Time: Tuesday 3:30-4:30pm
Place: Wean Hall 5409
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
09/22/98 - Herb Simon Carnegie Mellon University
Brand Names and Cumulative Science
In AI, we typically (and properly) implement our research with large complex systems that employ a collection of interacting mechanisms to achieve their results. Hence, we often think of advances in AI knowledge in terms of the names of such programs: e.g., LT, GPS, EPAM, PRODIGY, Soar, BACON, Act-R --just to mention a few of local origin. However, it can be argued that the real "action" lies largely in the mechanisms embedded in these programs, and in issues about how such mechanisms can be combined effectively.
The "brand names" tend to make difficult the analysis and comparison of these mechanisms or the exchange of knowledge between research groups. One can argue that it has caused and causes an enormous amount of duplication of effort. Physicists did not divide quantum mechanics into the Heisenberg Brand, the Schrodinger Brand,and the Dirac Brand, but analyzed in detail the relations among these and use one or the other according to their computational power in particular situations. When specific "brand name" choices have arisen (wave v. particle theories of light, Ampere's v. Faraday"s theories of electromagnetism, phlogisten v. oxygen theories of production), they used experimental techniques to analyze both similarities and differences and to sort them out.
How can we develop AI theory that goes beyond Brand Names? What techniques do we have for comparing theories (more than staging horse races between them in particular tasks)? How do we trace particular mechanisms and their development through the theories in which they develop? The aim of the talk will not be to provide answers to all of these questions but to stimulate constructive discussion about new forms of research activity that will address them.
09/29/98 - Sebastian Thrun Carnegie Mellon University
When Robots Meet and Lead People
A few weeks ago, a mobile robot named Minerva was "hired" by the Smithsonian's National Museum of American History, to give tours to visitors. During a two-week period, the robot successfully traversed 44km through the "unmodified" museum, interacting with more than 50,000 people along the way.
This talk presents some of the underlying scientific methods for navigation and human-robot interaction. The speaker will argue that Minerva's ability to negotiate crowds of people in natural environments arises from three aspects: an ability to learn, an ability to handle uncertainty in a principled manner, and the use of any-time algorithms for decision making. In addition, special emphasis was placed on a "lifelike" appearance when designing the robot, using speech, a moving head, and a motorized face to communicate simple "emotions" to museum visitors. Videos will be shown!
The work was jointly carried out with Dieter Fox, Nicholas Roy, Greg Armstrong, Chuck Rosenberg, Jamie Schulte, and Anne Watzman from CMU, and Wolfram Burgard, Dirk Haehnel, Dirk Schulz, and Maren Bennewitz from the University of Bonn (Germany).
Sebastian Thrun is affiliated with Carnegie Mellon University's
Computer Science Department, pursuing research on AI, machine
learning, and robotics.
10/02/98 - Andrew Moore Robotics Institute and School of Computer Science, Carnegie Mellon University
Multi-Value-Functions: Efficient Automatic Action Hierarchies for Multi-Goal Markov Decision and Reinforcement Learning Tasks
When you've planned a strategy to get to some goal in an uncertain, noisy, world, but then someone asks about a different goal, what do you do?
In this talk, we will review some approaches, and then come up with a somewhat surprising result: it's actually not much harder to plan in advance for all possible start/goal combinations than for just one start and/or goal. This statement is subject to a number of important caveats, described in the talk.
An interesting feature of this technique is that it is based on natural and automatic generation of a hierarchy of abstract Markov decision process controllers.
The talk will describe previous approaches to the "all-goals" problem. It will then describe the new hierarchy, how to construct it, how to prove its performance guarantees and some animated examples of its empirical performance. We describe its role in motion planning and inventory management. We believe the greatest use will come in hierarchical control systems in which a higher level controller wishes to not merely call one of a set of atomic lower level controllers, but instead one of a set of lower level controllers that can be parameterized with a subgoal. At the end of the talk we will discuss how this is used for a task with hunters and prey which would not have been computationally feasible with previous methods.
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.
10/13/98 - R. Ravi Operation Research and Department of Computer Science, Carnegie Mellon University
Computational Challenges in Molecular Biology
The past few decades have witnessed an explosion of knowledge in the area of genomic sequences - their content, organization and function. The assembly and analysis of this sequence information involves solving several challenging computational problems en route.
After briefly describing some basic biological concepts and protocols, I will outline the steps of a typical gene-hunting project to highlight and describe the computational problems that arise in the course of such a search. These problems include physical mapping, sequence assembly, sequence alignments, phylogeny reconstruction, and gene prediction; A variety of techniques from combinatorial optimization and machine learning have been used to address these problems.
R. Ravi is an assistant professor of Operations Research at GSIA with a
courtesy appointment in the Computer Science Department. He has been at CMU since Fall 95. His primary
research interests
are Combinatorial Optimization, Approximation Algorithms and
Computational Molecular Biology. Before joining Carnegie Mellon,
Ravi was a DIMACS post-doctoral fellow at Princeton University during
94-95 as part of the special year on Mathematical Support for Molecular
Biology. Ravi's training and work in computational biology began during
a year of postdoctoral work at UC Davis in 93-94. Ravi has a Ph. D. in
Computer Science from Brown
University in '93, and a B. Tech. in Computer Science and Engineering
from the Indian Institute of Technology (IIT), Madras, in '89.
10/27/98 - Moises Goldszmidt SRI International
Getting the Best of Both Discretization and Parametric Fitting
Machine learning problems often confront users with many representation choices including the level of granularity for discrete features, and the family of distributions to be used for continuous features. Needless to say that these choices involve tradeoffs that influence the resulting model and its performance. In this talk I will describe an approach, based on Bayesian networks, for simultaneously representing features in different forms. I will illustrate this approach in the context of pattern classification, where continuous features are simultaneously represented in both discrete and (semi)parametric form. This dual representation frees the classifier from committing to one or the other, and enables different features to correlate to either representation in the same model. Our empirical results show that this classifier usually achieves performance that is as good as or better than similar classifiers that commit to a single representation, or that include both representations without modeling their relation. During the talk I will place this method in the general context of inducing Bayesian networks from data and discuss a set of open problems and future work.
Part of this talk includes joint work with Nir Friedman of the Hebrew University and Tom Lee of SRI.
Dr. Moises Goldszmidt is currently a senior computer scientist with SRI
International performing research in the area of learning and adaptive
systems. Previous to that he was a member of the technical staff and
principal
investigator at the Rockwell Science Center in Palo Alto. He holds an
Electrical Engineering degree from Simon Bolivar University (Venezuela), a MSc
in Electrical Engineering from UC Santa Barbara, and a PhD in Computer Science
from UC Los Angeles. He has numerous publications on topics related to
representation and reasoning under uncertainty, decision making, and automatic
induction of probabilistic models. His research interests include: decision
making under uncertainty, automatic induction of probabilistic models, pattern
classification, data mining, planning and control, machine learning, knowledge
representation and AI.
01/26/99 - Kristian Hammond Department of Computer Science, Northwestern University
Anticipating Users' Needs: Redeeming Big Brother in the Information Age
One of the core problems running through work in information systems is "the query". Few people know how to build them and even fewer systems provide aids that help to construct them. What is needed, is a new way of communicating with information systems.
The research at Northwestern's Intelligent Information Laboratory (The InfoLab) is aimed directly at this problem of reshaping the way in which users interact with information to better fit the goals on the human side of the equation. In this talk, I will discuss three of our approaches to interaction with information access. The first---Q&A---takes its lead from work in case-based reasoning and involves restructuring of the communication to reflect user goals. The other two---Watson and the Content and Control Project---take a somewhat more radical approach in which explicit communication with the information systems vanishes, and is replaced by a set of applications that observe users and provide information as it is needed rather than requested. In all three, the ultimate goal is to remove the query from the information access equation altogether.
Kristian Hammond received his Ph.D. in Computer Science from Yale University in May of 1986. From 1986 until September of 1998, he was the Director of the Artificial Intelligence Laboratory at the University of Chicago and guided the development of intelligent agents in domains ranging from radiation treatment therapy to geometry problem solving and real-time game playing. His focus has been on the development of models of cognition based on episodic memory and the view of reasoning as reminding.
Four years ago, Dr. Hammond formed The Chicago Intelligent Information Laboratory (InfoLab). The InfoLab's mission is focused on issues of information access and management that rise out of the high-speed computer connectivity of the modern world. Its charter is to invent and creatively exploit new information technologies in the development of systems that are responsive to and supportive of human goals and their achievement within complex computer environments. The InfoLab's mandate is to seize the opportunities of the information age and develop technologies that forward rather than impede human endeavors.
In the Summer of 1998, Professor Hammond and the InfoLab moved to
Northwestern University's Computer Science Department and the Institute for
Learning Sciences.
02/23/99 - Paul Viola MIT AI Laboratory
A Non-parametric Multi-scale Statistical Model for Texture: from Psychophysics to Graphics
Several theories of texture perception use Gabor-like image representations to model human performance on texture segmentation and recogntion tasks. I will present a new generative statistical model for texture which explictly incorporates what we have learned from these models. The generative process is hierarchical, conditioning fine scale details on coarser scales. The conditional distributions are non-parametric and can be estimated from example textures. Our model is quite general and can be used to model textures which are "noisy" as well as those that contain textons. The model can be used to recognize textures, but interestingly it can also generate novel textures. While these textures are different from the original textures, they are almost indistinguishable from the original. The model is useful for computer vision (recognition), psychophysics and computer graphics (texture mapping).
Paul Viola is an Assistant Professor of Computer Science and Engineering at the Massachusetts Institute of Technology. Professor Viola received his Ph.D. from MIT in 1995. Before returning to MIT as a professor he spent two years as a visiting scientist in the Computational Neurobiology of the Salk Institute in San Deigo. He has a broad background in advanced computational techniques, publishing in the fields of computer vision, neurobiological vision, medical imaging, mobile robotics, machine learning, and automated drug design.
Professor Viola's research focuses on statistical models of images and
the underlying structures that create them. He has also done work on
the grammatical structure of shapes and image database retrieval.
03/16/99 - Pietro Perona California Institute of Technology
Recognition of Object Classes via Photometry and Probabilistic Geometry
Joint work with Mike Burl, Thomas Leung and Markus Weber.
Objects that share a similar visual appearance, e.g. dogs, sneakers, sedan cars, are commonly grouped into classes by human observers. Despite the fact that no two faces/dogs are identical we can easily detect, and recognize as such, a previously unseen face/dog amongst a clutter of other objects and background texture. We develop a model for these `visual' classes of objects. It consists of `features' and `shape'. The features are localized texture patches at different scales of resolution (e.g. eye-corners, eyes, mouth, jawline). The shape is the mutual position of those features. The intrinsic variability of an object class is represented by a probability density which assigns a likelihood to any variation of the appearance of the textured patches and shape. The shape is constructed in such a way to be invariant with respect to translation, rotation and scale - and to affine deformations if needed. Occlusion and deformation are handled in a principled way.
03/30/99 - Bernardo A. Huberman Xerox Palo Alto Research Center
The Laws of the Web
The World Wide Web has become in a short period a standard source of information for a large part of the world's population. Its exponential growth has rapidly transformed it into an ecology of knowledge in which a highly diverse quantity of information is linked in extremely complex and arbitrary fashion.
In spite of the sheer size of the Web and its highly interactive nature, there exist strong statistical regularities that govern the way it grows, and how people use it and interact with each other. This talk will discuss the existence and nature of such laws and their experimental verification.
Dr. Bernardo A. Huberman is a Research Fellow at the Xerox Palo Alto Research Center, processes in social organizations and computational systems. He received his Ph.D. in Physics from the University of Pennsylvania, and is currently a Consulting Professor of Physics at Stanford University. He has worked in condensed matter physics and the theory of critical phenomena, and is one of the discoverers of chaos in a number of physical systems. He established a number of universal properties in nonlinear dynamics, and his research into the dynamics of complex systems led to the discovery of ultradiffusion in hierarchical structures.
In the field of computation, he predicted the existence of phase transitions in artificial intelligence and large scale computational systems, which have now been observed in a number of computationally hard problems. Dr. Huberman is also one of the pioneers in the field of ecology of computation, and editor of a book on the subject.
Recently, Dr. Huberman has been studying the power of cooperation in collective problem solving by groups, and the dynamics of the Internet, where he discovered a number of strong regularities, such as the law of surfing and the signature of social dilemmas underlying congestion.
At the practical level his team designed and implemented Spawn, a market system for the allocation of idle resources among machines in computer networks, and a thermal market mechanism for the control of building environments. He also holds a number of patents on self-repairing parallel computers, a parallel motion detector and an electronic auction for distributed document services over the Internet.
Dr. Huberman is a Fellow of the American Physical Society, a former trustee of the Aspen Center for Physics and Fellow of the Japan Society for the Promotion of Science. He is co-winner of the 1990 CECOIA prize in Economics and Artificial Intelligence and he recently received the IBM Prize of the Society for Computational Economics. He is also the Chairman of the Council of Fellows at Xerox Corporation, and a faculty member of the Symbolics Systems Program at Stanford University. He has held visiting professorships at the University of Paris and the University of Copenhagen.
04/06/99 - Yoram Singer AT&T Shannon Laboratory
"80 and tea: ho May I hell Pooh ?": Boosting Algorithms for Text and Speech Categorization
When an AT&T customer dials 00, she gets an operator who asks: ``AT&T: how may I help you?''. A possible answer might be: ``Ah yes, I would like to call 908-508-1464 and charge it to my Visa.'' Being able to automatically determine what the caller actually wants (in this case a `dial-for-me' request and a `credit-card-call' request) might be a difficult task especially since there is more than one action to be taken due to the caller's request.
The talk focuses on algorithms which learn from examples to perform text and speech categorization tasks such as the one above. I will first describe a new family of algorithms for classification problems. These algorithms use a machine-learning technique called boosting, a method based on the idea of combining many simple and only moderately accurate rules into a single highly accurate classification rule. I will give a simple analysis of new boosting algorithms and describe various extensions to multiclass settings. I will conclude the talk with a demonstration of a system for automatic call-type identification from unconstrained spoken/written requests.
This is joint work with Rob Schapire (AT&T Labs).
Yoram Singer is a member of the technical staff at AT&T Shannon Laboratory.
He received his PhD in computer science from the Hebrew University, Jerusalem,
Israel. His research focuses on statistical models and algorithms for written
and spoken language understanding.
04/13/99 - Usama Fayyad Microsoft Research
Data Mining and Databases: Recent Developments
Systems for extracting interesting structure from databases, especially large data stores are becoming a necessity. The existing data access model is clearly hitting its limits. Data Mining methods provide a way to address some of these problems. These methods have their origins in statistics, databases, pattern recognition, learning, visualization, and parallel computing. I'll outline some recent advances towards scaling mining algorithms to large database, and cover the research challenges and opportunities posed by the problem of extracting models from massive data sets. The talk will particularly focus on the decomposition of classification and clustering algorithm so they work effectively with a database system backend.
Usama Fayyad is a Senior Researcher at Microsoft Research. His research interests
include scaling data mining algorithms to large databases,
learning algorithms, and statistical pattern recognition, especially
classification and clustering. After receiving the Ph.D. degree from
The University of Michigan, Ann Arbor in 1991, he joined the Jet
Propulsion Laboratory (JPL), California Institute of Technology,
where (until 1996) he headed the Machine Learning Systems Group and
developed data mining systems for automated science data analysis. He
received the 1994 NASA Exceptional Achievement Medal and the JPL 1993
Lew Allen Award for Excellence in Research for his work on developing
data mining systems to solve challenging science analysis problems in
astronomy and remote sensing. He remains affiliated with JPL as a
Distinguished Visting Scientist. He is a co-editor of Advances
in Knowledge Discovery and Data Mining (AAAI/MIT Press, 1996) and is an
Editor-in-Chief of the journal: Data Mining and Knowledge Discovery. He was
program co-chair of KDD-94 and KDD-95 (the First International Conference on Knowledge Discovery and Data Mining) and a general chair of KDD-96. He
co-chaired the 1997 Workshops on the role of KDD in Visualizations held at
KDD-97 and IEEE Vis-97 conferences.
04/20/99 - Dana Ballard University of Rochester
A Model of Predictive Coding Based on Spike Timing
It is remarkeable that so much progress has been made in understanding cognitive processes without a comprehensive understanding of how neurons work. Several decades of research have made many advances towards the goal of interpreting the neural spike train but a comprehensive understanding remains elusive. A new class of neural models termed predictive models characterize the cortex as a memory whose parameters can be used to predict its input. This allows the input to be economically coded as a residual difference between itself and the prediction. Such models have had considerable success in modeling features of visual cortex. We show that the predictive coding model can be extended to a lower level of detail that includes individual neural spikes as primitives. This is a significant improvement in perspicuity compared to the firing rate variables used by most current models. The specific model we describe exploits the use of coincidence of spike arrival times and the fact that neural representations can be distributed over large numbers of cells.
Dana Ballard Ph.D. (1974) University of California at Irvine. Visiting Consultant, Laboratorio Technol. Biomediche, Rome, Italy (74-75). Assistant Professor of Computer Science and Radiology (75-82), Associate Professor of Computer Science (82-87), Professor (87-present); University of Rochester.
Co-author of Computer Vision and author of Introduction to Natural Computation.
Dana Ballard's main research interest is in computational theories of the brain with emphasis on human vision. In 1985 with Chris Brown, he led a team that designed and built a high speed binocular camera control system that is capable of simulating human eye movements. The system is mounted on a robotic arm that allows it to move at one meter per second in a two meter radius workspace. This system has led to an increased understanding of the role of behavior in vision. The theoretical aspects of that system were summarized in a paper "Animate Vision", which received the Best Paper Award at the 1989 International Joint Conference on Artificial Intelligence.
Dana is also interested in models of the brain that relate to detailed neural models. A position paper on this work appeared in the Behavioral and Brain Sciences and can be accessed.
Webpage maintained by Stella X. Yu