Carnegie Mellon University
Department of Computer Science
AI Seminar 99/00 Schedule
Time: Tuesday 3:30-4:30pm
Place: Wean Hall 5409
The AI seminars are open to the public and will
be held on most Tuesdays at 3:30pm in Wean Hall, Carnegie Mellon University.
Special AI seminars can be arranged for visitors on dates other than Tuesdays.
The schedule will be updated daily. To volunteer to give an AI seminar
or to nominate an outside speaker, contact Dr.
Tai Sing Lee at tai@cs.cmu.edu.
All AI seminars will be posted to both cboards
and bboards (cs & robotics). Meanwhile, if you would like to/not
to be reminded by email, please subscribe/unsubscribe ai-seminar-announcements
mailing list by sending email to Michael.Lewicki@cs.cmu.edu.
Simply put "Subscribe/Unsubsribe ai-seminar-announcements" in
the subject of your message.
10/05/99 - Xiaolan Zeng
Department of Electrical Engineering, Yale University.
3D Image Segmentation Using a Generic Shape Constraint with Applications to Cortical Shape Analysis
A novel approach has been developed in this thesis for the problem of
segmenting volumetric layers, a type of structure often encountered in
medical image analysis. This approach is aimed towards the use of
structural information to enhance the performance of the segmentation
process. While some organs have a more consistent global shape and can be
characterized using a specific shape model, other anatomical structures
possess much more complex shape with possibly high variability which needs
a more generic shape constraint. The three-dimensional(3D) nature of
anatomical structures makes necessary the use of volumetric approaches
that exploit complete spatial information and therefore are far superior
to the non-optimal and often biased 2D methods. Our method takes a
volumetric approach, and incorporates a generic shape constraint -- in
particular, a thickness constraint. The resulting coupled surfaces
algorithm with a level set implementation not only offers segmentation
with the advantage of minimal user interaction, robustness to
initialization and computational efficiency, but also facilitates the
extraction and measurement of many geometric features of the structures of
interest.
The algorithm was applied to 3D Magnetic Resonance (MR) brain
images for skull-stripping, cortex segmentation and various feature
measurements including cortical surface shape and cortical thickness.
Validation of the model was done through both synthetic images with
"ground truth" and a wide range of real MR images with expert tracing
results. As a natural follow up of the segmentation work, a new approach
was developed for the extraction of sulcal ribbon surfaces which are
distinctive sulcal landmarks of the brain. This effective and efficient 3D
method of sulcal ribbon extraction has potentials in a variety of
applications such as the automatic parcellation of cortical regions and
the geometric-constrained brain atlas building. The tools of cortical and
sulcal shape analysis developed in this work are of great importance to
studies of neuroanatomy through medical imaging, and are bringing about
new understanding of brain anatomy and function.
Xiaolan Zeng
received B.S.E.E. from University of Science and Technology
of China in 1995. She has been working toward her Ph.D. degree in the
Department of Electrical Engineering at Yale University since 1994, and
has just defended her thesis. Her research interest includes medical image
processing and computer-assisted intervention.
10/12/99 - Andrew Moore CMU Robotics Institute, CALD, and Schenley Park Research, Inc.
Inner-Loop Statistics in Automated Scientific Discovery
Intensive statistical analysis of massive data sources ("data mining")
has been embraced as one of the final areas with a need for massive
computation beyond that available on a $2000 computer or $200
videogame. We begin this talk with two examples of software, instead
of hardware, giving 1000-fold speedups over traditional
implementations of statistical algorithms for prediction, density
estimation, and clustering.
We then pause to examine directions in which these software solutions
when faced with Physics, Biology and commercial scientific data seemed
blocked by a curse of dimensionality and limitations on machine main
memories. This is followed by four examples of new pieces of research
that circumvent these barriers: Komarek's lazy cached sufficient
statistics, Pelleg's exact accelerated k-means, multiresolution
ball-trees for very high dimensional real-valued data, and Gordon's
filament identifier.
We then reveal the reason for our new-found respect for
super-computation: when an algorithm you previously ran overnight
executes in seconds, you find yourself wanting to run it ten thousand
times. We show the impact of being able to run intensive statistics as
an inner loop has had on our analysis of cosmology data (preliminary
data from the Sloan Digital Sky Survey) and biotoxin identification,
where desirable but hopelessly extravagant operations such as model
selection, bootstrapping, backfitting, randomization and graphical
model design now become somewhat non-hopeless.
* Joint work with Andy Connolly, Geoff Gordon, Paul Komarek, Bob Nichol,
Dan Pelleg and Larry Wasserman
Andrew
Moore received a Phd in Computer Science from the University of Cambridge
in 1991 (thesis topic: Robot Learning). He worked as a post-doc with Chris
Atkeson at MIT AI lab for three years before joining CMU as an assistant
professor in Robotics/CS in 1993. His research interests include: Autonomous
learning systems for manufacturing, efficient algorithms for machine learning
and reinforcement learning, finite production scheduling, and machine learning
applied to optimization.
10/19/99 - Pat Langley
Institute for the Study of Learning and Expertise, 2164 Staunton Court, Palo Alto, CA 94306.
The Computer-Aided Discovery of Scientific Knowledge
Contrary to common opinion, AI research on computational discovery
has already led to the discovery of new knowledge in a variety of
scientific domains, and in this talk I review progress on this
front. I begin by characterizing five historical stages of the
scientific discovery process, which I then use as an organizational
framework to describe applications. I also identify five distinct
steps during which developers or users can influence the behavior
of a computational discovery system. Rather than criticizing such
intervention, as done in the past, I recommend it as the preferred
approach to using discovery software. As evidence for the advantages
of such human-computer cooperation, I report some examples of novel,
computer-aided discoveries that have appeared in the scientific
literature, along with the role that humans played in each case. I
close with a proposal that future systems provide more explicit
support for human intervention in the discovery process.
Pat Langley
Dr. Pat Langley's research focuses on machine learning and knowledge
discovery. He has published over 100 papers on this topic and related
aspects of AI, he has edited or authored five books in the area,
including the textbook, Elements of Machine Learning, and he was the
founding editor of the journal Machine Learning. Dr. Langley's work
has contributed to methods for rule induction, probabilistic learning,
and case-based reasoning, and he has applied these techniques to a
variety of problem areas. His current research emphasizes adaptive
user interfaces, which invoke machine learning to construct user models
based on interaction with their users. Dr. Langley received his PhD
from Carnegie Mellon University in 1979, and he has worked in academia,
in government, and in industry. He currently serves as Director of
the Institute for the Study of Learning and Expertise, as Head of
the Adaptive Systems Group at the Daimler-Benz Research & Technology
Center, and as a Consulting Professor at Stanford University.
10/21/99 - Toby Walsh
Division of Informatics, The University of Edinburgh.
Problems on the Knife-edge
In recent years, there has been an explosion of research
in so-called "phase transition" behaviour in search, especially
in the area of propositional satisfiability (or SAT). Problems that
are under-constrained tend to be relatively easy to solve as you can
easily stumble across a solution. On the other hand, problems
that are over-constrained are also often easy to solve as the
bottlenecks tend to be very obvious. The hardest problems tend to
be in between when problems are on the "knife-edge". I will
survey work in this area, and show how insights gained from such
analysis can be used to inform the design of algorithms.
Toby Walsh
has a PhD and MSc from the Dept of AI (Edinburgh),
and an MA in theoretical physics and mathematics from St John's
College (Cambridge). He is currently an EPSRC Advanced Research Fellow
at the Dept of CS (York). He has previosuly held postdoctoral positions
at the Universities of Edinburgh and Strathclyde, as well as the
research labs of INRIA (Nancy) and IRST (Trento). His research interests
span a range of topics in automated reasoning, but most concern search
and representation. He is editing a special issue of the Journal of Automated
Reasoning entitled "SAT2000: satisfiability at the start of the year 2000".
11/09/99 - John Lafferty
School of Computer Science, CMU
Three Rivers: Graphs, Coins, and Codes
The past five years have seen spectacular progress on the problem that
Claude Shannon formulated fifty years ago: devise codes that can
communicate efficiently over a noisy channel. Shannon used random
coding arguments to show the existence of codes that will do the job;
yet good codes and practical decoding algorithms have proven very
difficult to find. The recent progress has come from thinking of
codes as graphs, building random graphs with special structure, and
using iterative decoding algorithms that are closely related to the
technique of "belief propagation" in the AI literature.
One important graphical representation of an error-correcting code is
the minimal trellis. This representation enables maximum likelihood
decoding, using statistical inference algorithms originally developed
for hidden Markov models, but with exponential lower bounds on
complexity. More efficient algorithms require graphs with cycles.
Great empirical and some theoretical success has been attained
recently using random, sparse bipartite graphs. We can also dispense
with the coins and get excellent performance from certain classes of
codes defined on algebraically constructed graphs.
In this talk we will give an introduction to this recent work, and
discuss those aspects that are promising for new applications. As an
example, we'll show how iterative decoding techniques can be used to
solve instances of a class of hard satisfiability problems.
Ultimately, our aim is to gain insight from coding theory that will
lead to more powerful statistical learning and inference methods based
on graphical probabilistic models.
John Lafferty Faculty member in the
Computer Science Department and Language Technologies Institute.
Affiliated faculty member of the Center for Automated Learning and Discovery, and
the Program in Algorithms, Combinatorics and Optimization.
Research interests: Statistical learning algorithms, language modeling,
information theory, codes and graphical models. Applications to natural language processing,
information retrieval, digital libraries.
11/16/99 - Masaru Tomita
Laboratory for Bioinformatics, Keio University, Japan. Adjunct faculty, SCS, CMU.
E-CELL: Towards Integrative Simulation of Whole Cell Metabolism
The E-CELL project was launched in 1996 at Keio University in order to
model and simulate various cellular processes with the ultimate goal
of simulating the cell as a whole. The first version of the E-CELL
simulation system, which is a generic software package for cell
modeling, was completed in 1997. The E-CELL system enables us to
model not only metabolic pathways but also other higher-order cellular
processes such as protein synthesis and membrane transport within the
same framework. These various processes can then be integrated into a
single simulation model.
Using the E-CELL system, we have successfully constructed a virtual cell
with 127 genes sufficient for ``self-support'' (ref. Science 284, 2 Apr
1999, pp.80). The gene set was selected from the genome of Mycoplasma
genitalium, the organism having the smallest known genome. The set
includes genes for transcription, translation, the glycolysis pathway for
energy production, membrane transport, and the phospholipid biosynthesis
pathway for membrane structure.
The E-CELL system has been made available for beta testing from our
website http://www.e-cell.org.
Masaru Tomita
1981 B.S in Mathematics, Keio University
1983 M.S in Computer Science, Carnegie Mellon University
1985 Ph.D in Computer Science, Carnegie Mellon University
1985-1994 Research Associate, Assistant Professor, Associate Professor Department of Computer Science, Carnegie Mellon University; Associate Director, Center for Machine Translation, Carnegie Mellon University.
1988 Presidential Young Investigators Award (National Science Foundation)
1998 Recieved another Ph.D in Molecular Biology from Keio University.
Present Professor, Laboratory for Bioinformatics, Department of Environmental Information, Keio University; Adjunct Professor, School of Medicine, Keio University;
Research Fields: Bioinformatics, Genome Informatics, Theoretical Molecular Biology, Natural Language Processing, Artificial Intelligence
11/23/99 - Stephen Robertson
Microsoft Research, St George House, 1 Guildhall Street, Cambridge CB2 3NH, UK
A Little Theory, Some Intuition and A Lot of Experiments:
Developing Probabilistic Models for Information Retrieval
The probabilistic approach to information retrieval has proved itself to be
a significant and valuable way of formulating and theorizing about problems
in IR. It has provided some valuable insights, concepts, and arguments,
which have contributed substantially to the state of the art. In addition,
it appears to provide a sound theoretical basis for IR models. However, in
common with other models, it fails to answer many questions, or provides
only partial answers. In many cases, indeed, it does little more than
suggest answers. There are several reasons for this; one is that it simply
has little to say about some of the phenomena which affect retrieval. For
example, while it may be the case that some linguistic phenomena could be
described in probabilistic terms, there are other aspects of language which
seem much less susceptible to probabilistic modeling. Some of these
questions (not all) may be answered by experimental means.
The combination of probabilistic models and experimentation can be more
than the sum of its parts -- theoretical modeling and experimentation can
be seen to feed on and contribute to each other in the best scientific
tradition. However, it sometimes seems that our approach to experimental
evaluation is antithetical to theory.
In this talk, I will discuss the experiences of some years' work in this
area. These experiences are mainly associated with City University London
and the team which developed the Okapi experimental retrieval system, as
well as with the huge international experimental programme known as TREC.
I will also discuss some recent trends in probabilistic IR and the variety
of ways of looking at retrieval probabilistically.
Stephen Robertson
First degree in mathematics from Cambridge; masters in information science
from City University; doctorate from University College London, with BC
Brookes (all a very long time ago now!).
Researcher at Aslib for five years, then held a research fellowship at
University College London. Began collaborations with Karen Sparck Jones and
Nick Belkin at this time. Then returned to City University. Three months on
a Fulbright at UC Berkeley (collaborated with Bill Cooper and Bill Maron).
Started the Centre for Interactive Systems Research at City, and built a
research group with a strong focus on the design and evaluation of
information retrieval systems (including Micheline Beaulieu and Stephen
Walker, and the Okapi system). Also head of department of information
science during part of this time.
Major break in 1998, when Microsoft asked me to join their new research lab
in Cambridge. Here I am building a new team, under rather different
conditions. I retain a part-time position at City.
Last year the Institute of Information Scientists awarded me the Tony Kent
STRIX award for good behaviour (well, something like that).
01/18/00 - Tommi S. Jaakkola MIT AI Lab.
Maximum Entropy Approach to Classification with Incomplete
Labels and Other Discrimination Problems
I will present a maximum entropy approach to discriminative
classification that includes problems with uncertain, missing, or
structured labels. The formalism also accomodates other seemingly
unrelated problems such as anomaly detection and feature selection
that are essential in many application domains. The maximum entropy
approach can be cast in terms of regularization theory (in the space
of distributions over parameters) and is fundamentally driven by large
margin classification. I will characterize how this approach relates
to, e.g., boosting and support vector machines. While the maximum
entropy formalism is generally applicable, it is not always
computationally feasible. Simple variational approximation methods can
be employed, however, and they appear to yield most of the benefit. I
will present experimental results to illustrate both the strengths and
limitations of our approach.
This is in part joint work with Marina Meila (CMU) and Tony Jebara
(MIT Media Lab).
Host: John Lafferty (Please contact Barbara Sandling x8-8860
for appointment with Dr. Jaakkola.)
Tommi S. Jaakkola
Research interests:
Statistical inference and estimation in the context of graphical models and the development of principled approximation methods such as variational methods
for problems with limited computational resources. Representation in machine learning in general. Theoretical analysis and development of algorithms for
regression/classification problems as well as for temporally extended decision problems. Recent interests include kernel methods and their combination with
graphical models. Primary application areas are molecular biology (protein sequence analysis, gene identification), large scale (medical) diagnosis problems,
and currently to a lesser degree problems in vision and speech.
01/25/00 - Morton Ann Gernsbacher University of Wisconsin-Madison
Imaging Higher-Level Cognition: Insights from Discourse Processing
In this talk I shall report two studies that used functional brain imaging
(FMRI) to explore the cognitive processes involved in comprehending
connected discourse. The data from these two studies are used to illustrate
two conclusions and a caveat: A picture is worth a thousand milliseconds;
one can find hay in a haystack; and a lesson should be learned from
drinking Scotch and water, whiskey and water, and rye and water.
Host: Jill Fain (Please contact contact Jill Lehman jef@cs.cmu.edu
for appointment with Dr. Jaakkola.)
Morton Ann Gernsbacher,
Sir Frederic C. Bartlett Professor
University of Wisconsin-Madison
1202 W. Johnson Street
Madison, WI 53706-1611
(608) 262-6989 [fax (608) 262-4029]
MAGernsb@facstaff.wisc.edu
??/??/?? - Alex Waibel
Professor, Computer Science; Director, Interactive Systems Laboratories; C
arnegie Mellon University
02/15/00 - Yiming Yang Carnegie
Mellon University
Combining Classifiers for Better Prediction in Event Detection and Tracking
Topic detection and tracking (TDT) from chronologically ordered
document streams presents new challenges for statistical text
classification. Parameter tuning through cross-validation becomes very
difficult when the validation set contains no examples of new classes
(events) in the evaluation set. We address this problem by selecting
a set of classifiers with different performance characteristics and
combining their output to make joint classification decisions. Using
an early-time TDT benchmark corpus for validation and a later one for
evaluation, we obtained a 38-65% error reduction in event tracking by
combining k-Nearest Neighbor, Rocchio and Language Modeling
classification schemes instead of applying each single method alone.
We also observed improved performance in event detection by combining
document clustering systems based on different similarity metrics.
Yiming
Yang is an Associate Professor at Language Technologies Institute and Computer
Science Department, Carnegie Mellon University
. She received her
B.S. degree in Electrical Engineering, and her M.S. and Ph.D. degrees
in Computer Science (Kyoto University, Japan). Her research has
centered on statistical learning approaches to text categorization,
retrieval, clustering, summarization, multimedia and translingual
information access, and intelligent search in digital libraries and
Internet environment. Dr. Yang received the Academy Prize of
Information Processing Society of Japan in 1985, the Best Theoretical
Paper Award at the Annual Symposium on Computer Application in Medical
Care in 1993 and 1994, and the Distinguished Paper Award at the
International Joint Conference of Artificial Intelligence in 1997.
03/07/00 - Robert F. Murphy Carnegie Mellon University
Computational Challenges in Cell and Molecular Biology
There has been explosive growth in biological knowledge over the past
twenty years, as challengesof increasing complexity have been met.
Currently, significant effort is directed towards mapping and sequencing
the genomes of animals and humans. In parallel, individual biologists
continue to study individual genes, proteins, and metabolic pathways. One
of the major challenges of the next decade will be to make use of
information from genome sequencing to study the macromolecules and
biochemical reactions in cells and tissues on a much larger scale. In the
first part of my talk, I will discuss challenges being faced in analyzing
the expression of genes and proteins , the determination of
proteinstructures, subcellular locations, and enzymatic activities, and the
large scale modeling of complex biological systems. I will discuss current
research problems being addressed by faculty in the Department of
Biological Sciences that have significant computational issues associated
with them.
In the second part, I will discuss particular computational research that
my group has carried out and our plans for the future. I am particularly
interested in methods for determining and predicting the subcellular
location of proteins. Proteins are often specialized to carry out their
function in a particular subcompartment (or organelle) in a cell. The
location or locations within a cell that a particular protein is found in
thus has profound implications for understanding its function. While some
preliminary efforts towards predicting protein location from primary
sequence have been made, current prediction accuracy is hindered by the
absence of a sound systematics of protein location. Information on a given
protein's subcellular location can be determined by fluorescence
microscopy, but there has been no careful analysis of the range of possible
subcellular patterns that can be displayed (even in a single cell type).
Our group has therefore developed new numerical features for describing and
classifying protein localization patterns, and explored the feature space
using a variety of classification methods. These methods will be applied
to many different proteins as part of a large new NIH-funded project. The
approach we have used, its extension to three-and four-dimensional images,
and its implications for creation of multimedia knowledge bases
(incorporating information from on-line full-text journals) will be
discussed. Offshoots of this approach include the ability to objectively
select one or more representative images from an experiment for publication
or presentation and to automate interpretation of microscopy experiments.
Robert F. Murphy, Ph.D.
Associate Professor of Biological Sciences
Undergraduate Research Advisor, Department of Biological Sciences
Director, Biological Sciences Summer Undergraduate Research Program
Program Leader, Undergraduate and Graduate Education,
Science and Technology Center for Light Microscope Imaging
Carnegie Mellon University
03/14/00 - Michael Collins AT & T
Discriminative Reranking for Natural Language Parsing
Machine-learning approaches to natural language parsing have recently
shown some success in complex domains such as newswire text. Many of
these methods fall into the general category of history-based
models, where a parse tree is represented as a derivation (sequence of
decisions) and the probability of the tree is then calculated as a
product of decision probilities. While these approaches have many
advantages, it can be awkward to encode some constraints within this
framework.
This paper considers approaches which rerank the output of an existing
probabilistic parser. The base parser produces a set of candidate
parses for each input sentence, with associated probabilities that
define an initial ranking of these parses. A second model then
attempts to improve upon this initial ranking, using additional
features of the tree as evidence. The strength of our approach is that
it allows a tree to be represented as an arbitrary set of features,
without concerns about how these features interact or overlap, and
without the need to define a derivation which takes these features
into account.
The problems with history-based models, and the desire to be able to
specify features as arbitrary predicates of the entire tree, have been
noted before. In particular, previous work (Abney 97; Della Pietra,
Della Pietra and Lafferty 97; Johnson 99) has investigated the use of
Markov Random Fields (MRFs), or log-linear models, as probabilistic
models for parsing and other NLP tasks. The first method we discuss is
based on a feature selection method within the MRF framework. The
second approach is based on the application of boosting models for
ranking problems (Freund et al 98). The boosting method gives a 13%
relative decrease in error rate over an existing parser.
Michael Collins
Email: mcollins@research.att.com
Phone: +1 973 360-8349
Fax: +1 973 360-8970
Room: A253
Address: 180 PARK AVE, P.O. BOX 971, FLORHAM PARK, NJ
07932-0000 UNITED STATES
04/18/00 - Michael Kearns
AT & T Labs
A Reinforcement Learning Dialogue System
Spoken dialogue systems communicate with users via automatic
speech recognition (ASR) and text-to-speech (TTS) interfaces,
and mediate the user's access to a back-end database. Designers
of such systems face a number of nontrivial choices in dialogue
strategy, including user vs. system initiative (the choice between
soliciting relatively open-ended vs. constrained user utterances),
and choices in confirmation strategy (when to confirm or re-prompt
for an ambiguous utterance). System design has typically been done
in an ad-hoc manner, with subsequent improvements to dialogue
strategy being fielded sequentially.
In this work, we apply the formalism of Markov decision processes
(MDPs) and the algorithms of reinforcement learning to the problem
of automated dialogue strategy synthesis. In this approach, an MDP
is built from training data gathered from an initial "exploratory"
system. This MDP provides a state-based statistical model of user
reactions to system actions, and is used to simultaneously evaluate
many dialogue strategies and choose the apparent optimal among
them. At AT&T Labs, we have applied this methodology in a dialogue
system for accessing a database of information on activities in New
Jersey, and have run controlled user experiments to evaluate the approach.
In this talk, I will describe our results, which include statistically
significant improvements in system performance, and discuss the issues
we faced in making the methodology work.
This talk describes joint work with Satinder Singh, Diane Litman, and
Lyn Walker of AT&T Labs.
Michael Kearns did his undergraduate studies at the University of California at Berkeley in math and
computer science, graduating in 1985. He received a Ph.D. in computer science from Harvard University in 1989;
the title of his dissertation was The Computational Complexity of Machine Learning, and Prof. L.G. Valiant was his advisor. Following postdoctoral positions at the Laboratory for
Computer Science at M.I.T. and at the International Computer Science Institute in Berkeley, in 1991 he joined the
research staff of AT&T Labs.
05/01/00 - Maja J Mataric'
University of Southern California
Robot Teams and Humanoids on Their Best Behavior:
Principled Behavior-Based Control and Learning
Behavior-based control, which exploits the dynamics of collections of
concurrent, interacting processes coupled to the external world, is
both biologically relevant and effective for problems featuring local
information, uncertainty, and non-stationarity. In this talk we
describe methods we have developed for principled behavior-based
control and learning in two problem domains: multi-robot team
coordination and humanoid imitation.
In the multi-robot domain the key challenges involve reconciling
individual and group-level goals and achieving scalable, on-line
real-time learning. How to do all of this in a distributed
behavior-based way in a timely and consistent fashion? We will
describe our results in making distributed, behavior-based systems
perform in a well-behaved fashion on problems of behavior selection at
the individual and group level, communication for dynamic task
allocation, and on-line model learning. We will describe the use of
Pareto-optimality and satisficing to make behavior selection both
principled and timely, the robust publish/subscribe messaging paradigm
for distributed communication, and augmented Markov models for on-line
real-time model building for adaptation. We will demonstrate the
results of these methods on groups of locally-controlled but globally
efficient cooperative mobile robots performing distributed collection,
multiple-target-tracking and capture, and object manipulation.
In the humanoid control domain the key challenges are the high
dimensionality of the problem and the choice of representation and
modularity that properly integrates the perceptual and motor systems.
We describe an imitation system modeled on psychophysical and
neuroscience evidence for selective movement attention, mirror
neurons, and motor primitives. The system employs direct
sensory-motor mappings within the behavior-based framework to address
how to understand, segment, and map the observed movement onto the
existing motor system. The same biologically motivated structure
serves for recognition, classification, prediction, and learning. We
will demonstrate the results of this model on a 20 DOF dynamic
humanoid imitating human dance and sports movements from visual data.
Maja Mataric
is an assistant professor in the Computer Science Department and the Neuroscience Program at the University of Southern California, the Director of the USC Robotics Research Labs and an Associate Director of IRIS (Institute for Robotics and Intelligent Systems). She joined USC in September 1997, after two and a half years as an assistant professor in the Computer Science Department and the Volen Center for Complex Systems at Brandeis University. She founded her Interaction Lab in 1995.
Maja Mataric received her PhD in Computer Science and Artificial Intelligence from MIT in 1994, her MS in Computer Science from MIT in 1990, and her BS in Computer Science from the University of Kansas in 1987. She is a recipient of the NSF Career Award and the MIT TR100 Innovation Award, and is featured in the upcoming movie about scientists, "Me & Isaac Newton". She has worked at NASA's Jet Propulsion Lab, the Free University of Brussels AI Lab, LEGO Cambridge Research Labs, GTE Research Labs, the Swedish Institute of Computer Science, and ATR Human Information Processing Labs, interacting with a variety of human and robotic colleagues (the latter ranging from LEGO robots to full-body humanoids). Her Interaction Lab at USC performs research in the areas of control and learning in behavior-based multi-robot systems, and skill learning by imitation based on sensory-motor primitives.
05/02/00 - Song-chun Zhu
Department of Computer and Information Sciences, Ohio State University.
Mathematical Modeling of Image Ensembles: Descriptive vs. Generative Models
In this talk, I shall discuss a few fundamental questions in computer
vision and pattern recognition: what are mathematical definitions
for stochastic visual patterns, such as a texture or a shape? why do we
have to deal with probabilistic models in CVPR? What are the
first principles for learning the "proper" models in a given application?
To answer these questions, I will discuss the concept of image ensemble,
and then compare two existing mathematical paradigms for modeling image
ensembles. The first paradigm studies descriptive models: such as
Markov random fields and the FRAME model learned under the principle of
minimax entropy. The second paradigm studies generative models, such
as PCA/ICA/TCA and other graphic models. Then I will discuss some
fundamental issues for both families of models and the intrinsic
links between the two paradigms. I will use texture and texton modeling
as examples to illustrate the concepts.
This is joint work with C. Guo, X. Liu at OSU, Y. Wu at UCLA and
D. Mumford at Brown.
Song Chun Zhu
received his BS degree from USTC in 1991. He received his
MS and PhD degrees in computer science from Harvard University in 1994
and 1996 respectively. In 1996-97, he was a research associate
in the Division of Applied Math at Brown University. In 1997-98, he was a
lecturer in the Computer Science Department in Stanford University. He
joined the faculty of the Department of Computer and Information Sciences
in The Ohio State University in 1998, where he leads the OSU Vision And
Learning (OVAL) group. His research is concentrated on the areas of
computer vision, statistical modeling, and stochastic computing.
He has published over 40 articles on object recognition, image
segmentation, texture modeling, visual learning, perceptual organization
and performance analysis. He received numerous research awards.
05/09/00 - Yoav Freund AT&T Research Labs
Decision Trees, Margins and Brownian Motion
Adaboost is a learning algorithm that, in recent years, has been
gaining a name as one of the best off-the-shelf learning algorithms.
In fact, Adaboost is not exactly a learning algorithm, rather, it is
a method for improving or "boosting" the performance of a given
learning algorithm, often called the "base learner". The most popular
base learners so far have been algorithms for learning decision trees,
such as CART and C4.5. In the first part of the talk I will describe a
new learning algorithm which integrates a tree-learning algorithm with
a boosting algorithm. This algorithm generates a new type of
classifier, which we call "alternating trees" and is a significantly
more powerful representation than decision trees or boosted decision
trees.
One of the surprising phenomena associated with boosting is that it is
much more resistant to overfitting than expected. In recent work we
have demonstrated that this behavior can be explained using the
notion of "margins". In fact, looking at adaboost from this
perspective it becomes apparent that adaboost is performing a type of
gradient descent with respect to a potential function that is
exponential in the margin. It also highlights the problem that the
exponential function is a poor approximation to the actual target
function, which is the classification error step function. In the
second part of the talk I will describe a new boosting algorithm which
is based on the equation for the time evolution of Brownian
motion. This algorithm approximates the classification step function
with an error function (erf) rather than an exponential.
Yoav Freund
Main area of research: computational learning theory and the related areas in probability theory, information theory, statistics and pattern recognition.
Webpage maintained by Stella
X. Yu
