Page 128 - Contribution To Phenomenology

CONNECTIONISMAND PHENOMENOLOGY 121

future based merely on similarities observed about their performances in
the past.'
One of the most striking features of connectionist systems is their
response when presented with problems of what CogSci folks call
"categorization," i.e., the recognition of something in terms of the class
of things it belongs to or the subsumption under concepts, and pattern
recognition. These responses do not fit into a framework of concepts in
terms of necessary and sufficient conditions for belonging or failing to
belong to a class. To take a simple example, consider a network where
activation of a particular unit represents recognition of something as
belonging to the category represented by that unit. The unit gets input
from a number of different connected units, which might represent
features of the object. Differences in weights from different input units
can correspond to differences in importance of the various features.
Whether a unit turns on might also depend upon its reaching a particular
threshold, so very different combinations of input might lead to "recogniz-
ing" an object as belonging to that category. This makes it possible to
recognize things not only as simply "in" or "out," but rather also
"typical," "atypical," or somewhere in-between.
Not all features end up counting equally, some end up counting much
more strongly than others, and there may not be any one feature or even
a specific set of features that is essential in the sense that the system will
consider its presence absolutely necessary in order to identify an object
as being a member of the class of things that usually exhibit that feature.
Thus, there is no strict essence of a thing, but rather cores of features
around which individuals gather. This does not imply that a connectionist
system could not be made to conform to a more rigid conceptual logic.
Rather, like most human beings, they do not naturally tend to do so.
That adds a flexibiUty for dealing with less than complete information
and for equivocal situations, but introduces a degree of indeterminacy and
ambiguity in the way that they are likely to classify specific objects.

' I am indebted to my colleague Terence Horgan for pointing this out to me (with
the usual qualifier about eventual mistakes in presenting it accurately not being
attributable to him, but to my having my weightings wrong when he presented me with
what should normally represent the proper input needed to allow one to produce a
correct version of it). One can get an idea (provided one has the proper initial settings)
about how this bears on explanation in the human and social sciences from Terence
Horgan and John Tienson, "Soft Laws," Midwest Studies in Philosophy, 14 (1990): 256-
279.