jagversm@praxis.cs.ruu.nl (Koen Versmissen) (03/20/90)
I'm to give a talk in a seminar on neural nets shortly, and I thought
Harmony theory would be an interesting subject. I have a few questions
though:
- I only know harmony theory through Paul Smolensky's chapter in the
Rumelhart & McClelland PDP-book. What have been the developments
since? References to important articles would be most welcome (even
though it may be too late for me to use them in the preparation of
my talk).
- I hope someone can shed some light on something that I found
particlularly puzzling in Smolensky's article:
Harmony theory claims to be psychologically relevant ("... these
dynamical systems can serve as models of human cognition", p. 195).
Now neurons are not (or at best hardly) probabilistic, right? How
then can harmony theory claim to provide a _micro-level_ description
of the mind? Shouldn't it rather be called an _intermediate level_
description? After all, an account of how probabilistic units can be
implemented (or rather, simulated) by means of, say, deterministic
threshold units, is still needed.
Does all this make some sense?
(N.B. even though Smolensky stresses the importance of taking small
steps, I find it strange that he didn't mention the above issue at
all).
- The same question as the above, but now concerning symmetric
connections in Harmony theory, as opposed to asymmetry in the brain.
Koen
----------------------------------------------------------------------
Koen Versmissen
(jagversm@praxis.cs.ruu.nl)bill@boulder.Colorado.EDU (03/21/90)
In article <2693@ruuinf.cs.ruu.nl> jagversm@praxis.cs.ruu.nl (Koen Versmissen) writes: > Now neurons are not (or at best hardly) probabilistic, right? I can't claim any particular expertise with respect to Harmony theory, but I can tell you that most neurons (and all the ones thought to be involved in cognition) are extremely probabilistic. There are a number of factors (variable amount of neurotransmitter released by a synapse; random opening and closing of channels; etc.) combining to make it unpredictable when a given cell will fire -- in addition, because of the all-or-nothing character of firing, any moderately complex system of neurons will form a chaotic system (in the dynamical systems sense), so that quantum uncertainties will quickly be amplified into global randomness. There was a book published a few years ago devoted to analyzing the stochastic behavior of neurons; the title was something like "Stochastic Analysis of Neural Activity." (I don't have the reference, and I don't know the author; it's not in our library.) > How > then can harmony theory claim to provide a _micro-level_ description > of the mind? Shouldn't it rather be called an _intermediate level_ > description? I believe you are right that the units of Harmony theory are not necessarily intended to correspond to biological neurons: the theory is "micro" level in comparison to the "macro" level of explicit rules, not in any physiological sense. >- The same question as the above, but now concerning symmetric > connections in Harmony theory, as opposed to asymmetry in the brain. The question of symmetry in the brain remains open. If indeed the brain makes use of Harmony-style relaxation-calculations, they probably take place in restricted subnetworks confined to relatively small portions of the cerebral cortex -- and there are in fact classes of local connections in cortex which are roughly symmetric. (Of course there are also a great many classes of asymmetric connections.) The "style" of computation in the brain is something neuroscience has at present few effective ways of investigating. -- Bill Skaggs