Boulder Computational Learning Group
Summer 1998 Schedule
Unless otherwise noted, meeting time is on Mondays, 1-3 p.m.,
in ECOT 831a (Engineering center, 8th floor, small conference
room). Dates in red are at the usual time and place.
Dates in blue are special meeting dates.
Schedules from previous semesters
June 19, 10:30, NCAR, Foothills Laboratory, Bldg 2, Room 1001: Tosh Munakata, Cleveland State
The State of AI: A Practical Perspective
The speaker has been active in the areas of practical applications
of AI. For example, he served as Guest Editor for two well-
received special issues on commercial and industrial AI for the
Communications of the ACM (March 1994 and November 1995), and is
currently the Chair of the Rough Control Group. Based on his
extensive experience, Dr. Munakata will present a seminar that
consists of the following two parts:
Part I. Macroscopic overview of the current state and future
perspectives of practical AI.
This part includes a quick review of the areas the CACM issues
cover, some common characteristics of practical AI, and some
forecasts for the near and distant future.
Part II. A Tutorial on Rough Sets and Rough Control (if time
allows)
A major feature of rough set theory in terms of practical
applications is the classification of empirical data and subsequent
decision making. There has also been much recent interest in rough
control: applications of rough set theory to control problems.
July 20, 3p.m., ECOT 831: Nicol Schraudolph, IDSIA
Accelerated Gradient Descent by Factor-Centering Decomposition
Gradient factor centering is a new methodology for decomposing neural
networks into biased and centered subnets which are then trained in
parallel. The decomposition can be applied to any pattern-dependent
factor in the network's gradient, and is designed such that the subnets
are more amenable to optimization by gradient descent than the original
network: biased subnets because of their simplified architecture,
centered subnets due to a modified gradient that improves conditioning.
The architectural and algorithmic modifications mandated by this
approach include both familiar and novel elements, often in prescribed
combinations. The framework suggests for instance that shortcut
connections --- a well-known architectural feature --- should work best
in conjunction with slope centering, a new technique described herein.
Our benchmark experiments bear out this prediction, and show that
factor-centering decomposition can speed up learning significantly
without adversely affecting the trained network's generalization ability.