In Pinciple Arguements Against Dynamic
Connectionism
Charles Wallis
California State University, Long Beach
There is a long standing debate in cognitive science regarding
how best to understand computational capacities of connectionist systems
and the brain. One side of this debate advocates an approach to understanding
the computational capacities of the brain would treat these systems as Turing-equivalent
computers. The other side of this debate urges that researchers treat
these systems dynamically. To treat connectionist systems dynamically,
is to treat them as computing devices that are fundamentally different from
Turing Equivalent computers. Specifically, it involves, as Smolenski
(1988) has suggested, treating these systems as ones in which computation
occurs according to differential equations describing the numerical evolution
of the activation patterns--not by means of the recursive application of
a finite set of read/write rules for manipulating symbols. (I ignore
a third position, dynamic systems theory as it is non-computational in approach.)
In this paper I review in principle arguments found in the philosophical
literature against treating connectionist systems as dynamic systems.
I suggest that these arguments fail to establish their intended conclusion.
My overridding theme in reviewing these arguments is that they conflate
issues concerning theoretical characterizations of computing with issues
concerning finite instantiations of theoretically decribed systems. I
then explore two arguments for treating connectionist systems dynamically.
One of these arguments settles decisively the question of theoretic
non-equivalence between connectionist systems and Turing-equivalent systems.
The second argument speaks decsively for treating connectionist systems
as finite instantiations of dynamically described systems.