NICK@AI.AI.MIT.EDU (Nick Papadakis) (06/08/88)
From: research!dlm@research.att.com
Date: Tue, 7 Jun 88 00:13 EDT
>From: allegra!dlm (D.L.McGuinness)
To: research!arpa!mc.lcs.mit.edu!ailist
Subject: Talk Announcement
______________________________________________________________________
TALK ANNOUNCEMENT
Speaker: Mark Derthick - Dept. of CS, Carnegie Mellon University
Title: Mundane Reasoning
Date: Tuesday, June 7
Time: 10:00
Place: AT&T Bell Laboratories MH 3D436
Abstract:
Frames are a natural and powerful conception for organizing knowledge.
Yet in most well-defined frame-based knowledge representation systems,
such as KL-ONE, the knowledge base must be logically consistent, no
guesses are made to remedy incomplete knowledge bases, and they
sometimes fail to return answers in a reasonable time, even for
seemingly easy queries. On the other hand are connectionist knowledge
representation systems, which are more robust in that they can be made
to always return an answer quickly, and knowledge is combined
evidentially. Unfortunately these systems, if they have a well
defined formal semantics at all, have had much less expressive power
than symbolic systems. The differing characteristics result from two
independent decisions. First, the statistical technique of Maximum a
Posteriori estimation is used as a semantic foundation rather than
logical deduction. Second, heuristic simplifications of the models
considered give rise to fast, but errorful behavior. Having made this
distinction, it is possible to use the same powerful syntax of
symbolic systems, but interpret it statistically and implement it with
a connectionist network. Although correct networks are exponentially
large, they serve as a basis from which architectural simplifications
can be made which preserve an intuitive connection to the formal theory.
The knowledge base must be tuned to alleviate errors caused by the
heuristic simplifications, so the system is intended for familiar
everyday situations in which past performance has been used for
training and in which the ramifications of wrong answers are not
serious enough to justify the exponential search time required for
provably correct behavior.
Sponsor: Ron Brachman & Deborah McGuinness (allegra!dlm)dlm@RESEARCH.ATT.COM (07/03/88)
From: dlm@research.att.com
Date: Fri, 1 Jul 88 09:10 EDT
>From: allegra!dlm (D.L.McGuinness)
To: arpa!mc.lcs.mit.edu!AIList
Subject: talk announcement
Title: Intermediate Mechanisms For Activation Spreading
or
Why can't neural networks talk to expert systems?
Speaker:Jim Hendler
University of Maryland Institute for Advanced Computer Studies
University of Maryland, College Park
Date: Tuesday, July 19
Time: 1:30
Place: AT&T Bell Laboratories - Murray Hill 3D-473
Abstract:
Spreading activation, in the form of computer
models and cognitive theories, has recently been under-
going a resurgence of interest in the cognitive science
and AI communities. Two competing schools of thought
have been forming. One technique concentrates on the
spreading of symbolic information through an associa-
tive knowledge representation. The other technique has
focused on the passage of numeric information through a
network. In this talk we show that these two tech-
niques can be merged.
We show how an ``intermediate level'' mechanism,
that of symbolic marker-passing, can be used to provide
a limited form of interaction between traditional asso-
ciative networks and subsymbolic networks. We describe
the marker-passing technique, show how a notion of
microfeatures can be used to allow similarity based
reasoning, and demonstrate that a back-propogation
learning algorithm can build the necessary set of
microfeatures from a well-defined training set. We
discuss several problems in natural language and plan-
ning research and show how the hybrid system can take
advantage of inferences that neither a purely symbolic
nor a purely connectionist system can make at present.
Sponsor: Diane Litman (allegra!diane)