ICML-99 Invited Talks

Learning in ACT-R

John R. Anderson, Carnegie Mellon University, Pittsburgh PA, USA

ACT-R is a unified theory of cognition that tries to satisfy Newell's 13
constraints on a theory of mind.  In addition it tries to be faithful to
the detailed empirical literature on human cognition in tasks ranging
from searching simple visual arrays to designing scientific experiments.
The basic control cycle in the ACT-R architecture involves (1)
selecting a production rule relevant to the current goal; (2) executing
the actions associated with that rule which can include retrieval of
information from a declarative memory; and (c) updating the goal with
the results of the actions and retrievals.  In addition to symbolic
learning of new productions and declarative information, critical
learning in the ACT-R theory involves a subsymbolic tuning of the
parameters that govern selection among productions and access to
declarative information.  I will describe in detail the subsymbolic
declarative learning processes and Christian Lebiere's work documenting
how they apply to the growth in arithmetic knowledge over a child's
life.  I will also describe the subsymbolic procedural learning
processes and Marsha Lovett's work documenting how they apply to
selecting problem-solving operators in a situation much like
paradigmatic reinforcement learning situations.

What can we learn from the Web?

William Cohen, Shannon Laboratory, AT&T Labs-Research, Florham Park, NJ, USA

Information on the Web is easy for a computer to access, but
frustratingly difficult for a computer to understand.  It would nice
if information from the Web could be used to automatically answer
structured queries -- in short, if the Web looked more like a
knowledge base.  Unfortunately, the information on the Web is hard to
represent with conventional knowledge-base and database formalisms:
problems with Web information include terminological differences
across sites, and the frequent interleaving of textual information
with structured, data-like information.

Over the last few years, I have developed a new "information
representation language" called WHIRL that addresses these problems by
incorporating ideas from both AI knowledge representation systems and
statistical information retrieval.  Specifically, WHIRL is a subset of
Prolog that has been extended by adding special features for reasoning
about the similarity of fragments of text.  WHIRL has many nice
properties: it strictly generalizes both logical deduction and ranked
retrieval of documents; it can be implemented fairly efficiently; and
it greatly facilitates the construction of question-answering systems
that use information found at multiple Web sites.

After introducing WHIRL I will discuss some related issues of direct
interest to the machine learning community.  In particular, I will
describe how certain "soft" inferences made by WHIRL can be evaluated
as inductive classification, and will outline ways in which WHIRL and
learning can be combined to both improve the accuracy of inference in
WHIRL, and expand the range of applicability of existing learning
algorithms.

Machine Learning Elements

J. Ross Quinlan, University of New South Wales, Sydney, Australia

How do we come to design a new ML system?  Nobody starts
completely from scratch -- we are guided by shared biases
towards certain kinds of algorithms, representations, and
heuristics.  These key concepts are the building blocks
from which new systems and approaches are synthesized,
the "elements" of the title.

This talk covers a sampler of such concepts that have
influenced many of us, including hypothesis spaces,
approaches to theory construction, tradeoffs of various kinds,
and the heuristics that drive them.  I will show why they
are important, and how they might suggest further research.