Page 138 - Contribution To Phenomenology

CONNECTIONISMAND PHENOMENOLOGY 131

someday would be capable of learning on their own and be less
dependent upon the machinations of an outside observer.
For most researchers in Cognitive Science and many of those attracted
to Connectionism, representation is still thought of as a relationship
between states of the system (activations of the units, in the case of
Connectionism) and objects in world, both being distinguishable as distinct
entities. The problem which then presents itself is how tell what states
represent which objects and why. The problem is similar to the one that
presents itself to Descartes. How do I know that my ideas represent
anything in the world if all I ever have access to is my ideas and never
the world? This seems to hold for cognitive systems other than human
beings as well, even though the problem there is somewhat different,
since we have no reason to beUeve that they have any self-conscious
states in which they ask whether there are is anything else other than
those states which are immediately accessible to them. But even for an
outside observer (which is a very different perspective) the problem still
poses itself about the states of the system as representations. What does
that even mean for a machine to "represent the world," for example?
And until we know that, how can we attribute "semantic import" to these
states?
Here I think a few themes borrowed from Phenomenology's confronta-
tion with Descartes' problem could help. What is it that knowers need
to do in order to be knowers from a phenomenological standpoint?
Surely, at least from the standpoint of Transcendental Phenomenology, it
cannot involve transcending the bounds of possible experience. Rather
it involves a structure of intention and fulfillment in which those things
count as genuine objects that fulfill the intentions directed toward them.
Applied to a cognitive system, that would mean that it would not have
to be constructed such that it could somehow jump out of the system,
nor that there would have to be isomorphic resemblances between parts
of the system and objects in the world, but merely that the system would
have to generate something that could count as expectations regarding
future input. What would give such a system "semantic import" would
not be its relationship to an external observer, nor the way that it maps
things completely outside the system, but rather the relationship it
establishes to future input such that this input could serve as feedback