Query: Dendritic Networks

Kevin Spencer kspencer at s.psych.uiuc.edu
Fri Feb 9 15:49:47 EST 1996


NobodySpecial <steve at cs.titech.ac.jp> writes:

>Hi,

>	I just wanted to point out that there has been a fair amount of abuse
>by all sides.  The problem is not modelling vs biology, or data vs
>understanding.  The problem is that there are a number of people who do
>not have a complete understanding of what they are talking about.  I
>have read "shadows of the mind" and was annoyed not by the quantum
>approach, but by penrose's ignorance of the neurosciences, etc.  In the
>same way there are many psychologist modellers who try to model
>psychological phenomenon with backprop trained mlp's.  This is a problem
>because the mlp basically learns to emulate the data.  So, in the end
>you have a black box that performs in accordance with the data.  So,
>what?  What does that explain about psychology. 

Neural-network modeling of psychological processes is not fundamentally
different from modeling with "traditional" statistical methods -- both
are instances of theory-building.  A theory that explains the data can
be OK or good depending on the range of data that it accounts for, but
in the end only "works" if it makes predictions that can be verified
in subsequent experiments.  This holds for all theories, whether or not
they can be implemented as algorithms.

I think that one "special" benefit you can get from modeling a
psychological process with a neural net is that if it works, you have an
existence proof that the process can indeed be implemented in a
computational architecture like the neural net.  Of course, it depends
on the theory whether or not this existence proof would be interesting.

Kevin
-----------------------------------------------------------
Kevin Spencer
Cognitive Psychophysiology Laboratory and Beckman Institute
University of Illinois at Urbana-Champaign
kspencer at p300.cpl.uiuc.edu / kspencer at psych.uiuc.edu
-----------------------------------------------------------




More information about the Neur-sci mailing list