[mod.ai] Seminar Series - Computer Science Open House

rapaport@BUFFALO.CSNET ("William J. Rapaport") (03/17/86)

            STATE UNIVERSITY OF NEW YORK AT BUFFALO

                 DEPARTMENT OF COMPUTER SCIENCE

                   GRADUATE STUDENT OPEN HOUSE

On Thursday, March 20, 1986, the graduate students of the
SUNY Buffalo Dept. of Computer Science will be presenting
an all-day conference on their recent research (most of which is
on AI).  A tech report with extended abstracts will be available;
for further information, contact James Geller (geller%buffalo@csnet-relay).

ABSTRACTS OF TALKS
  
9:00 - 9:30 
  
JON HULL, A Theory of Hypothesis Generation in Visual Word Recognition
 
An algorithm is presented that generates hypotheses about
the identity of a word of text from its image.  This
algorithm is part of an effort to develop techniques for
reading images of text that possess the human capability to
adapt to variations in fonts, scripts, etc.  This
methodology is being pursued by using knowledge about the
human reading process to direct the development of
algorithms for reading text.  The algorithm discussed in
this talk locates a set of hypotheses about the identity of
an input word (called the  neighborhood  of the input
word).  
 
Results are reported in this talk on the size of
neighborhoods for words printed in lower case that are drawn
from a large text.  Several statistical measures are
computed from subsets of a text of over 1,000,000 words and
their corresponding dictionaries.  These results show that
the average neighborhood in the dictionary of the entire
text contains only 2.5 words.  The feasibilty of this method
is also shown by experimentation with a database of lower
case word images.  The application of this approach to 8700
word images taken from 29 different fonts, in three
conditions of noise, shows that the correct neighborhood is
determined in 80% to 100% of all cases. 
 
9:30 - 10:00
  
GEORGE SICHERMAN, Databases that Refuse to Answer Queries
 
Question-answering systems must often keep certain information secret.
One way they can do this is by refusing to answer some queries.  But
if the user may be able to deduce information from the system's refusal
to answer, the secrecy of the information is broken.
 
In this talk I present a categorization of answer-refusing systems
according to what they know, what the user knows, and when the system
refuses to answer.  I also give two formal results about when the user
can deduce secrets from the system's refusals to answer, depending on
how much she knows about the system.
  
10:00 - 10:30
  
JANYCE WIEBE, Understanding De Re and De Dicto Belief Reports
		in Discourse and Narrative
 
Belief reports can be interpreted "de re" or "de dicto", and we
investigate the disambiguation of belief reports as they appear in
discourse and narrative.  In earlier work by Rapaport and Shapiro
[1984], representations for "de re" and "de dicto" belief reports were
presented, and the distinction between them was made solely on the
basis of their representations.  This analysis is sufficient only when
belief reports are considered in isolation.  We need to consider more
complicated belief structures in order to sufficiently represent "de
re" and "de dicto" belief reports as they appear in discourse and
narrative.  Further, we cannot meaningfully apply one, but not the
other, of the concepts "de re" and "de dicto" to these more complicated
belief structures.  We argue that the concepts "de re" and "de dicto"
apply not to an agent's conceptual representation of her beliefs, but
to the utterance of a belief report on a specific occasion.  A
cognitive agent interprets a belief report such as `` S believes that
 N  is  F '', or `` S  said, ` N  is  F ' '' (where  S  and  N are
names or descriptions, and  F  is an adjective) "de dicto" if she
interprets it from  N 's perspective, and "de re" if from her own.

10:45 - 11:15
  
MINGRUEY TAIE, Device Representation Using Instantiation Rules
		and Structural Templates
 
A device representation scheme for automatic electronic device fault 
diagnosis is described.  Structural and functional descriptions of 
devices (which are central to design-model-based fault diagnosis) are 
represented as instantiation rules and structural templates in a 
semantic network.  Device structure is represented hierarchically to 
reflect the design model of most devices in the domain.  Each object 
of the device hierarchy has the form of a module.  Instead of 
representing all objects explicitly, an expandable component library 
is maintained, and objects are instantiated only when needed.
The component library consists of descriptions of component "types"
used to construct devices at all hierarchical levels.  Each component
"type" is represented as an instantiation rule and a structural
template.  The instantiation rule is used to instantiate an object of
the component "type" as a module with I/O ports and associated
functional descriptions.  Functional description is represented as
procedural attachments to the semantic network; this allows the
simulation of the behavior of objects.  Structural templates describe
sub-parts and wire connections at the next lower hierarchical level of
the component "type".  Advantages of the representation scheme are
compactness and reasoning efficiency.

11:15 - 11:45

JAMES GELLER, Towards a Theory of Visual Reasoning

Visual Knowledge Representation has not yet found the treatment it
deserves as its own subfield of AI.  Visual reasoning is fundamentally
different from predicate calculus type logical reasoning and is of
central importance for the field of Visual Knowledge Representation.  A
systematization of different types of visual reasoning requires the
differentiation between purely geometrical reasoning and different
types of knowledge-based reasoning.  Knowledge-based reasoning in turn
can use knowledge about the world, knowledge about abstract
hierarchies, or knowledge about normality.  Research on visual
knowledge is directly applicable to graphics interface design for
intelligent systems.  The VMES maintenance expert system for circuit
board repair uses such a user interface which is designed in analogy to
a language generation program.

1:15 - 1:45
  
MICHAEL ALMEIDA, The Temporal Structure of Narratives
  
Narratives are a type of discourse used to describe sequences of
events. In order to understand a narrative, a reader must be able to extract
the ``story'', that is, the described events and the temporal relations
which hold between them, from the text. Our principle research goal has been
to develop a system which can read a narrative and produce a model of
the temporal structure of its story.
 
The principle heuristic used in constructing such a model is the
Narrative Convention:  unless we are given some signal to the contrary,
we assume that the events of the story occurred in the order in which
they are presented in the text. In addition, however, a reader must deal
with: (1) tense - in a standard past tense narrative the principle
distinction is between the past and the past perfect tenses, (2) aspect -
the distinction between events viewed perfectively or imperfectively,
(3) aspectual class - the intrinsic temporal properties of various
types of events, (4) time adverbials - these can be used to place
events within various calendrical intervals, give their durations,
or relate them directly to other events, and to some extent (5)
world-knowledge.
  
1:45 - 2:15
  
WEI-HSING WANG, A Uniform Knowledge Representation for Intelligent CAI Systems
  
In examining the current situation of Computer Aided Instruction
(CAI), we find that Intelligent CAI (ICAI) and its authoring system are 
necessary.  By studying the knowledge representation methods and expert
system concepts, we choose a frame representation method to construct
an Intelligent Tutor, called ITES.  We show that a frame can be used
to represent knowledge in semantic nets, procedures and production 
rules.  Furthermore, this method is very convenient in authoring 
system creation.

2:15 - 2:45
  
RICK LIVELY, Semantics for Abstract Data Types
  
An abstract data type is  often defined as a
pair < A ,  S >, where  A  is a set (of objects) and
 S  is a set of operations defined on cartesian
products of the types of the objects.  Axiomatic
methods are used to develop specifications for
the defined data type.
 
Semantics for abstract data types have
been treated by Adj using initial algebras, and
by Janssen (inspired by Montague semantics)
using many-sorted algebras.  A comparison
is made of the mathematical properties
and applicability to computer science of
these approaches.

3:00 - 3:30
  
SCOTT CAMPBELL, Using Belief Revision to Detect Faults in Circuits
  
To detect faults in electrical circuits, 
programs must be able to reason about whether
the observed inputs and outputs are consistent 
with the desired function of the circuit.
The SNePS Belief Revision System (SNeBR) is designed to reason about 
the consistency of rules and hypotheses defined within a particular 
context or belief space.
This paper shows how belief revision can be used for fault detection 
in circuits, and so leads to a unification of the fields of belief 
revision (also known as truth maintenance) and fault detection.

3:30 - 4:00
  
DOUGLAS H. MacFADDEN, DUNE: A Demon Based Expert System Architecture
			for Complex and Incompletely Defined Domains
  
Traditional expert system architectures use the rule (an `` if ...
then ''  data structure) as the primary unit of knowledge.  The primary
unit of knowledge in the DUNE system architecture is the demon.  Each
DUNE  demon  is  an  individual processing element that can contain a
variety of types of data and can perform a variety of  operations  on
its  data.   Each demon can communicate with any other demon or with
the user via messages.  Typical data for these demons may be a 
traditional type rule, a list of weight values for the features in the
left-hand-side of the rule, an (English) description of each feature,
a  list  of  related  demons,  etc.   Typical operations that these
demons may perform are: calculating the ``closeness'' of  the  rule  to
firing,  calculating the most important feature of the rule yet to be
resolved, telling the system  to  not  consider  this  demon  anymore
(entering a sleep state), telling other demons (and the user) that the
demon is either satisfied or will never be satisfied, etc.
 
We hope to show that  these  features  of  DUNE  demons  can  be
exploited  to  express the knowledge of many expert domains that have
proven unfeasible to traditional expert system architectures.

4:00 - 4:30
  
JOYCE DANIELS, Understanding Time and Space in Narrative Text
  
The Graduate Group in Cognitive Science at SUNY at Buffalo is an 
interdisciplinary group of faculty and graduate students.  Participants
in the group's activities come from over seventeen departments within
the university and local colleges in Western New York and Canada.
There are six core faculty and 
their graduate students, comprising a standing research group investigating
how we understand movement through time and space in narrative text.
This research addresses both the general issue of how time
and space are expressed in language, and specific individual disciplinary
interests such as identifying the exact lexical items signaling movement;
developing experiments to collect data on the 
psychological validity of the supposed influence of suspected lexical items;
examining the problems encountered by speech pathologists when a client
cannot understand spatial or temporal concepts in language; and
artificial intelligence program models of human and linguistic data on
the SNePS network.
 
Research conducted by group members has resulted in the identification
of what we term the ``Deictic Center'' (DC).  This contains a WHO-point,
a WHEN-point, and a WHERE-point.  It is the locus of a 
particular point in conceptual space-time.
We will explain the significance of the DC concept in greater detail.
and present some results of our linguistic and psychological
investigation.