Universität Bielefeld - Sonderforschungsbereich 360
Probabilistic Methods and Learning in Behaving Systems
(This is strictly an approximate title to what will probably be a
dynamically-configured talk.)
Chris Brown
University of Rochester
Rochester, NY, USA
Systems that perceive and act seem to need several capabilities that
make a homogeneous and uniform approach to their architecture
difficult. Candidate technologies abound: Symbolic AI and its
logic-based representation and planning, the probabilistic (Bayes net
and inference diagram) version of inexact reasoning,
representation-less subsumptive or reflexive architectures that hope
for emergence of intelligence, optimization and modeling with
continuous mathematics, Hebbian learning, ``Darwinian'' evolution,
associative memory, increasingly faster processing and thus
perception, etc. etc.
Perhaps these technologies could components of a heterogeneous
architecture like Fodor's model of brain modules for special functions
like vision and speech, tied together by an associational process that
deals with their outputs. Certainly the question of learning versus
programming looms large in any practical system, since the
unpredictability of real-world situations makes it hard to anticipate
adequate responses in all circumstances. What models of learning are
appropriate? Is ``learning'' really the appropriate term for what
goes on in certain important steps of biological development?
Anke Weinberger, 1997-12-01