(from Peter Cariani)
"Psychoanalysis is like the Russian Revolution -- it isn't clear at
which point the whole thing turned bad." --SchizoCulture Semiotext
I just wanted to second (with a rant of my own) Ray Allis' remarks (9/9)
on the symbolic logic straitjacket that has enveloped most "AI" since
the 1956 Dartmouth conference. It was at this conference that
it was decided (erroneously, in my opinion) by the proto-AI community
that all processes are amenable to digital, rule-governed description,
and that therefore the route to artificial intelligence would
necessarily and solely involve logic-based problem-solving. (Alternately,
one could lay the blame on Carnap's conversion to Platonism in the
1930's (with Tarski and Godel and so many others)).
The problem with this Platonization of intelligence is that the
semantic categories needed to solve a given problem, the linkages
between the symbols that are manipulated in the machines and what is
going on in the outside (material) environment, were not included in
the formulation of intelligence (these people were not biologists or
psychologists, but mathematicians). In biological systems these
(external semantic) linkages are made through the senses. In scientific
models via they are implemented using measuring devices. To any
practicing scientist, it is inconceivable that we could do away with
making measurements, and to any practicing engineer (besides computer
scientists) it would be inconceivable that we could do away with
effectors that act on the world, but this is exactly what the
AI community did by excluding real perception and action.
If we want to reintroduce perception and action into the general
project of building intelligent and adaptive devices, then we need
to situate our devices in real world environments with sensors and
effectors, performing real tasks, with explicitly stated performance
criteria. In my doctoral work I proposed evolutionary robotic devices
which would adaptively construct their own sensors and effectors, thereby
selecting which external semantic categories they would use to solve
the problems at hand (e.g. survival). Perceptual control theory, with its
insistence on organism-environment feedback perception-action loops and
analog mechanisms, similarly moves in the right direction.
Thus it is not "induction" per se that we need to put into our
devices -- it is grounding to the world by perception, action, and selection
that is needed. We need to make this grounding as flexible as possible, so that
the categories in which learning takes place are as adaptable as possible.
The sensory systems of living animals are capable of changing the external
semantics of central representations by both attentional mechanisms, by
plasticity of processing, and by action (e.g. moving the sense organ or
by breaking an object to expose its interior).
Functionally the sensors do not end at our retinas or our cochleas, the
analog representation that they provide extends all the way up to the cortex,
permitting a high-quality analog representation there from which any relevant
properties can be (adaptively) extracted. This is distinct from the "pixel
theory" of perception, in which the outside world is encoded into little,
discrete sensory atoms early on,the information is conveyed more centrally, and
then reassembled
via multitudes of feature detectors.
Many mixed digital-analog pulse-coded schemes are possible alternatives to
this discrete and static universe, and I am currently trying to mentally
work out how temporally-based processing might be carried out in the
central auditory system to recognize complex perceptual forms (periodicity
pitch,
musical intervals, speech). We'll see how it goes ....
Peter Cariani