in our EEA happened to be enough to generalise and handle
even today's categories (cars, jet engines, integrated circuits,
cancer cells, etc.). Maybe it only SEEMS that our feature-detectors
are flexible beyond any possible predictions originating from our EEA.
Maybe it seems that way because we are totally unaware of all the
categories that we cannot and never will be able to detect, because
they do NOT conform to the flexible predictions formed in our EEA.
If that's true, then our seemingly flexible features are not as
remarkably flexible as they seem to be.

But the other possibility is that the combination of (1) our flexible
sensory feature-detecting capacity applied to the stimulation we get
from the CONCRETE, perceptible world, supplemented by (2) the
technology we have managed to invent for extending our sensory powers
(camera, microscope, microphone, telescope, seismometer), plus,
especially, our (3) capacity for defining and discovering ABSTRACT
categories through language and symbols, categories that no longer depend
directly on our sensory feature detectors, even when our senses are
enhanced by instruments -- perhaps all this has left all specific
genetic expectations behind long ago, and launched us into a world of
concrete and abstract categories that...

Now you have to fill in your part: that (a) give us a more detailed and
accurate picture of reality than our genes and their predictions could
ever have done or that (b) are constructed out of our own imagination
and invention, genetically grounded, but all-purpose (Whorfian
Constructivism).

>ab> Schyns demonstrates which French and Weaver acknowledge
> >that there is a aquisition of new features during catagory learning.
>ab> French and Weaver do not follow that the aquisition of new features
>ab> is a typical part of catagory learning.
gv>
gv> What is therefore possible is that a general inborn facility to
gv> distinguish dangerous animals with non dangerous animals, could be
gv> possible (being this an evolutionary advantage), but that finer
gv> subgrouping has to be acquired using a "feature subgrouping" type of
gv> system.

And what does that mean? And/Or/If/Not combinations of the fixed
features and detectors? or the "construction" of new ones?

>ab> French and Weaver argue that working assumption is reasonable because
> >if you look at the evolving constraints and universally
> >shared experiences with objects
> >in the world this suggests that by the time of adulthood,
> >most of the primitives
> >humans need to handle concrete categories will already have been
> >learned which is why we argue that fixed feature approach for the
> >learning of concrete categories is a reasonable working assumption.
>
gv> Following this argument, it is understandably so.

But perhaps this is too quick: Do children then have flexible features,
but they harden by the time they become adults? Surely this would
depend on experimental tests to see whether it is so, not on arguments.

>ab> Feature based theoriesd of concept formation have two dilemmas.
>ab> 1= for many natural concepts it is hard to see how the concept of the
>ab> features could be developmentally more basic.
>ab> 2= the concept formation must be guided by abstraction heuriotics.
>ab>
>ab> Schyns theory argues that the concepts of the
>ab> features need not be developmentally more basic. They don't explain
>ab> how novel features might be formed.
>ab>
>ab> Yet all agree that on one thing
> > a theory of conceptual development has to explain how children learn
> > to apply category words to roughly the same sort of things as'
> > others who have learned language.
>
gv> yet their application seems to follow a prototype similarity approach,
gv> therefore indicating a prototype formation of the concept, as a product of
gv> learning guided by an inborn "generalising concept". In fact, children
gv> make the mistake of calling whales and dolphins as fishes, as opposed to
gv> their correct classification as mammals. They do this probably because the
gv> general habitat of these animals as well as their form is more
gv> prototype-like of fishes than other groups. The same goes for penguins and
gv> birds.

Not clear what this is meant to imply...

What are "prototypes"?

>ab> No two peoples abstracts general concepts form exactly the same
>ab> clas of instances. Yet people agree on categories.
>ab> These categories, it is being argues, maybe innate or shared
>ab> principles of abstraction.
gv>
gv> Or it is an innate ability for category formation, using immediately
gv> identifiable variables (such as form an environment of interaction).

Probably best to think of specific examples, data and models here,
because this is a bit too vague and general...

>ab> Grossberg uses adaptive resonance theory ( ART )
>ab> There are two levels.
> >the lower level deals with new features that
> >contribute to the larger categories.
> >This higher category level can selectively bind together or fuse
> >certain combinations of new features in one contexed and different
> >combinations in another.
> >
>ab> This process generates large categories and conserves memory
>ab> reserves.
>
gv> This could be a very good representation of the actual human module, but
gv> it limits itself to explain the categorisation of external concrete
gv> objects, leaving emotional states and complex situations unaccounted

Why would ART not work on emotional categories as well as on external
ones? And what complexity do you have in mind?

gv> Definitely a greater investigation into the matter is required, were the
gv> answer will probably lie in-between the two opposing viewpoints.

Please, on the exam, everyone avoid comments like "the truth is in the
middle" and "we need more data." That's obvious, and could be said
about anything and everything. You must give kid-sib evidence that you
are thinking critically, and about the devil in the details, not just
a bland ecumenism...

--------------------------------------------------------------------
HARNAD Stevan harnad@cogsci.soton.ac.uk
Professor of Psychology harnad@princeton.edu
Director, phone: +44 1703 592582
Cognitive Sciences Centre fax: +44 1703 594597
     
Department of Psychology http://www.cogsci.soton.ac.uk/~harnad/
University of Southampton http://www.princeton.edu/~harnad/
Highfield, Southampton ftp://ftp.princeton.edu/pub/harnad/
SO17 1BJ UNITED KINGDOM ftp://cogsci.soton.ac.uk/pub/harnad/



This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:21 GMT