[Date Prev][Date Next][Date Index]
CSLI Calendar, Wednesday, 18 September 2002, vol. 18:2
CSLI CALENDAR OF PUBLIC EVENTS
______________________________________________________________________
18 September 2002 Stanford Vol. 18, No. 2
______________________________________________________________________
A weekly publication of the
Center for the Study of Language and Information (CSLI)
Stanford University, Ventura Hall, Stanford, CA 94305-4115
http://www-csli.stanford.edu/
____________
ACTIVITIES FROM 18 SEPTEMBER 2002 TO 27 SEPTEMBER 2002
WEDNESDAY, 18 SEPTEMBER 2002
5:00pm UC Berkeley Dissertation Defense
405 Soda Hall (UC Berkeley)
Shaping and policy search in reinforcement learning
Andrew Y. Ng
UC Berkeley
Abstract below
http://www.cs.berkeley.edu/~ang/
THURSDAY, 19 SEPTEMBER 2002
4:00pm PARC Forum
George Pake Auditorium at PARC
Could we - or Should we - try to Solve Global Warming?
Stephen H. Schneider
Biological Sciences, Stanford University
http://www.parc.com/forum/
4:00pm UC Berkeley CIS Seminar
Soda Hall 310 (UC Berkeley)
Group Symmetry and Multiple View Geometry
Yi Ma
Electrical and Computer Engineering, U. Illinois at Urbana-Champaign
http://www.cs.berkeley.edu/~eyal/cis-seminar
Abstract below
THURSDAY, 26 SEPTEMBER 2002
4:00pm SRI AI Seminar Series
EJ228, SRI International
Teamwork: Practice and New Theory
Milind Tambe
University of Southern California
http://www.ai.sri.com/seminars/
Abstract below
4:00pm PARC Forum
George Pake Auditorium at PARC
Technology in Golf Equipment Design and Development
Steve Ehlers
Callaway Golf Research and Development
http://www.parc.com/forum/
4:15pm Seminar on Computational Learning and Adaptation (SCLA)
Cordura Hall, room 100
Lessons for the Computational Discovery of Scientific Knowledge
Pat Langley
CSLI
http://www-csli.stanford.edu/cll/scla.html
Abstract below
4:15pm Brain Research Center Fall Series
Munzer Auditorium, Beckman B060
Dissecting neural circuits in the visual system using targeted
cell class ablation
Sheila Nirenberg
UCLA
FRIDAY, 27 SEPTEMBER 2002
12:30pm CS547: Human-Computer Interaction Seminar
Gates B01
Video Interfaces for Entertainment
Richard Marks
Sony
http://www-pcd.stanford.edu/cs547/
Abstract below
____________
Stanford Blood Bank status: Shortage of O-, O+, A-. For an
appointment: http://bloodcenter.stanford.edu/ or call 650-723-7831.
It only takes an hour of your time.
____________
FUNDING OPPORTUNITY
A reminder of the significant funding opportunity to support research
in the general areas of information processing and cognitive science
presented by the new DARPA program in "Cognitive Information
Processing Technology". This major new program is extremely broad in
its scope, and could cover almost any research done at CSLI, and a
great deal of research under the new, broader umbrella of Media X.
The first stage in submitting a proposal is to prepare a brief outline
or white paper. DARPA will evaluate the white paper, and only if their
response is positive is a full proposal required. The deadlines are
April 4, 2003 for white papers and June 6, 2003 for full
proposals. But individuals or groups intending to submit are strongly
encouraged to submit their white paper as soon as possible, to ensure
sufficient time for preparation of a full proposal. DARPA intends to
respond to white papers as they come in.
For full details, see DARPA BAA #02-21, available at
http://www.darpa.mil/ipto/Solicitations/PIP_02-21.html
There is no pre-set maximum limit to budgets for projects in this program.
CSLI will offer assistance to Stanford individuals or groups wishing
to submit a proposal to this program through CSLI or Media
X. Interested researchers should contact Keith Devlin, Executive
Director, CSLI (devlin@csli.stanford.edu).
____________
CLASS/SEMINAR ANNOUNCEMENTS
CS 528
Broad Area Colloquium in
AI, Geometry, Graphics, Vision and Robotics
http://robotics.stanford.edu/ba-colloquium
Mondays 4:15, TCSeq 200
The Broad Area Colloquium series begins again this quarter on Monday
afternoons. This colloquium is intended to bring established and
senior researchers from the fields of AI, Geometry, Graphics,
Robotics, and Vision to discuss and explain broad considerations and
high-level tasks with which the relevant communities are dealing. The
talks are intended to create awareness and interest for all of the
members of these communities, hopefully bridging gaps and creating
collaborations.
Those who would like to receive weekly reminder announcements of the
seminar may subscribe to the mailing list
broad-area-cs-colloquium@lists.stanford.edu. To be added, send e-mail to
majordomo@lists.stanford.edu with a blank subject line and "subscribe
broad-area-cs-colloquium" in the body.
Students who wish to attend this series and receive credit (1 unit)
may enroll in cs528. Those who miss not more than one of the talks
given during the quarter will receive credit. Students who wish to
enroll in this course should join the cs528seminar mailing list to
receive relevant course information. To be added to the cs528 mailing
list, send e-mail to majordomo@lists.stanford.edu with a blank subject
line and "subscribe cs528seminar" in the body.
The first lecture of the quarter will be Monday, 9/30/02.
Check the web site for details: http://robotics.stanford.edu/ba-colloquium
Questions may be sent to bac-coordinators@cs.stanford.edu
____________
UC BERKELEY DISSERTATION DEFENSE
on Wednesday, 18 May 2002, 5:00pm
405 Soda Hall (UC Berkeley)
Shaping and policy search in reinforcement learning
Andrew Y. Ng
University of California, Berkeley
http://www.cs.berkeley.edu/~ang/
I will present two ideas that have enabled the successful application
of reinforcement learning in the domains of four-legged robot
locomotion and the control of helicopter flight.
In reinforcement learning, "shaping" refers to the important practice
of giving a learning algorithm "hints" by modifying the reward
function. While often crucial to making learning tractable, shaping
frequently changes the problem in unanticipated ways that cause poor
solutions to be learned. In this talk, I will present a theory of
shaping that shows how these problems can be eliminated, and also give
guidelines for designing good shaping functions that in practice
result in significant speedups of the learning process.
A second issue in reinforcement learning is that naive algorithms
often scale exponentially in the number of state variables, and are
thus frequently impractical. I will present PEGASUS, a method for
evaluating and finding good controllers. The key insight of this
method is that, when using a computer simulation to evaluate policies,
we can use the same set of random samples to evaluate different
policies. This leads to an efficient algorithm---one whose "sample
complexity" is polynomial, rather than exponential, in the dimension
of the problem---for which we can give strong performance guarantees.
____________
UC BERKELEY CIS SEMINAR
on Thursday, 19 September 2002, 4:00pm
Soda Hall 310 (UC Berkeley)
http://www.cs.berkeley.edu/~eyal/cis-seminar
Group Symmetry and Multiple View Geometry
Yi Ma
University of Illinois Urbana-Champaign
http://black1.csl.uiuc.edu/~yima/
In this talk, we provide a principled explanation of how knowledge in
global 3-D structural invariants, typically captured by a group action
on a symmetric structure, can dramatically facilitate the task of
reconstructing a 3-D scene from one or more than one image. More
importantly, since every symmetric structure admits a canonical
coordinate frame with respect to which the group action can be
naturally represented, the "absolute" pose between the viewer and this
canonical frame can be recovered too, which explains why symmetric
objects (e.g., buildings) provide us overwhelming clues to their
orientation and position. We give the necessary and sufficient
conditions in terms of the (group) symmetry admitted by a structure
under which this pose can be uniquely determined. We also
characterize, when such conditions are not satisfied, exactly to what
extent this pose can be recovered. We show how algorithms from
conventional multiple view geometry, after being properly modified and
extended, can be directly applied to perform such recovery, from all
"hidden images" of one image of the symmetric structure. We also apply
our results to a wide range of problems such as camera
self-calibration, global orientation for robot navigation, image based
rendering, as well as visual illusions (caused by symmetry if
wrongfully imposed).
____________
SRI AI SEMINAR SERIES
on Tuesday, 14 April 2002, 4:15pm - 5:30pm
EJ228, SRI International
http://www.ai.sri.com/seminars/
Teamwork: Practice and New Theory
Milind Tambe
University of Southern California
For the past several years, we have been conducting research in
teamwork, a critical capability in a range of multiagent domains. I
will present results from two key aspects of this research. First,
inspired by the need for team-coordination flexibility and
reusability, we have been developing STEAM/TEAMCORE, a general
teamwork model (team coordination algorithm). I will discuss the
STEAM/TEAMCORE algorithm, the associated software infrastructure, and
present results from its reuse across several different domains, e.g.,
synthetic helicopter pilot teams, RoboCup soccer teams, "Electric
Elves" personal assistant teams, recent work on human-robot teams etc.
Second, as we develop general team-coordination specifications and
algorithms, there is now a critical need for a new theoretical
framework to analyze complexity-optimality tradeoffs in competing
algorithms. To this end, I will present Communicating Markov Team
Decision Problems (COM-MTDPs), that are based on decentralized,
communicating POMDPs. COM-MTDPs provide us complexity results for key
types of teamwork domains, and allow us to compare complexity and
optimality of different coordination algorithms.
* Portions of this research have been conducted in collaboration with
members of the TEAMCORE research group ( http://www.isi.edu/teamcore ).
About the speaker: Milind Tambe is an Associate Professor of Computer
Science at University of Southern California(USC), and a project
leader at USC's Information Sciences Institute. He received his
Ph.D. from the School of Computer Science at Carnegie Mellon
University. His interests are in the areas of multi-agent systems,
specifically multi-agent teamwork, adjustable autonomy, and
distributed negotiations and he has published extensively in these
areas. A current member of the board of directors of the International
foundation for multiagent systems, he has also served on the board of
trustees of RoboCup, the Robot World Cup Federation. He is currently
on the editorial board of the Journal of Artificial Intelligence
Research (JAIR), Journal of Autonomous Agents and Multi-agent Systems
(AAMAS) and IEEE Intelligent Systems. He was also the chair of the
organizing committee for the first Americas Agents School and program
co-chair of the International conf on multi-agent systems (ICMAS)
2000.
____________
SEMINAR ON COMPUTATIONAL LEARNING AND ADAPTATION (SCLA)
on Thursday, 26 March 2002, 4:15pm to 5:30pm
Cordura 100
http://www-csli.stanford.edu/cll/scla.html
Lessons for the Computational Discovery of Scientific Knowledge
Pat Langley
Computational Learning Laboratory, CSLI
mailto:langley@csli.stanford.edu
In this talk, I review early analyses of machine learning
applications, along with more recent treatments of successful
discoveries of scientific knowledge. Although the two problem areas
have much in common, I use recent work on computational discovery in
Earth science and microbiology to illustrate some important
differences. The lessons that emerge from these efforts run counter to
some rhetorical claims and assumptions that are widespread in the
machine learning and data mining communities. For example, for many
scientific problems, it is more desirable to revise models than to
construct them from scratch, as emphasized by most data mining
researchers. Another difference is that scientific data are often rare
rather than plentiful, despite traditional claims by the data mining
community about the abundance of data. These observations and others
suggest the need to explore research paths which are quite distinct
from those that currently dominate the field.
This talk repeats material from a presentation at the ICML-2002
Workshop on Data Mining Lessons Learned. The associated paper is
available at http://www.isle.org/~langley/papers/discovery.dmll02.ps .
____________
CS547: HUMAN-COMPUTER INTERACTION SEMINAR
on Friday, 27 October 2002, 12:30-2:00pm
Gates B01
http://www-pcd.stanford.edu/cs547/
Video Interfaces for Entertainment
Richard Marks
Manager R&D Special Projects, Sony
Natural, versatile man-machine interfaces can be created by processing
live video input from a digital camera. Movements of either the user
or simple hand-held props drive an engaging entertainment
experience. The greatest level of interactivity can be produced by
mixing live video of the user with computer-generated graphics. The
low cost of digital cameras and processors has recently made such
computer vision interfaces viable, even for a cost-sensitive market
such as console gaming.
About the speaker: Richard Marks was an Avionics major at MIT before
getting his PhD at Stanford in the area of visual sensing for
underwater robotics. He then joined Teleos Research, a computer
vision start-up that was later acquired by Autodesk. He departed and
consulted for a year, before the unveiling of the PlayStation2
hardware inspired him to join PlayStation R&D. His research focus has
been studying real-time video input to the PS2, and he now manages R&D
Special Projects, which includes Man-Machine Interfaces and Physical
Simulation.
____________
END MATERIAL
The CSLI Calendar appears weekly on Wednesdays throughout the academic
year. Announcements, abstracts, and other information to appear in
the Calendar should be submitted to the editor, who reserves the right
to decide what does or does not go in the calendar
mailto:incalendar@csli.stanford.edu
Requests to be added to the mailing list should be sent to
majordomo@csli.stanford.edu. With the lines in the body of the text
of either
subscribe csli-calendar
for the long form or
subscribe csli-short-calendar
for the short form (i.e., no abstracts). Problems with subscribing or
unsubscribing should be sent to
owner-csli-calendar@csli.stanford.edu.
The full current issue is at
http://www-csli.stanford.edu/Archive/calendar/current.shtml
and the archives at
http://www-csli.stanford.edu/Archive/calendar/
People on most of the CSLI computers can type 'help csli-calendar' to
see the current issue.
The CSLI Calendar is also posted each week to
news://nntp-csli.stanford.edu/csli.bboard.
and
news://news.stanford.edu/su.events
Information about CSLI's research program is available at
http://www-csli.stanford.edu/
For maps to the Stanford University campus see
http://www.stanford.edu/home/visitors/maps.html
____________