Posted   17 Aug 1996
Newsgroups: comp.ai.philosophy
References: <4rcir5$ioh@usenet.srv.cis.pitt.edu> <811652916wnr@chatham.demon.co.uk> <4ub0nn$oes@usenet.srv.cis.pitt.edu> <4ubpjm$qm1@usenet.srv.cis.pitt.edu> <4uoqao$icg@usenet.srv.cis.pitt.edu> <418424767wnr@chatham.demon.co.uk>
Subject: Is there non-symbolic processing? (Was: Note on truth-values, Frege -- Re Sloman Part 7)

Oliver Sparrow <ohgs@chatham.demon.co.uk> writes in response to Anders N
Weinstein.

> Date: Wed, 14 Aug 1996 08:04:56 +0100
> Organization: Royal Institute of International Affairs

Oliver seems to be invoking a distinction between symbolic and
non-symbolic processing which is apparently widely accepted, but
which I think is fundamentally misconceived: there is no
distinction.

[AW]
> > Now would you try to explain the nature of promising or writing a
> > check by looking inside human brains? Is this a task for brain science?
> > Even if you found correlates of your promises there, that would tell
> > you nothing about the nature of promising as a social act.

[OS]
> Nicely put. As usual.
>
> One can then ask, (a) where to these "agencies" come from; and (b),
> given that they act, how can I layer my language of description so
> as to take into account a nested, layered operating environment?
>
> (a) takes one into the issues of unbounded emergent structures.
> (b) suggests that symbolic or analytical reductionism is a silly place
>       to start for hard AI.
>
> Taking (b), as I have bored on about (a) quite sufficiently, there is a
> tendency to see the representation of cognitive processes within
> artificial constructs - companies, societies as well as computers -
> as being
>
> (1) the collection of a number of tokens of meaning: symbols, analysis.
> (2) the creation of a grammar by whch these symbols and tokens are
>       to interact.
>
> This approach grows naturally from how symbolic thought (that tiny
> fraction of overall cognition) is supposed to be undertaken; and the
> official view of how science proceeds. I have written a piece under
> the 'opacity' thread which addresses this orthodoxy, so I shall not
> repeat what I said.
>
> I suspect that symbolic thought represents an emergent peak that
> develops from massive underlying geological processes.
> ....

> ... If you do not have this infrastructure,
> then manipulation will fail when it is taken outside of a very restricted
> domain. The symbolic representation will be unable to detect these
> limits, however, whilst a system of which the symbols are emergent
> tokens of deeper connectivity will (may) be able to detect this, insofar
> as the emergent processes no longer support the generation of the
> symbols themselves. What looked like a cat was, in fact, a discarded
> fur hat.

Now I am not totally clear what point is being made but it looks
like a version of the common contrast between symbolic and
non-symbolic or sub-symbolic processes, which characterised some of
the (often unclear) debates between connectionists and other AI folk
in the 80s.

I occasionally find the term "sub-symbolic" useful as a mnemonic
when contrasting the operation of a set of condition-action rules
with other mechanisms which they may invoke, e.g. neural nets, low
level numerical or symbolic procedures, etc.

But the idea that there is some major ontological or epistemological
contrast between the symbols that occur in e.g. a set of condition
action rules, or a planner, or a theorem prover on the one hand, and
the symbols that are used in low level vision routines, or motor
control routines, or content addressible memory, or neural net
simulators on the other, strikes me as based on a failure to look
closely at all the deep similarities.

In all these cases there are symbols which hold information used by
the system. The information may be of different sorts and scales,
and the uses may be of different types.

For instance the information may be about something in the
environment, that table, or the edge of that table, or the slope of
a fragment of the edge of that table.

Or it may be about some internal structure that is involved in the
processing of sensory data, e.g. information about the sensory input
in a particular pixel or about the the boundary between two regions
found in a region detector, or about the contrast across a fragment
of an edge feature.

Or it may be about some other sort of internal structure, e.g.
information about the structure of a sentence, the gaps in a plan,
the scope of database, a procedure that can be used by the
system, etc.

Or it may be control information regarding whether or not to do
something or when to do it or how much weight to give to this option
as against other options when selecting what to do. The latter could
be implemented as a synaptic weight in a neural net, a numerical
weight associated with a condition action rule, or some conditional
probability in an explicit decision making system -- but essentially
they are ALL doing the same general sort of thing, i.e. symbolising
a weight (or probability) to be used for control purposes.

I once read a book by a philosopher who thought symbols in an AI
system had to correspond to words that are used in a public
language, like English. That's the sort of misinformation that comes
from teaching people AI courses which merely include such things as
simple expert systems or simple natural language or problem solving
systems, without saying anything about robotics, vision, motor
control, etc.

A slightly different assumption seems to be made by Oliver above.
Apparently he thinks that symbols have to be the contents of
thoughts, or things we are conscious of using, whereas the "massive
underlying geological processes" are totally different and
non-symbolic. I don't know how closely he has looked at AI work on
robotics.

It seems to me that unless the underlying processes involve symbols
performing the sorts of functions I have listed above, and others,
then they cannot provide the geological infrastructure he thinks
they provide.

Of course they may be implemented in an even lower level which
involves synapses or transistors, or electrical spikes, or chemical
concentrations, but if we are to understand the functional role of
those physical mechanisms we'll have to know something about what
information they are manipulating and how that manipulation is
controlled, usually by other information. That means bringing in a
level of symbolic processing, even if it is very different in some
ways from conscious thought processing or the "highest level"
symbolic processing in AI systems.

(It's not quite that simple as not all physical components directly
perform a symbolic functional role, e.g. when there are more
sophisticated mappings between virtual machines and underlying physical
machines, e.g. in sparse arrays, virtual memory systems, infinite lists
implemented as memo functions, etc. But even then my point remains that
there is a level, probably well below the level of what we are conscious
of, at which information is used and manipulated and which is part of
the explanation of how the total system works: this is commonplace in
computing systems and seems to be a working hypothesis of many neuro
scientists..)

The difference is concerned with

    o which information is symbolised,
    o what control decisions are represented and
    o what sort of syntax is used,

not with whether it is symbolic (in the most general sense,
i.e. involving structures with some sort of meaning) or not.

Another way of putting this is that there's no very deep difference
between the sorts of processes in our minds that we are aware of and
those we are not and cannot become aware of. It's not a particularly new
idea.

Now I'll have both Oliver and Anders jumping on me.

Aaron
==