On Sun, 02 Nov 2008 17:11:54 -0500, Christopher D. Green wrote:
> Mike Palij wrote in response to the following comment by Chris:
>>> You may be right that parallel distributed connectionist models (a less 
>>> tendentious name for "neural networks") 
>>
>> "Tendentious"?  Not to argue the point but considering that 
>> one can trace these models back to McCulloch & Pitts (1943)
>> and other models of the nervous system (and even earlier
>> connectionist conceptions), couldn't one claim that "neural
>> network" has priority in describing these types of models?
>>   
> Origin is not essence, as they say. Yes, McCulloch & Pitts were one 
> (two, actually) of the first to use these kinds of models (you'll find 
> something similar slightly earlier in Rashevsky), but there was nothing 
> particularly "nervous" about them (apart from McCulloch's use of the 
> term). There are at least as many disanalogies between them and neural 
> structure as there are analogies, and there are all kinds of ways to 
> interpret their activity that has nothing whatever to do with neural 
> modeling (See Green, 1998, 2001).

Okay, a few points:

(1)  As James Anderson presents in his 1995 text "An Introduction to 
Neural Networks", it could be argued that William James proposed
an early form of neural networks in his 1892 "Psychology: Briefer
Course".  See Anderson's pages 148-149 and Figure 6.1 which
illustrates a kind of two layer association system.  Anderson presents
this in the context of explaining Hebbian learning rules.  One could 
probably make the argument that such conceptions could be traced 
back to Herbart who tried to work out the mathematics of excitation 
and inhibition of ideas in order to explain why some ideas were in 
consciousness and other were not (but the changing patterns of 
excitation and inhibition could make ideas change  places).  I realize, 
however, that associationism's long history has to be seen as primarily c
oncerned with assciations among "ideas" and not "neurons" though 
this does change in the 20th century (I believe, in part, because 
of the work of McCulloch & Pitts).

(2)  Paul Cull's brief biography of Nicolas Rashevsky puts his work 
and McCulloch and Pitt's work into appropriate historical context (see:
Cull, P. (2007). The mathematical biophysics of Nicolas Rashevsky.
Biosystems, 88, 178-184 -- available at the ScienceDirect database).
In 1933 Rashevsky did provide a model for the behavior of a neuron
and how a neural network might operate.  McCulloch & Pitts (who
also resided in Chicago while Rashevsky was there and published their
1943 paper in the "Bulletin of Mathematical Biophysics" of which
Rashevsky was the editor), proposed an alterantive model.  Where
Rashevsky's model was more concerned with continuous activity,
the McCulloch & Pitts (M&P) model was concerned with discrete 
states and the rules that could operate on them.  As James Anderson 
(see ref above) and others characterized this, M&P seemed to interested
in showing how the logical relationships of the "first order predicate 
calculus" could be modeled by their type of neurons (M&P 1943
represent these networks in form of neurons in their Figure 1 while
James Anderson givens a more abstract version including the "truth
table" associated with a M&P neuron in his Fig 2.9 on page 49).  
The lasting importance of the M&P idealized neuron, as both Cull
and Anderson point out,  is that John von Neuman used the ideas 
associated with M&P neurons to develop his own ideas for the
digital computer, meaning that our contemporary computers can
be thought of as the "mutated" offspring of M&P. ;-)

(3)  James Anderson points out, the M&P neuron had properties
generally consistent with what was known about the properties of
neurons circa 1940s (see Anderson p49-50).  Given that it is an 
abstraction which is supposed to embody certain mathematical ideas, 
it does not have to exactly mirror real neurons but be credible 
representations.  In your papers you seem to argue that because 
artificial neurons do behave "exactly" like real neurons, one can rule 
them out as credible models, even as crude approximations.  I'm 
sure you'll correct me if this interpretation is wrong.  However, 
given the growth of knowledge about the properties of neurons and
the nervous system since 1943, I think one might want to guard
against a "presentism" tendency to discount the importance of M&P's
contribution.

(4)  Finally, I acknowledge that the following should be interpreted
tentatively subject to further confirmation but I think that it effectively
deals with the issue of whether M&P neurons were supposed to
view as "real neurons".  From Ken Aizawa's entry on the history of
connectionism in the University of Wateloo's "Dictionary of Philosophy
of Mind":

|During the 1930's, Nicolas Rashevsky proposed to use differential 
|equations and physical concepts, such as energy minimization, to 
|describe how the behavior of nerves and networks of nerves that 
|might be related to psychological processes, such as Pavlovian 
|conditioning. (See, for example, Rashevsky, 1931a, 1931b, 1935). 
|Rashevsky's work was part of a larger project of developing a mathematical 
|biophysics that would mirror the methods of mathematical physics.
|
|Rashevsky was instrumental in bringing together Warren S. McCulloch 
|and Walter Pitts, who in 1943 published a seminal contribution to many 
|fields, "A Logical Calculus of Ideas Immanent in Nervous Activity". 
|This work described how networks of binary threshold neurons might 
|be described in terms of sentences of first-order logic. Although 
|popularly remembered for having contributed the idea that networks 
|might carry out logical inferences, McCulloch and Pitts were themselves 
|more interested in the description of networks containing closed loops. 
|*****This interest stemmed in part from McCulloch's work in tracing neural 
|pathways using strychnine neuronography.***** Neuroscientific findings by 
|Rafael Lorente de No also increased the interest in closed neural circuits.
http://philosophy.uwaterloo.ca/MindDict/connectionismhistory.html

>> It might be just me but it seems that J.J. Gibson might be becoming 
>> increasingly relevant to cognitive psychology.  Aren't Gibsonians fond
>> of saying "it's not what's inside your head that's important, it's what
>> your head is inside of that is".  Or something like that.
> 
> I've always thought that J. J. Gibson was important (but, then again, my 
> own supervisor was one of his students, so I got a heavy dose of 
> ecological psych when I was a grad student). 

Several researchers in AI and elsewhere who are interested in the role
of perception and action seem to be working in a Gibsonian framework
but rarely acknowledge him as a source (e.g., Steven Sloman in his
book "Causal models" spending several pages describing perceptual
processes that I immediately recognized as being Gibsonian in nature
but he only provides a brief reference to Gibson at the end of the chapter).

> IMHO, Gibsonians (like 
> Brooks) overstate the case -- the trick is not to abolish cognition, but 
> to conceive of it in an ongoing interaction with the environment --  but 
> they are still on to something important. 

Ulric Neisser in his attempts to develop an ecological approach to
cognition had to ultimately admit that mental representations that
Gibson eshewed had to be accepted as being legitimate at some
level.  I think the question becomes how much information can an
organism directly extract (perceive) from stimulus information before
a mental representation has to be retrieved.  

>Now if we could only cash out  the fascinating insight of "affordances." :-)

This is a frustrating concept but I believe that one way of thinking
about affordances is that it involves the perception of functional properties
of stimulus or environmental configuration that can be associated with
previous stimulus-action sequences.  For example, chairs are associated
with sitting so what other objects in the environment will "afford" us the
ability to sit on them?  Of course, there is the recognition that some things
may be better things to sit on (e.g., a garbage can) than others (e.g., a
road cone ;-).

-Mike Palij
New York University
[EMAIL PROTECTED] 

> Refs:
> Green, C. D. (1998). Are connectionist models theories of cognition? 
> <http://psycprints.ecs.soton.ac.uk/archive/00000553/> /Psycoloquy 
> <http://psycprints.ecs.soton.ac.uk/>, 9/ (4).
> Green, C. D. (2001). Scientific models, connectionist networks, and 
> cognitive science <http://www.yorku.ca/christo/papers/models-TP2.htm>. 
> /Theory and Psychology, 11/, 97-117.


---
To make changes to your subscription contact:

Bill Southerly ([EMAIL PROTECTED])

Reply via email to