Forum for Artificial Intelligence

About FAI

The Forum for Artificial Intelligence meets every other week (or so) to discuss scientific, philosophical, and cultural issues in artificial intelligence. Both technical research topics and broader inter-disciplinary aspects of AI are covered, and all are welcome to attend!

If you would like to be added to the FAI mailing list, or have any questions or comments, please send email to Prem Melville, Misha Bilenko, or Nick Jong.

Past talks

11/29 • Michael A. Arbib
11/19 • Thomas Dietterich
11/18 • Tina Eliassi-Rad
11/12 • Michael Gelfond
11/11 • Foster Provost
10/28 • Dieter Fox
10/22 • Michael Witbrock
10/15 • Martha Palmer
10/1 • Astro Teller
9/30 • Michael Pazzani
9/3 • Jeffrey Mark Siskind

Monday, November 29th, 12:00pm

Coffee at 11:45am

ACES 2.302 Avaya Auditorium

Modeling the Mirror System: From Hand Movements to Language

Prof. Michael A. Arbib   [homepage]
Departments of Computer Science and Neuroscience
University of Southern California

The mirror system in the macaque monkey is a set of neurons each of which is active both when the monkey makes certain actions and observes others (human or monkey) make similar actions. Brain imaging suggests that humans have such a mirror system as well, with a key part in or near Broca's area, a key player in the brain's mechanisms for language. This grounds the mirror system hypothesis tracing an evolutionary path from a mirror system for grasping via imitation, pantomime, protosign and protospeech to language. The talk will provide an exposition and critique of these ideas which will be informed by analysis of computational models which probe the development and function of the mirror system in the macaque. The debate between Arbib and Davis & MacNeilage on whether protospeech needed the scaffolding of protosign will be reviewed, briefly.

About the speaker:

Michael A. Arbib is the Fletcher Jones Professor of Computer Science, as well as a Professor of Biological Sciences, Biomedical Engineering, Electrical Engineering, Neuroscience and Psychology at the University of Southern California (USC). He has also been named as one of a small group of "University Professors" at USC in recognition of his contributions across many disciplines. He received his Ph.D. in Mathematics from MIT in 1963. He is the author or editor of more than 30 books, including "Brains, Machines and Mathematics" (McGraw-Hill, 1964), "Neural Organization: Structure, Function, and Dynamics" (with Peter Erdi and John Szentagothai, MIT Press, 1998), and the edited volume "The Handbook of Brain Theory and Neural Networks" (MIT Press, Second Edition, 2003). Jointly sponsored by the Departments of Computer Sciences and Communication Sciences and Disorders.

Friday, November 19th, 3:00pm

Coffee at 2:45pm

ACES 2.302 Avaya Auditorium

Three Challenges for Machine Learning Research

Prof. Thomas G. Dietterich   [homepage]
School of Electrical Engineering and Computer Science
Oregon State University

[Sign-up schedule for individual meetings]

Over the past 25 years, machine learning research has made huge progress on the problem of supervised learning. This talk will argue that now is the time to consider three new directions.

The first direction, which is already being pursued by many groups, is Structural Supervised Learning in which the input instances are no longer independent but instead are related by some kind of sequential, spatial, or graphical structure. A variety of methods are being developed, including hidden Markov support vector machines, conditional random fields, and sliding window techniques.

The second new direction is Transfer Learning in which something is learned on one task that can help with a second, separate task. This includes transfer of learned facts, learned features, and learned ontologies.

The third new direction is Deployable Learning Systems. Today's learning systems are primarily operated offline by machine learning experts. They provide an excellent way of constructing certain kinds of AI systems (e.g., speech recognizers, handwriting recognizers, data mining systems, etc.). But it is rare to see learning systems that can be deployed in real applications in which learning takes place on-line and without expert intervention. Deployed learning systems must deal with such problems as changes in the number, quality, and semantics of input features, changes in the output classes, and changes in the underlying probability distribution of instances. There are also difficult software engineering issues that must be addressed in order to make learning systems maintainable after they are deployed.

About the speaker:

Dr. Dietterich (AB Oberlin College 1977; MS University of Illinois 1979; PhD Stanford University 1984) joined the Oregon State University faculty in January 1985. In 1987, he was named a Presidential Young Investigator for the NSF. In 1990, he published, with Dr. Jude Shavlik, the book entitled Readings in Machine Learning, and he also served as the Technical Program Co-Chair of the National Conference on Artificial Intelligence (AAAI-90). From 1992-1998 he held the position of Executive Editor of the journal Machine Learning. The American Association for Artificial Intelligence named him a Fellow in 1994, and the Association for Computing Machinery did the same in 2003. In 2000, he co-founded a new, free electronic journal: The Journal of Machine Learning Research. He served as Technical Program Chair of the Neural Information Processing Systems (NIPS) conference in 2000 and General Chair in 2001. He currently President of the International Machine Learning Society, a member of the DARPA Information Science and Technology Study Group, and he also serves on the Board of Trustees of the NIPS Foundation.

Thursday, November 18th, 3:30pm

Coffee at 3:15pm

Taylor 3.128

A Probabilistic Approach to Accelerating Path-Finding in Large Semantic Graphs

Dr. Tina Eliassi-Rad   [homepage]
Center for Applied Scientific Computing
Lawrence Livermore National Laboratory

[Sign-up schedule for individual meetings]

The majority of real-world graphs contain semantics. That is, they encode meaningful entities and relationships in their vertices and edges, respectively. Moreover, such graphs have semantic types associated with their vertices and edges. These types provide an ontology (or a schema) graph (i.e., they encode the types of the vertices that may be connected via a given edge type). In this paper, we use ontological information, probability theory, and artificial intelligence (AI) search techniques to reduce and prioritize the search space between a source vertex and a destination vertex for path-finding tasks in large semantic graphs. Specifically, we introduce two probabilistic heuristics that utilize a semantic graph's ontological information. We embed our heuristics into A* and compare their performances to breadth-first search and the simple non-probabilistic A* search. We test our heuristics on large synthetic and real ontologies and semantic graphs with real-world properties (such as graphs with "scale-free" or "small-world" topologies). Our experimental results illustrate the merits of our approach.

About the speaker:

Tina Eliassi-Rad joined the Center for Applied Scientific Computing at Lawrence Livermore National Laboratory as a computer scientist in September 2001. Her research interests include machine learning, knowledge discovery and data mining, artificial intelligence, text and web mining, information retrieval and extraction, intelligent software agents, bioinformatics, intrusion detection, and E-commerce.

She earned a Ph.D. in Computer Sciences (with a minor in Mathematical Statistics) at the University of Wisconsin-Madison in 2001. She completed her M.S. in Computer Science at the University of Illinois at Urbana-Champaign in 1995, and her B.S. in Computer Sciences at the University of Wisconsin-Madison in 1993.

Friday, November 12th, 3:00pm

Coffee at 2:45pm

ACES 2.402

Answer Set Programming and Design of Deliberative Agents

Prof. Michael Gelfond   [homepage]
Department of Computer Science
Texas Tech University

[Sign-up schedule for individual meetings]

Answer set programming (ASP) is a new declarative programming paradigm suitable for solving a large range of problems related to knowledge representation and search. ASP begins by encoding relevant domain knowledge as a logic program, P, whose connectives are understood in accordance with the answer set (stable model) semantics of logic programming. In the second stage of the ASP programming process, a programming task is reduced to finding the answer sets of a logic program P + R where R is normally a simple program corresponding to this task. The answer sets are found with the help of answer set solvers - programming systems implementing various answer set finding algorithms. During the last few years the answer set programming paradigm seems to have crossed the boundaries of AI and has started to attract people in various areas of computer science. In this talk I will discuss the use of ASP for the design and implementation of software components of deliberative agents capable of reasoning, planning and acting in a changing environment. The basic idea will be illustrated by discussing the use of ASP for the development of a decision support system for the Space Shuttle.

About the speaker:

Michael Gelfond received his PhD from the Institute of Mathematics of the Academy of Sciences, St. Petersburg, in 1974. He is currently a professor in the Computer Science Department at Texas Tech University. Michael Gelfond is a fellow of AAAI, Area Editor for the International Journal of Logic Programming, and the Executive Editor of the Journal of Logic and Computation.

Thursday, November 11th, 4:00pm

Coffee at 3:45pm

CBA 3.202

Classification and Learning with Networked Data: some observations and results

Prof. Foster Provost   [homepage]
Dept. of Information, Operations and Management Sciences
Leonard N. Stern School of Business
New York University

[Signup schedule for individual meetings]

As information systems record and provide access to increasing amounts of data, connections between entities become available for analysis. Customer accounts are linked by communications and other transactions. Organizatons are linked by joint activities. Text documents are hyperlinked. Such networked data create opportunities for improving classification. For example, for detecting fraud a common and successful strategy is to use transactions to link a questionable account to previous fraudulent activity. Document classification can be improved by considering hyperlink structure. Marketing can change dramatically when customer communication is taken into account. In this talk I will focus on two unique characteristics of classification with networked data. (1) Knowing the classifications of some entities in the network can improve the classification of others. (2) Very-high-cardinality categorical attributes (e.g., identifiers) can be used effectively in learned models. I will discuss methods for taking advantage of these characteristics, and will demonstrate them on various real and synthetic data sets.
(Joint work with Claudia Perlich and Sofus Macskassy)

About the speaker:

Foster Provost is Associate Professor of Information Systems and NEC Faculty Fellow at New York University's Stern School of Business. He is Editor-in-Chief of the journal Machine Learning, and a founding board member of the International Machine Learning Society. Professor Provost's recent research focuses on mining networked data, economic machine learning, and applications of machine learning and data mining. Previously, at NYNEX/Bell Atlantic Science and Technology, he studied a variety of applications of machine learning to telecommunications problems including fraud detection, network diagnosis and monitoring, and customer contact management.

Thursday, October 28th, 11:00am

Coffee at 10:45am

ACES 2.402

Location Estimation for Activity Recognition

Prof. Dieter Fox  [homepage]
Department of Computer Science and Engineering
University of Washington

Knowledge of a person's location provides important context information for many pervasive computing applications. Beyond this, location information is extremely helpful for estimating a person's high-level activities. In this talk we show how Bayesian filtering can be applied to estimate the location of a person using sensors such as GPS, infrared, or WiFi. The techniques track a person on graph structures that represent a street map or a skeleton of the free space in a building. In the context of GPS, we show how such a graph representation can be embedded into a hierarchical activity model that learns and infers a user's daily movements through the community. The model uses multiple levels of abstraction in order to bridge the gap between raw GPS measurements and high level information such as a user's mode of transportation or her goal.

About the speaker:

Dieter Fox is an Assistant Professor of Computer Science & Engineering at the University of Washington, Seattle. He obtained his Ph.D. from the University of Bonn, Germany. Before joining UW, he spent two years as a postdoctoral researcher at the CMU Robot Learning Lab. His research focuses on probabilistic state estimation in robotics and activity recognition. He received various awards, including an NSF CAREER award and best paper awards at major robotics and artificial intelligence conferences.

Friday, October 22nd, 3:00pm

Coffee at 2:45pm

ACES 2.402

Learning From Knowledge - Getting Cyc to Build Itself

A recording of this talk is available, please contact one of the organizers if you would like to borrow it.

Dr. Michael Witbrock  [homepage]
Cycorp, Inc.

For the past twenty years, human beings have been painstakingly adding formally represented knowledge to Cyc. While this knowledge base has been usefully applied to several real-world and research problems, it is insufficient to approach the eventual goal of a fully functioning Artificial Intelligence. One of the original premises of the Cyc project was that one could only acquire knowledge from a base of knowledge; you can't learn anything unless you know something. We're now in a position to put that premise to the test. In this talk, I'll describe the Cyc system, and how we are applying its current knowledge base, NL capability, and inference power to the problem of automated knowledge acquisition.

About the speaker:

Dr. Michael Witbrock (Cycorp), has a PhD in Computer Science from Carnegie Mellon University, and currently is Vice President for Research at Cycorp. Before joining Cycorp, in 2001, to direct its knowledge formation and dialogue processing efforts he had been Principal Scientist at Terra Lycos, working on integrating statistical and knowledge based approaches to understanding web user behavior, a research scientist at Just Systems Pittsburgh Research Center, working on statistical summarization, and a systems scientist at Carnegie Mellon on the Informedia spoken document information retrieval project. He also performed dissertation work in the area of speaker modeling. He is author of numerous publications in areas ranging across neural networks, parallel computer architecture, multimedia information retrieval, web browser design, genetic design, computational linguistics and speech recognition.

Friday, October 15th, 3:00pm

Coffee at 2:45pm

ACES 2.302 Avaya Auditorium

Putting Meaning into Your Trees

Prof. Martha Palmer  [homepage]
Computer and Information Sciences Department
University of Pennsylvania

The current success of applications of machine learning techniques to tasks such as part-of-speech tagging and parsing has kindled the hope that these same techniques might have equal or greater success in other areas such as lexical semantics. Advances in automated and semi-automated methods of acquiring lexical semantics would release the field from its dependence on well-defined sub-domains and enable broad-coverage natural language processing. However, supervised machine learning requires large amounts of publicly available training data, and a prerequisite for this training data is general agreement on which elements should be tagged and with what tags. With respect to lexical semantics, this type of general agreement has been strikingly elusive. A recent consensus on a task-oriented level of semantic representation to be layered on top of the existing Penn Treebank syntactic structures has been achieved. This level, know as the Proposition Bank, or PropBank, consists of argument labels for the semantic roles of individual verbs and similar predicating expressions such as participial modifiers and nominalizations. This talk will describe the PropBank verb semantic role annotation being done at Penn for both English and Chinese. The annotation process will be discussed as well as the use of existing lexical resources such as WordNet, Levin classes and VerbNet. Similar projects include the FrameNet Project at Berkeley and the Prague Tectogrammatics project. PropBank annotation is shallower than the Prague Tectogrammatics project and more broad coverage than FrameNet, in that every verb instance in the corpus has to be annotated. The talk will also briefly describe progress in developing automatic semantic role labelers based on this training data and investigations into the role of sense distinctions in improving performance.

About the speaker:

Martha Palmer is an Associate Professor in the Computer and Information Sciences Department of the University of Pennsylvania. She has been a member of the Advisory Committee for the DARPA TIDES program, the Chair of SIGLEX, the Chair of SIGHAN, and is now Vice-President of the Association for Computational Linguistics. Her early work on lexically based semantic interpretation formed the basis of the successful DARPA-funded message processing system, Pundit, and fostered a continuing interest in Information Extraction (ACE) and Machine Translation (TIDES). Her interest in lexical semantics and verb classes also led to her involvement in SENSEVAL and the development of English VerbNet and the English, Chinese and Korean Proposition Banks.

Friday, October 1st, 11:00am

Coffee at 10:30am

ACES 2.302 Avaya Auditorium

Building a New Kind of Body Monitoring Company around Machine Learning

A recording of this talk is available, please contact one of the organizers if you would like to borrow it.

Dr. Astro Teller  [homepage]
BodyMedia, Inc.

One trillion dollars of US healthcare costs per year are directly attributable to people's lifestyle choices and our country spends less than 5% of that addressing this issue. What if there was an unobtrusive, accurate way to gather the physical and mental states of people in their natural environments, in real time and over long periods of time? If such information could be obtained, we could start to address the fundamental issue in health and wellness: behavior modification. This talk is a tour through five years of challenges and discoveries building a wearable body monitoring business using machine learning techniques. The talk will cover challenges gathering data, building body state models, validating the models with the medical community, and will place AI within the larger context of the company, BodyMedia, and the healthcare, wellness, and fitness industries.

About the speaker:

A respected scientist, seasoned entrepreneur, and award-wining novelist, Dr. Astro Teller's endeavors all grow out of a passion for the transformative nature of intelligent technologies. Dr. Teller is currently the CEO of BodyMedia, Inc, the leading company in unobtrusive wearable body monitoring. Past work has taken him through a previous CEO position, teaching and researching at Stanford University, numerous patents, a Hertz fellowship, a range of technical and non-technical articles and books, and $22M in raised capital. Dr. Teller holds a BS in computer science and an MS in symbolic and heuristic computation, both from Stanford University. Dr. Teller completed his Ph.D. in computer science at Carnegie Mellon University.

Thursday, September 30th, 2:00pm

Coffee at 1:30pm

ACES 2.302 Avaya Auditorium

Machine Learning for Personalized Wireless Portals

Dr. Michael Pazzani  [homepage]
Information and Intelligent Systems Division
National Science Foundation

People have access to vast stores of information on the World Wide Web ranging from online publications to electronic commerce. All this information, however, used to be accessible only while users are tethered to a computer at home or in an office. Wireless data and voice access to this vast store allows unprecedented access to information from any location at any time. The presentation of this information must be tailored to the constraints of mobile devices. Although browsing and searching are the acceptable methods of locating information on the wired web, those operations soon become cumbersome and inefficient in the wireless setting and impossible in voice interfaces. Small screens, slower connections, high latency, limited input capabilities, and the serial nature of voice interfaces present new challenges. This talk focuses on personalization techniques that are essential for the usability of handheld wireless devices.

About the speaker:

Michael J. Pazzani is the Director of the Information and Intelligent Systems Division of the National Science Foundation. He received his Ph.D. in Computer Science from UCLA and is on leave from a full professorship at the University of California, Irvine where he also served as department chair of Information and Computer Science at UCI for five years. Dr. Pazzani serves on the Board of Regents of the National Library of Medicine. He is a fellow of the American Association of Artificial Intelligence and has published numerous papers in machine learning, personalization, information retrieval, and cognitive science.

Friday, September  3rd, 3:00pm

Coffee at 2:30pm

ACES 2.302 Avaya Auditorium

Stochastic Spatio-Temporal Grammars for Images and Video

Prof. Jeffrey Mark Siskind  [homepage]
School of Electrical and Computer Engineering
Purdue University

Probabilistic Context-Free Grammars (PCFGs) induce distributions over strings. Strings can be viewed as observations that are maps from indices to terminals. The domains of such maps are totally ordered and the terminals are discrete. We extend PCFGs to induce densities over observations with unordered domains and continuous-valued terminals. We call our extension Spatial Random Tree Grammars (SRTGs). While SRTGs are context sensitive, the inside-outside algorithm can be extended to support exact likelihood calculation, MAP estimates, and ML estimation updates in polynomial time on SRTGs. We call this extension the center-surround algorithm. SRTGs extend mixture models by adding hierarchal structure that can vary across observations. The center-surround algorithm can recover the structure of observations, learn structure from observations, and classify observations based on their structure. We have used SRTGs and the center-surround algorithm to process both static images and dynamic video. In static images, SRTGs have been trained to distinguish houses from cars. In dynamic video, SRTGs have been trained to distinguish entering from exiting. We demonstrate how the structural priors provided by SRTGs support these tasks.

Joint work with Charles Bouman, Shawn Brownfield, Bingrui Foo, Mary Harper, Ilya Pollak, and James Sherman.

About the speaker:

Jeffrey Mark Siskind received the B.A. degree in computer science from the Technion, Israel Institute of Technology in 1979, the S.M. degree in computer science from MIT in 1989, and the Ph.D. degree in computer science from MIT in 1992. He did a postdoctoral fellowship at the University of Pennsylvania Institute for Research in Cognitive Science from 1992 to 1993. He was an assistant professor at the University of Toronto Department of Computer Science from 1993 to 1995, a senior lecturer at the Technion Department of Electrical Engineering in 1996, a visiting assistant professor at the University of Vermont Department of Computer Science and Electrical Engineering from 1996 to 1997, and a research scientist at NEC Research Institute, Inc. from 1997 to 2001. He joined the Purdue University School of Electrical and Computer Engineering in 2002 where he is currently an associate professor. His research interests include machine vision, artificial intelligence, cognitive science, computational linguistics, child language acquisition, and programming languages and compilers.

Past Schedules

Spring 2004

Fall 2003

Spring 2003

Fall 2002

Spring 2002

Fall 2001

Spring 2001

Fall 2000

Spring 2000