Page 139 - Contribution To Phenomenology

132 TOMNENON

about the correctness or incorrectness of the response generated in
response to past input.^
This suggestion does not entail that cognitive scientists all necessarily
have to embrace a version of transcendental philosophy (although I think
they should, and not only for the sake of their machines and models) in
order to accomplish this, since adopting a transcendental standpoint
involves far more than what I have just suggested. It would, however,
point the way to some other practical advances. Once one begins to
think of representation no longer as the problem of how to compare
parts of the system to things outside of it, one might also be able better
to conceive what learning would be and how this could become part of
the architecture. (Although for the purposes of us outside the machine
who want to know what it is doing and gain information from it, we will
certainly want to continue to make some such comparisons or assign
interpretations from the outside until part of its output is put into a
conventional language we can understand.) Learning would then simply
involve shifting the weights whenever the received input did not
correspond to what was expected under such circumstances. Representa-
tion would be the representation of some specific input under certain
conditions, and learning would be the system's self-adjustment so that the
two eventually come to coincide more closely. This would remove the
necessity for an external operator to constantly readjust the machine
according to a learning algorithm outside the system in the way that this
is currently done using techniques such as back-propogation.^^ The point
holds generally for any cognitive system, whether designed in a classical
way or according to a connectionist model, but at least as far as learning
is concerned, the prospects for a connectionist system are rosier, since the
flexibility of a connectionist system should in principle be better able to
accommodate the inherent vagueness of such terms as "under the proper
circumstances" or "similar input" which will certainly never able to be
completely eliminated. In this case also, what I am offering in this paper

^ In a number of places, e.g., in §8 of Experience and Judgment, Husserl explicitly
identifies anticipation as one of the key features of consciousness as intentional.
^^ If this is indeed theoretically possible, this would also have a direct bearing on
Evans's criticism that connectionists systems are just as essentially as insensitive to context
as traditional AI has been (121-22), since the changes would not be enacted by an outside
manipulator, but by the system and its interaction with its environment, according to which
would change itself in response to discrepancies between its "expectations" and the input
actually submitted to it from the outside.