From Aaron Sloman Thu Mar 14 09:59:50 GMT 1996
Newsgroups: comp.ai.philosophy
References: <4g367o$p62@usenet.srv.cis.pitt.edu> <4ga4jb$14eq@zen.hursley.ibm.com> <4gama8$mue@mp.cs.niu.edu> <4gc51s$vis@zen.hursley.ibm.com> <4ge0cg$j1j@usenet.srv.cis.pitt.edu>
Subject: JJ Gibson and AI (Was  Physics vs. computation)

andersw+@pitt.edu (Anders N Weinstein) writes:

> Date: 21 Feb 1996 02:34:56 GMT
> Organization: University of Pittsburgh

> ....[much deleted]

> I do think, however, that the Gibsonian approach works best for for
> non-linguistic animals, and it is far from obvious how to adapt it to
> rational creatures like ourselves. After all, we are capable of reasoning
> before acting. Nevertheless I take that even here, the AI community is
> recognizing that the traditional view vastly exaggerates the extent
> to which conscious, explicit ratiocination precedes intelligent action.

I have never taken AI, even in the so called "classical" period when I
first got involved with it (1969 onwards), as having anything much to do
with "conscious, explicit rationocination." That was obvious especially
from the early work in natural language processing and vision which
investigated processes of which we *obviously* were totally unconscious.

(I think most current views of GOFAI are just a distortion of history
based on partial knowledge of what was really going on in AI -- for
which some practitioners with narrow viewpoints were partly to blame. NB
I am not accusing Anders of that distortion.)

Some years ago, I tried to reconcile what's good in Gibson's
theories of perception with an AI-informed mechanistic viewpoint in:

A. Sloman,
    `On designing a visual system: Towards a Gibsonian computational
    model of vision'
    Journal of Experimental and Theoretical AI
    1,4, 289-337 1989

Also available as
    ftp://ftp.cs.bham.ac.uk/pub/groups/cog_affect/
in the file
    Aaron.Sloman_vision.design.ps.Z

(apart from a few pictures -- well known ambiguous figures -- which
were hand drawn for the original article and not included in the
postscript file).

Abstract:

This paper contrasts the standard (in AI) "modular" theory of the
nature of vision with a more general theory of vision as involving
multiple functions and multiple relationships with other sub-systems
of an intelligent system. The modular theory (e.g. as expounded by
Marr) treats vision as entirely, and permanently, concerned with the
production of a limited range of descriptions of visible surfaces,
for a central database; while the "labyrinthine" design allows any
output that a visual system can be trained to associate reliably
with features of an optic array and allows forms of learning that
set up new communication channels. The labyrinthine theory turns out
to have much in common with J.J.Gibson's theory of affordances,
while not eschewing information processing as he did. It also seems
to fit better than the modular theory with neurophysiological
evidence of rich interconnectivity within and between sub-systems in
the brain. Some of the trade-offs between different designs are
discussed in order to provide a unifying framework for future
empirical investigations and engineering design studies. However,
the paper is more about requirements than detailed designs.


NOTE 1:
My interpretation of Gibson's rejection of representations was that
he had been too much influenced by a particular type of philosopher
into thinking that representations must be things of which we are
conscious. So he was, for example, objecting to epistemological
theories that claim that we somehow consciously *infer* what's out
there from what's "given" in experience. He also seemed to think
that computation had to involve explicit, maybe even conscious,
execution of something like computer programs. I got the impression
from reading his work that he knew nothing about AI approaches to
vision, and might not have objected to them if he had heard about
them, except for one point: they tend to assume that vision starts
from images on a retina, whereas Gibson thought (rightly in my view)
that vision starts from the structure of the optic array, which is
sampled by the combination of lens and retina (and frequently
re-sampled). It was Steve Draper who first drew my attention to this
point, in the late 70s I think.

I have not kept up with recent AI work in vision, and maybe now more
of it is Gibsonian in this respect.

NOTE2
Gibson's affordances are, I think, a special case of something
important and general, i.e. that for an intelligent agent the
environment is not just composed of structures (or structured
objects in a variety of spatio-temporal relationships), but is
largely composed of things like abilities, propensities,
dispositions, capabilities, powers, tendencies, limitations,
constraints, opportunities, risks, dangers, etc. etc. and these
things have to play a key role in the agent's ontology. By that I
mean that reference to such things will pervade the whole
information processing system, including perceptual mechanisms,
long term memories, plans, goals, decisions, thoughts, etc.

My provocative summary of this is: reality is inherently modal.

At present my ideas about these things are still not very clear. Maybe
I'll post some notes later. I believe that this ontological point will
ultimately be the basis for a new convincing functionalist account of
mind, consciousness, or whatever, including explaining the painfulness
of pains, the niceness of enjoyable things, etc. More anon.

Cheers
Aaron
(This note is http://www.cs.bham.ac.uk/~axs/misc/gibson.vision)
---