Fundamental Compressionist Philosophy.

[Gerry Wolff]

----- Original Message -----
From: "Andrew Stanworth" <andrew.stanworth at bigfoot.com>
To: "casc mail list" <casc at sanna.com>
Sent: 23 May 2001 04:47
Subject: Re: Fundamental Compressionist Philosophy.

...


> Gerry wrote;-
>
> | The ICMAUS ideas are intended to be relatively abstract, more abstract
> than
> | the distinction between 'neural' and 'symbolic' etc. That said, they
have
> | borrowed a lot from cognitive psychology so in that sense they have a
> | 'neural' flavour. I am currently planning an article about how the
ICMAUS
> | framework could be realised using neural mechanisms.
>
> I don't understand what you mean by more abstract, since 'symbolic' is as
> abstract as I can conceive.  Your system seems to work by automatically
> uncovering symbolic interrelations - turning input data into a mixture of
> raw and symbolic elements, but I can't see anything more abstract than
> this - though I would like to!  If you mean more abstract than 'neural' I
> can relate to that, though from what you say of the origins of your
system,
> and your current intentions, I take this as being a tentative agreement
that
> it conceptually belongs in the neural learning family of algorithms -
though
> I plainly accept that it exists as the concepts abstracted away from any
> physical or mechanical idea of neuronal models.  You may instead be saying
> that if 'neural' is 'a', 'symbolic' is 'b' and 'ICMAUS' is 'c', then c is
> greater than b minus a (c > b - a) which still doesn't make sense to me.
> Life in this group just gets harder and harder!!
>

These are fair points. Let's see whether I can be a bit clearer about what I
had in mind.

A convention has grown up that the kind of computing done by conventional
computers is 'symbolic' (because they use letters, numbers etc) and the kind
of computing done by a neural net is 'sub-symbolic', 'non-symbolic' or
'connectionist' because (artificial) neural nets (and perhaps natural ones
too) don't seem to use the kinds of chunky symbols that ordinary computers
use. But most neural nets are actually run on ordinary digital computers
(which use the symbols '0' and '1' at their basic level). So the distinction
is not a very precise one.

When I said the ICMAUS ideas were intended to be relatively abstract, it was
this rough-and-ready distinction that I had in mind. I was really trying to
respond to your suggestion that "your system belongs in the 'neural network'
family of architectures." The framework uses symbols, and current
implementations run on an ordinary digital computer (which uses binary
symbols). So in that sense, the framework is symbolic. But the underlying
concepts (information compression by matching and unification of patterns)
are not, in themselves, symbolic (you may dispute this). And it seems
entirely possible to me that the whole framework could be implemented using
neural mechanisms (which, according the rough-and-ready distinction) are not
symbolic. In short, I was trying to say that the system is not intrinsically
'neural' and is not fundamentally symbolic either (in the rough sense of
that term).

"I don't understand what you mean by more abstract, since 'symbolic' is as
abstract as I can conceive". Are analogue computers (using things like
continuously-varying voltages) symbolic? If not, then the features that are
shared by digital and analogue computers are, presumably, more abstract than
'symbolic'.

Regards,

Gerry