Forum for Artificial Intelligence


Forum for Artificial Intelligence

[ About FAI   |   Upcoming talks   |   Past talks ]



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, subscribe here. If you have any questions or comments, please send email to Karl Pichotta or Craig Corcoran.

The current schedule is also available as a Google Calendar or alternatively in iCal format.



[ Upcoming talks ]



[ Past talks]

Tue, August 4
4:00PM
ACES 2.402
James A. Bednar
University of Edinburgh, Institute for Adaptive and Neural Computation
Building a Comprehensive Model for the Development and Function of the Visual Cortex
Tue, September 1
11:00AM
ACES 2.302
Justinian Rosca
Siemens Corporate Research
Doing More with Less: Mutual Interdependence Analysis
Fri, September 11
11:00AM
ACES 2.302
Peter Stone
University of Texas at Austin
Teammates in Ad Hoc Teams or What I did on my sabbatical
Fri, October 9
11:00AM
TAY 3.128
Ubbo Visser
University of Miami, Department of Computer Sciences
Statistics-based real-time sports simulation
Thu, October 22
3:30PM
ACES 2.302
Manuela Veloso
Carnegie Mellon University, Computer Science Department
A Visitor's Companion Robot: Localization, Navigation, and Symbiotic Human-Robot Interaction
Fri, November 6
11:00AM
TAY 3.128
Cynthia Thompson
PriceWaterhouseCoopers
Comparing versions of electronic or paper-based documents with Magic Lens
Fri, November 20
11:00AM
ACES 2.402
Regina Barzilay
MIT CSAIL
Embracing Language Diversity: Unsupervised Multilingual Learning?
Fri, December 4
11:00AM
ACES 2.302
Robin R. Murphy
Texas A&M
Remote Presence: Autonomy Can Be Shared (or Blamed)
Fri, January 29
11:00AM
TAY 3.128
Esra Erdem
Sabanci University
Querying Biomedical Ontologies in a Controlled Natural Language
Fri, February 5
11:00AM
ACES 2.302
Henry Kautz
University of Rochester
Behavior Recognition and Interaction for Cognitive Assistance
Thu, February 11
11:00AM
ACES 2.402
Tal Tversky
Apple, Inc.
Detecting Online Credit Card Fraud: A Data Driven Approach
Fri, February 19
11:00AM
ACES 2.402
Robert Holte
University of Alberta
Using Machine Learning to Iteratively Improve a Search Heuristic
Thu, February 25
11:00AM
ACES 2.402
Astro Teller
Cerebellum Capital, Inc.
Time Traveling And Other Small Problems of AI-Based Finance
Fri, March 5
11:00AM
ACES 2.302
David Forsyth
UIUC
Words and Pictures
Thu, March 11
2:00PM
ACES 2.402
Burr Settles
CMU
Asking the Right Questions: New Query Types for Active Learning
Fri, March 12
11:00AM
ACES 2.402
Dan Klein
University of California at Berkeley
Phylogenetic Models for Natural Language
Fri, March 26
11:00AM
ACES 2.402
James Lester
North Carolina State University
Narrative-Centered Learning Environments
Fri, April 9
11:00AM
TAY 3.128
Noah Smith
CMU
Text-Driven Forecasting: Meaning as a Real Number
Wed, April 14
11:00AM
ACES 2.302
Stephanie Forrest
University of New Mexico
Automatic Program Repair with Evolutionary Computation
Fri, April 23
11:00AM
TAY 3.128
Martial Hebert
CMU
Representations for object and action recognition
Thu, April 29
11:00AM
ACES 2.402
Cynthia Breazeal
MIT
Robots as Social Learners
Fri, April 30
11:00AM
ACES 2.402
Eric Horvitz
Microsoft Research
Learning, Reasoning, and Action in the Open World
Mon, May 24
11:00AM
ACES 2.402
Louis-Philippe Morency
USC
Computational Study Of Nonverbal Social Communication

Tuesday, August 4, 2009, 4:00PM



ACES 2.402

Building a Comprehensive Model for the Development and Function of the Visual Cortex

James A. Bednar   [homepage]

University of Edinburgh, Institute for Adaptive and Neural Computation

Vision is a standard system for studying cortical sensory processing. Previous computational models of adult primary visual cortex (V1) have been able to account for many of the measured properties of V1 neurons, but not how or why these particular properties arise. Previous developmental models have been able to reproduce the overall organization of specific feature maps in V1, such as orientation maps, but the neurons in the simulated maps behave quite unlike real V1 neurons, and in many cases are too abstract even to be testable on actual visual stimuli. I believe that the complex adult circuitry only makes sense when considering the developmental process that created it, and conversely, that the developmental process only makes sense if leading to a system that can perform behaviorally relevant visual tasks.

Accordingly, in this talk I outline a long-term project to build the first model to explain both the development and the function of V1. To do this, researchers in my group are building the first developmental models with wiring consistent with V1, the first to have realistic behavior with respect to visual contrast, the first to include all of the various visual feature dimensions, and the first to include all of the major sources of connectivity that modulate V1 neuron responses. The goal is to have a comprehensive explanation for why V1 is wired as it is in the adult, and how that circuitry leads to the observed behavior of the neurons during visual tasks. This approach leads to experimentally testable predictions at each stage, and can also be applied to understanding other sensory cortices, such as somatosensory and auditory cortex.

About the speaker:

Jim Bednar leads the Computational Systems Neuroscience research group at the University of Edinburgh, and is the deputy director of the Edinburgh Doctoral Training Centre in Neuroinformatics and Computational Neuroscience. His 2002 Ph.D. in Computer Science is from the University of Texas at Austin, and he also has degrees in Philosophy and Electrical Engineering. His research focuses on computational modeling of the development and function of mammalian visual systems. He is a co-author of the monograph "Computational Maps in the Visual Cortex" (Springer, 2005), and is the lead author of the Topographica cortical modeling software package (see topographica.org). He is also a member of the Board of Directors for the annual international Computational Neuroscience Meeting.

Tuesday, September 1, 2009, 11:00AM



ACES 2.302

Doing More with Less: Mutual Interdependence Analysis

Justinian Rosca   [homepage]

Siemens Corporate Research

The mean of a data set is one trivial representation of data from a class. Recently, mutual interdependence analysis (MIA) has been successfully used to extract more involved representations, or "mutual features", accounting for samples in the class. For example a mutual feature is a speaker signature under varying channel conditions or a face signature under varying illumination conditions. A mutual representation is a linear regression that is equally correlated with all samples of the input class. We present the MIA optimization criterion from the perspectives of regression, canonical correlation analysis and Bayesian estimation. This allows us to state and solve the MIA criterion concisely, to contrast the MIA solution to the sample mean, and to infer other properties of its closed form, unique solution under various statistical assumptions.

This work has been done in collaboration with Heiko Claussen (Siemens Corporate Research, Princeton, NJ) and Robert Damper (Univ. Southampton, UK).

About the speaker:

Justinian Rosca received his Ph.D. in Computer Science from the University of Rochester, NY. He is Program Manager in Audio, Signal Processing and Wireless Communications at Siemens Corporate Research in Princeton, NJ, and also Affiliate Professor, Department of Electrical Engineering of University of Washington, Seattle. He conducts research in signal processing and radio management, with an emphasis on topics involving acquisition, management and processing of data with uncertainties, statistical audio processing, blind signal separation, and probabilistic inference. Dr. Rosca has more that two dozen US and international patents awarded. He co-authored two books in mathematics and signal processing, most recently he co-edited the Proceedings of the 6th International Conference on Independent Component Analysis and Blind Signal Separation. He is presently on the editorial board of the Journal of Signal Processing Systems from Springer. Within Siemens, he has lead the Audio Signal Processing program in the development of state-of-the-art algorithms and implementations in microphone array signal processing, blind signal separation, sound analysis and identification with applications in hearing aids.

Friday, September 11, 2009, 11:00AM



ACES 2.302

Teammates in Ad Hoc Teams or What I did on my sabbatical

Peter Stone   [homepage]

University of Texas at Austin

Teams of agents, defined as agents operating in the same environment with identical utility functions, are typically developed in a planned, coordinated fashion. However, such coordinated development is not always possible. Rather, as deployed agents become more common in robotics, e-commerce, and other settings, there are increasing opportunities for previously unacquainted agents to cooperate in ad hoc team settings. In such scenarios, it is useful for individual agents to be able to collaborate with a wide variety of possible teammates under the philosophy that not all agents are fully rational. This talk considers an agent that is to interact repeatedly with a teammate that will adapt to this interaction in a particular suboptimal, but natural way.

We formalize this "ad hoc team" framework in two ways.

First, in a fully cooperative normal form game-theoretic setting, we provide and analyze a fully-implemented algorithm for finding optimal action sequences, prove some theoretical results pertaining to the lengths of these action sequences, and provide empirical results pertaining to the prevalence of our problem of interest in random interaction settings.

Second, we consider a cooperative k-armed bandit in which cooperating agents have access to different actions (arms). In this setting we prove some theoretical results pertaining to which actions are potentially optimal, provide a fully-implemented algorithm for finding such optimal actions, and provide empirical results.

About the speaker:

Dr. Peter Stone is an Alfred P. Sloan Research Fellow, Guggenheim Fellow, Fulbright Scholar, and Associate Professor in the Department of Computer Sciences at the University of Texas at Austin. He received his Ph.D in Computer Science in 1998 from Carnegie Mellon University. From 1999 to 2002 he was a Senior Technical Staff Member in the Artificial Intelligence Principles Research Department at AT&T Labs - Research. Peter's research interests include machine learning, multiagent systems, robotics, and e-commerce. In 2003, he won a CAREER award from the National Science Foundation for his research on learning agents in dynamic, collaborative, and adversarial multiagent environments. In 2004, he was named an ONR Young Investigator for his research on machine learning on physical robots. In 2007, he was awarded the prestigious IJCAI 2007 Computers and Thought award, given once every two years to the top AI researcher under the age of 35.

Friday, October 9, 2009, 11:00AM



TAY 3.128

Statistics-based real-time sports simulation

Ubbo Visser   [homepage]

University of Miami, Department of Computer Sciences

Autonomous agents in real-time and dynamic adversarial environments offer numerous research challenges. Perception, localization, decision-making, communication, and locomotion are good examples. The novel modern sports simulator we will discuss integrates results from ten years of research in the area of autonomous soccer playing robots (both softbots and physical robots) with RoboCup as a testbed.

We will explore the problem of enabling autonomous agents in finding the right passing point or in making a complex decision within a soccer game while dealing with time constrains, hostile opponents, and dynamic environments. We propose a framework for spatio-temporal real-time analysis of dynamic scenes. We introduce a knowledge processing pipeline ranging from relevance-driven compilation of a qualitative scene description to a knowledge-based detection of complex event and action sequences, conceived as a spatio-temporal pattern matching problem. We present experimental results from the RoboCup 3D soccer simulation that substantiate the online applicability of our approach under tournament conditions.

We then take a closer look at new generation sports simulation online games where the results of this research have been integrated. 175,000 users currently operate the 'Official Bundesliga Manager' created by the University of Bremen's spin-off company aitainment GmbH and which was adopted by the German Bundesliga (one of the most prestigious soccer leagues in the world). The 'Official Bundesliga Manager' is a complex real-time soccer simulator that is based on user-models from actual data (e.g. passing performance, scoring accuracy) of real soccer players from the German Bundesliga. The underlying hierarchical three-tier multiagent system consists of autonomous BDI agents that allows dynamic group structures (e.g. an emergent situation for a wing attack). The online game runs seamlessly in a web browser with a new and state-of-the-art 3D visualization engine. Latest developments include research results from a motion capturing lab and face generators to enhance the believability of the players and the users' visualization experience.

About the speaker:

Dr. Ubbo Visser is currently a Visiting Associate Professor in the Department of Computer Science at the University of Miami. He received his Habilitation in Computer Science (qualification for Full Professor) from the University of Bremen in 2003, his PhD in Geoinformatics from University of Muenster in 1995, and his MSc in Landscape-ecology from University of Muenster in 1988. His research specialization is in artificial intelligence, more specifically on knowledge representation and reasoning. He is interested in the combination of symbolic and sub-symbolic technologies in the domain areas of "Semantic Web" and "Multiagent Systems (RoboCup, Games)". His focus in the Semantic Web area lies in the development of methods that combine terminological logics and spatio-temporal representation and reasoning techniques. The focus in the Multiagent Systems area lies in the development of techniques for agents that act in highly dynamic and real-time environments. He won several awards for research and development of innovative AI software (e.g. Best AI Award from the Society for Informatics (GI) in Germany) and is a co-founder of innovative software companies both in Europe and in the United States.

Thursday, October 22, 2009, 3:30PM



ACES 2.302

A Visitor's Companion Robot: Localization, Navigation, and Symbiotic Human-Robot Interaction

Manuela Veloso   [homepage]

Carnegie Mellon University, Computer Science Department

We have been developing an indoor robot, CoBot, as a visitor's companion robot. CoBot moves in buildings for which it knows the map, but that offer dynamic challenges, such as people moving and obstacles without fixed placements. I will present the wifi-based localization and navigation algorithms, illustrating them with examples of multiple hours-long autonomous runs of the robot. I will conclude with our symbiotic interaction algorithm in which the robot and the human have complementary limitations and expertise.

This work is joint with my students Joydeep Biswas, Stephanie Rosenthal, and Nick Armstrong-Crews. The robot was designed and constructed by Michael Licitra, as an omnidirectional four-wheeled robot, inspired by his own previous platform of our small-size soccer robots.

About the speaker:

Manuela M. Veloso is the Herbert A. Simon Professor of Computer Science at Carnegie Mellon University. She researches in artificial intelligence and robotics, in the areas of planning and learning for single and multirobot teams in uncertain, dynamic, and adversarial environments. Veloso is a Fellow of the American Association of Artificial Intelligence, and the President of the RoboCup Federation. She was recently awarded the 2009 ACM/SIGART Autonomous Agents Research Award. Veloso is the author of one book on "Planning by Analogical Reasoning" and editor of several other books. She is also an author in over 200 journal articles and conference papers.

Friday, November 6, 2009, 11:00AM



TAY 3.128

Comparing versions of electronic or paper-based documents with Magic Lens

Cynthia Thompson   [homepage]

PriceWaterhouseCoopers

PricewaterhouseCoopers (PwC) professionals review, compare, and opine on thousands of documents which originate in multiple file formats, with many available only on paper. Magic Lens is a technology enabled document review tool that automatically highlights differences in text when comparing similar documents, reducing document review time by up to 80%. This unique application compares more than just Word documents; it also handles Adobe PDF, scanned paper documents, and even compares documents of different formats. PwC professionals use Magic Lens to compare and review revenue contracts, 10-Ks, invoices, financial agreements and other critical documents with increased accuracy and efficiency.

I will discuss the novel and efficient algorithm for text comparison that we created for Magic Lens, which is capable of finding matching text even when it moves to a different location in a document. We also designed and implemented a novel document comparison user interface. Magic Lens has been used by over 5000 PwC partners and staff.

I will also provide a brief overview of other current & past CAR projects.

The PricewaterhouseCoopers Center for Advanced Research (CAR) conducts PwC-sponsored research and development on business problems that have no known solution in the marketplace. Since innovation often results from combining widely-varied insights, an important part of CAR's strategy is to hire researchers with backgrounds in a variety of fields, including computer science, mathematics, economics, statistics, mechanical engineering, and software development, and further cross pollinate their knowledge with professionals and subject matter specialists from various PwC lines of service. The diversity in this collaboration results in fresh perspectives on real business problems. Located in the heart of Silicon Valley in PwC's San Jose office, CAR began operation in 2003 under the leadership of PricewaterhouseCoopers partner and technology entrepreneur, Sheldon Laube.

We are now accepting applications for spring and summer internships, and Dr. Thompson will be available to meet with interested students.

About the speaker:

Dr. Thompson (Cindi) is a Senior Research Manager at PwC. She led the Magic Lens project for two and a half years and is now leading a new project investigating the impact of interruptions on productivity and quality. Prior to leading Magic Lens, she contributed to the Connection Machine project, an adaptive expertise locator system that, and which was also deployed to the US Firm. Prior to joining PwC, Dr. Thompson was an Assistant Professor in the School of Computing at the University of Utah. Her research focused on the development and application of techniques from machine learning to natural language understanding, scientific time series, and recommendation systems. Cindi's teaching included courses on Machine Learning, Artificial Intelligence, and Discrete Mathematics. Her publications include book chapters in Learning Language in Logic and articles for the Journal of Artificial Intelligence Research, and many conference paper publications.

Cindi also served as Consulting Researcher at Stanford University's Institute for the Study of Learning and Expertise, and during the 2002-03 academic year served as Visiting Assistant Professor in the Computer Science Department at Stanford. She was a Postdoctoral Research Fellow at Stanford University's Center for the Study of Language and Information. Her postdoctoral research subject was spoken dialogue systems that change their interaction behavior based on past interactions with users.

Cindi received her Ph.D. and M.A. in Computer Science from the University of Texas at Austin under Professor Ray Mooney, and her B.S. in Computer Science from North Carolina State University.

Friday, November 20, 2009, 11:00AM



ACES 2.402

Embracing Language Diversity: Unsupervised Multilingual Learning?

Regina Barzilay   [homepage]

MIT CSAIL

For centuries, the deep connection between human languages has fascinated scholars, and driven many important discoveries in linguistics and anthropology. In this talk, I will show that this connection can empower unsupervised methods for language analysis. The key insight is that joint learning from several languages reduces uncertainty about the linguistic structure of each individual language.

I will present multilingual generative unsupervised models for morphological segmentation, part-of-speech tagging, and parsing. In all of these instances we model the multilingual data as arising through a combination of language-independent and language-specific probabilistic processes. This feature allows the model to identify and learn from recurring cross-lingual patterns to improve prediction accuracy in each language. I will also discuss ongoing work on unsupervised decoding of ancient Ugaritic tablets using data from related Semitic languages.

This is joint work with Benjamin Snyder, Tahira Naseem and Jacob Eisenstein.

Friday, December 4, 2009, 11:00AM



ACES 2.302

Remote Presence: Autonomy Can Be Shared (or Blamed)

Robin R. Murphy   [homepage]

Texas A&M

This talk describes a spectrum of three models of teleoperation for remote presence applications, such as emergency response, law enforcement, and military operations in urban terrain, where humans use a robot to obtain real-time perception at a distance. These enterprises are treated as joint cognitive systems and examined in terms of roles, information flow, and team processes. The state of the practice, where the robot has no autonomy, is captured by the Remote Tool Model. The Taskable Agent Model, where the robot has full autonomy and human involvement is negligible represents the other extreme of the spectrum but is not a desirable goal for remote presence applications. A third novel model occupying the space between the two extremes is posited, the Shared Roles Model, which incorporates semi-autonomy and increased communications connectivity. Shared roles provide a naturalistic, explicit representation of the requisite responsibilities and whether the division of functions between the robot and human conserves those responsibilities. The talk discusses whether advances in technology will obviate the Shared Roles Model, what the model implies about the human-robot ratio and whether the ratio can be reduced by merging roles, and identifies open research issues in team processes.

About the speaker:

Robin Roberson Murphy is the Raytheon Professor of Computer Science and Engineering at Texas A&M. She received a B.M.E. in mechanical engineering, a M.S. and Ph.D in computer science in 1980, 1989, and 1992, respectively, from Georgia Tech, where she was a Rockwell International Doctoral Fellow. Her research interests are artificial intelligence, human-robot interaction, and heterogeneous teams of robots. In 2008, she was awarded the Al Aube Outstanding Contributor award by the AUVSI Foundation for her insertion of ground, air, and sea robots for urban search and rescue (US&R) at the 9/11 World Trade Center disaster, Hurricanes Katrina and Charley, and the Crandall Canyon Utah mine collapse. She is a Distinguished Speaker for the IEEE Robotics and Automation Society, and has served on numerous boards, including the Defense Science Board, USAF SAB, NSF CISE Advisory Council, and DARPA ISAT.

Friday, January 29, 2010, 11:00AM



TAY 3.128

Querying Biomedical Ontologies in a Controlled Natural Language

Esra Erdem   [homepage]

Sabanci University

Recent advances in health and life sciences have led to generation of a large amount of data. To facilitate access to its desired parts, such a big mass of data has been represented in structured forms, like biomedical ontologies. On the other hand, representing biomedical ontologies in a formal language, constructing them independently from each other and storing them at different locations have brought about many challenges for answering queries over them. One of the challenges for the users is to be able represent a complex query in a natural language, and get its answers in an understandable form. We address this challenge by introducing a controlled natural language --- a subset of natural language with a restricted grammar and vocabulary --- specifically for expressing biomedical queries towards drug discovery; we call this language BioQueryCNL. The idea is then to transform a biomedical query in BioQueryCNL into a program in answer set programming (ASP) --- a formal framework to automate reasoning about knowledge --- and compute answers to the given query over some biomedical ontologies using a state-of-the-art ASP system. We have developed some algorithms to realize these ideas, and illustrated the applicability of our methods over some biomedical ontologies obtained from various information repositories, such as PHARMGKB, UNIPROT, GO and DRUGBANK.

About the speaker:

Esra Erdem received her Ph.D. in Computer Sciences at the University of Texas at Austin in 2002. She was a post-doctoral fellow in the Cognitive Robotics Group at the University of Toronto (2002-2003), and a research scientist in the Knowledge Based Systems Group at Vienna University of Technology (2003-2006). Since September 2006, she has been an assistant professor at Sabanci University.

Friday, February 5, 2010, 11:00AM



ACES 2.302

Behavior Recognition and Interaction for Cognitive Assistance

Henry Kautz   [homepage]

University of Rochester

The synthesis of research in artificial intelligence, pervasive computing, and human-computer interaction is giving us the means to create systems that recognize human activity from low-level sensor data, interpret that information in light of commonsense theories of behaviors, plans, and goals, and ultimately provide help to the users in a natural and contextually appropriate manner. This talk describes how this approach has been used to develop assistive technologies for persons with cognitive disabilities.

Thursday, February 11, 2010, 11:00AM



ACES 2.402

Detecting Online Credit Card Fraud: A Data Driven Approach

Tal Tversky   [homepage]

Apple, Inc.

A consistent problem plaguing online merchants today is the growth and evolution of online credit card fraud. Thieves harvest credit card numbers from a myriad of sources, place online orders from all over the world, and ship to countless drop locations. Current estimates place the problem at 1 to 1.5 billion dollars annually of which the online merchant holds complete liability.

In this talk, I will describe the problem of eCommerce fraud and outline various detection measures that online merchants employ. Like many merchants, Apple Computer leverages techniques from data mining, machine learning, and statistics to efficiently discover fraud patterns and adapt to new trends. Some topics that I will discuss include evaluating fraud patterns in historic data, building fraud predictive models, inferencing through order linkages, and anomaly detection.

About the speaker:

Dr. Tal Tversky is a Data Mining Scientist at Apple, Inc. He received his Ph.D. in Computer Science from the University of Texas at Austin in 2008. At Apple he works primarily on fighting online credit card fraud, but he has dabbled in forecasting sales and inventory optimization.

The Data Mining Team at Apple in Austin is looking to hire a Ph.D. or master's student as an intern for the summer of 2010. If you are interested in this position, you can sign up to meet with Dr. Tversky or contact him directly at tal "at" apple.com.

Friday, February 19, 2010, 11:00AM



ACES 2.402

Using Machine Learning to Iteratively Improve a Search Heuristic

Robert Holte   [homepage]

University of Alberta

This presentation describes a method for using machine learning to create heuristics for search algorithms such as IDA*. The idea is to iteratively improve upon an initial weak heuristic H1. The easy problems that can be solved using H provide training examples for learning a new heuristic H2 that is expected to be stronger than H1. If H1 is so weak that none of the given instances can be solved with it a method is used to create a sequence of successively more difficult instances starting with ones that are guaranteed to be solvable using H1. The entire process is then repeated using H2 in lieu of H1 to produce H3, and so on, until a sufficiently strong heuristic is produced. In our experimental tests this method produces a heuristic that allows IDA* to solve randomly generated problem instances extremely quickly with solutions very close to optimal.

About the speaker:

Dr. Robert Holte is a professor in the Computing Science Department and Vice Dean of the Faculty of Science at the University of Alberta. He is a well-known member of the international machine learning research community, former editor-in-chief of a leading international journal in this field ("Machine Learning"), and past director of the Alberta Ingenuity Centre for Machine Learning (AICML). His main scientific contributions are his seminal works on the performance of very simple classification rules and a technique ("cost curves") for cost-sensitive evaluation of classifiers. In addition to machine learning he undertakes research in single-agent search (pathfinding), in particular, the use of automatic abstraction techniques to speed up search.

Thursday, February 25, 2010, 11:00AM



ACES 2.402

Time Traveling And Other Small Problems of AI-Based Finance

Astro Teller   [homepage]

Cerebellum Capital, Inc.

Imagine a domain so hard that not only do people believe computers can't do it - many people believe even people can't do it. And this despite existence proofs. A domain structured as a worldwide 24/7 competition in which more than 2 billion people compete directly or indirectly. A single competition that allows participants tens of thousands of choices thousands of times per second and allows competitors to act anywhere on a time frequency range from a microsecond to a year. A domain in which data is everything and no one has ground truth. A domain which is entirely path dependent, has only one history, and in which control experiments are nearly impossible. A competition in which more than $20 trillion dollars are at stake. The field of quantitative finance, while it has attracted criticism over the past few years, is an enormously rich domain for computer science - full of unsolved problems that require both new science and innovative engineering. Cerebellum Capital was founded not to create a platform for smart people to build fast arbitrage systems for making money on the markets, but to architect and build a software system capable of structural search in trading strategy space - a system capable of autonomously finding, testing, refining, launching, improving, and when necessary decommissioning novel trading strategies. This talk will give a tour of the challenges and opportunities we have found as a group of outsiders approaching this domain as a computer science problem.

About the speaker:

Dr. Astro Teller is co-founder and CEO of Cerebellum Capital, a hedge fund management firm whose investments are continuously designed, executed, and improved by a software system based on techniques from statistical machine learning. Astro is also co-founder and Chairman of BodyMedia, Inc, a leading wearable body monitoring company. From 1999 to 2007, Dr. Teller was also Chief Executive Officer of BodyMedia, Inc. Prior to starting BodyMedia in 1999, Dr. Teller was co-founder, Chairman, and CEO of Sandbox Advanced Development, an advanced development technology company. Before his tenure as a business executive, Dr. Teller taught at Stanford University and was an engineer and researcher for Phoenix Laser Technologies, Stanford's Center for Integrated Systems, and The Carnegie Group Incorporated. Dr. Teller holds a Bachelor of Science in computer science from Stanford University, Masters of Science in symbolic and heuristic computation, also from Stanford University, and a Ph.D. in artificial intelligence from Carnegie Mellon University, where he was a recipient of the Hertz fellowship. As a respected scientist and seasoned entrepreneur, Teller has successfully created and grown five organizations and holds numerous patents related to his work in hardware and software technology. Dr. Teller's work in science, literature, art, and business has appeared in international media from the New York Times to CNN to NPR's "All Things Considered." Teller regularly gives invited talks for national and international technology, government, and business forums on the subject of the future of intelligent technology.

Friday, March 5, 2010, 11:00AM



ACES 2.302

Words and Pictures

David Forsyth   [homepage]

UIUC

Many pictures appear with text nearby. The relations between this text and the pictures are complicated and interesting. There are several reasons to study these relations. We could try and predict word annotations from pictures. A good solution to this problem would make it possible to search for unlabelled pictures using keywords - we just label the pictures automatically. Images that have multiple labels are cues to what objects co-occur and what do not; but they are also cues to what image features are stable and what are not. Linguistic phenomena can give strong information about what it means to do object recognition. For example, sentences written about pictures can reveal spatial relations between objects. As another example, adjectives in captions can help focus search for training data. As yet another example, captions sketch the major interesting phenomena in pictures, with reference to a shared discourse between the viewer and the caption writer which reveals information about the picture. A picture titled "This man died in a fire" would show a man, but a picture titled "A man died in this fire" would show a fire. Information of this form might tell us what is worth recognizing and labelling in pictures.

I will review a large set of methods for reasoning about relations between words and pictures, placing a special emphasis on methods that link linguistic phenomena to image observations.

Thursday, March 11, 2010, 2:00PM



ACES 2.402

Asking the Right Questions: New Query Types for Active Learning

Burr Settles   [homepage]

CMU

The key idea behind active learning is that a machine learning algorithm can achieve greater accuracy with less training if it is allowed to choose the data from which it learns. In this talk, I present two recent active learning paradigms in which learning algorithms may pose novel types of "queries" of human annotators to great effect. We call these new paradigms "multiple-instance active learning" and "feature active learning."

In traditional active learning, a partially-trained model selects new data instances to be labeled by a human annotator, which are then added to the training set and the process repeats. In a text classification task, for example, the learner might query for the labels of informative-looking documents. However, having a human read an entire document can be an inefficient use of time, particularly when only certain passages or keywords are relevant to the task at hand. Multiple-instance active learning addresses this problem by allowing the model to selectively obtain more focused labels at the passage level in cases where noisy document-level labels might be available (e.g., from hyperlinks or citation databases). This active learning approach provides a direct training signal to the learner and is also less cumbersome for humans to read. Likewise, feature active learning allows the learner to query for the labels of salient words (e.g., the query word "puck" might be labeled "hockey" in a sports article classification task), which naturally exploits the annotator's inherent domain knowledge. We show that such alternative query paradigms, especially when combined with intuitive user interfaces, can make more efficient use of human annotation effort.

[Joint work with Mark Craven, Soumya Ray, Gregory Druck, and Andrew McCallum.]

Friday, March 12, 2010, 11:00AM



ACES 2.402

Phylogenetic Models for Natural Language

Dan Klein   [homepage]

University of California at Berkeley

Languages descend in a roughly tree-structured evolutionary process. In historical linguistics, this process is manually analyzed by comparing and contrasting modern languages. Many questions arise: What does the tree of languages look like? What are the ancestral forms of modern words? What functional pressures shape language change? In this talk, I'll describe our work on bringing large-scale computational methods to bear on these problems.

In the task of proto-word reconstruction, we infer ancestral words from their modern forms. I'll present a statistical model in which each word's history is traced down a phylogeny. Along each branch, words mutate according to regular, learned sound changes. Experiments in the Romance and Oceanic families show that accurate automated reconstruction is possible; using more languages leads to better results.

Standard reconstruction models assume that one already knows which words are cognate, i.e., are descended from the same ancestral word. However, cognate detection is its own challenge. I'll describe models which can automatically detect cognates (in similar languages) and translations (in divergent languages). Typical translation-learning approaches require virtual Rosetta stones -- collections of bilingual texts. In contrast, I'll discuss models which operate on monolingual texts alone.

Finally, I'll present work on multilingual grammar induction, where many languages' grammars are simultaneously induced. By assuming that grammar parameters vary slowly, again along a phylogenetic tree, we can obtain substantial increases in grammar quality across the board.

About the speaker:

Dan Klein is an associate professor of computer science at the University of California, Berkeley (PhD Stanford, MSt Oxford, BA Cornell). His research focuses on statistical natural language processing, including unsupervised learning methods, syntactic analysis, information extraction, and machine translation. Academic honors include a Marshall Fellowship, a Microsoft New Faculty Fellowship, a Sloan Fellowship, an NSF CAREER award, the ACM Grace Murray Hopper award, and best paper awards at the ACL, NAACL, and EMNLP conferences.

Friday, March 26, 2010, 11:00AM



ACES 2.402

Narrative-Centered Learning Environments

James Lester   [homepage]

North Carolina State University

The long-term goal of the intelligent tutoring systems community is to create adaptive learning technologies that bring about fundamental improvements in education. For the past several years we have been investigating a family of intelligent tutoring systems that have a dual focus on learning effectiveness and student engagement: narrative-centered learning environments. Narrative-centered learning environments marry the inferential capabilities of intelligent tutoring systems with the rich gameplay supported by commercial game engines. In this talk we will introduce the principles motivating the design of narrative-centered learning environments, describe their roots in interactive narrative, explore the role of computational models of affect recognition and affect expression in their interactions, and discuss their cognitive and affective impact on students through empirical studies conducted in public school systems. The discussion will be illustrated with two narrative-centered learning environments, Crystal Island (elementary science education, middle school science education), and the Narrative Theatre (middle school writing).

About the speaker:

Dr. James Lester is Professor of Computer Science at North Carolina State University. He received his Ph.D. in Computer Sciences from the University of Texas at Austin in 1994. He has served as Program Chair for the ACM International Conference on Intelligent User Interfaces (2001), Program Chair for the International Conference on Intelligent Tutoring Systems (2004), Conference Co-Chair for the International Conference on Intelligent Virtual Agents (2008), and on the editorial board of Autonomous Agents and Multi-Agent Systems (1997-2007). His research focuses on intelligent tutoring systems, computational linguistics, and intelligent user interfaces. It has been recognized with a CAREER Award by the National Science Foundation and several Best Paper Awards. His current interests include intelligent game-based learning environments, affective computing, creativity-enhancing technologies, computational models of narrative, and tutorial dialogue. He is Editor-in-Chief of the International Journal of Artificial Intelligence in Education.

Friday, April 9, 2010, 11:00AM



TAY 3.128

Text-Driven Forecasting: Meaning as a Real Number

Noah Smith   [homepage]

CMU

We take inspiration from recent research on sentiment analysis that interprets text based on the subjective attitude of the author. We consider related tasks where a piece of text is interpreted to predict some extrinsic, real-valued outcome of interest that can be observed in non-text data. Examples include:

  • The interpretation of an annual financial report from a company to its shareholders is the risk incurred by investing in the company in the coming year.
  • The interpretation of a critic's review of a film is the film's box office success.
  • The interpretation of a political blog post is the response it garners from readers.
  • The interpretation of a day's microblog feeds is the public's opinion about a particular issue.
In all of these cases, one aspect of the text's meaning is observable from objective real-world data, although perhaps not immediately at the time the text is published (respectively: return volatility, gross revenue, user comments, and traditional polls). We propose a generic approach to text-driven forecasting that is expected to benefit from linguistic analysis while remaining neutral to different theories of language. A highly attractive property of this line of research is that evaluation is objective, inexpensive, and theory-neutral. This approach introduces some methodological challenges, as well.

We conjecture that forecasting tasks, when considered in concert, will be a driving force in domain-specific, empirical, and extrinsically useful natural language analysis. Further, this research direction will push NLP to consider the language of a more diverse subset of the population, and may support inquiry in the social sciences about foreknowledge and communication in societies.

This talk includes joint work with Ramnath Balasubramanyan, William Cohen, Dipanjan Das, Kevin Gimpel, Mahesh Joshi, Shimon Kogan, Dimitry Levin, Brendan O'Connor, Bryan Routledge, Jacob Sagi, and Tae Yano.

About the speaker:

Noah Smith is an assistant professor in the School of Computer Science at Carnegie Mellon University. He received his Ph.D. in Computer Science, as a Hertz Foundation Fellow, from Johns Hopkins University in 2006 and his B.S. in Computer Science and B.A. in Linguistics from the University of Maryland in 2001. His research interests include statistical natural language processing, especially unsupervised methods, machine learning for structured data, and applications of natural language processing. He serves on the editorial board of the journal Computational Linguistics and received a best paper award at the ACL 2009 conference. His ten-person group, Noah's ARK, is supported by the NSF, DARPA, Qatar NRF, Portugal FCT, and gifts from Google, HP Labs, and IBM Research.

Wednesday, April 14, 2010, 11:00AM



ACES 2.302

Automatic Program Repair with Evolutionary Computation

Stephanie Forrest   [homepage]

University of New Mexico

There are many methods for detecting and mitigating software errors but few generic methods for automatically repairing errors once they are discovered. The talk will describe an automated method for repairing errors in off-the-shelf, legacy programs without formal specifications, program annotations, or special coding practices. The method uses an extended form of genetic programming to evolve a program variant that retains required functionality but is not susceptible to the error. We use existing test suites to encode both the error and required functionality. The talk will describe the algorithm and summarize experimental results on 15 programs totaling 1.2M lines of C code. If time permits, the talk will also describe recent results combining the method with intrusion detection to form a closed-loop repair system and extensions of the method to assembly code programs.

About the speaker:

Stephanie Forrest is Professor and Chairman of the Computer Science Department at the University of New Mexico in Albuquerque. She is also an External Professor of the Santa Fe institute and has served as its Vice President and a member of the Science Board. Her research studies adaptive systems, including immunology, evolutionary computation, biological modeling, and computer security. Professor Forrest received M.S. and Ph.D. degrees in Computer and Communication Sciences from the University of Michigan (1982,1985) and a B.A. from St. John's College (1977). Before joining UNM in 1990 she worked for Teknowledge Inc. and was a Director's Fellow at the Center for Nonlinear Studies, Los Alamos National Laboratory.

Friday, April 23, 2010, 11:00AM



TAY 3.128

Representations for object and action recognition

Martial Hebert   [homepage]

CMU

Recognizing objects and actions is a key objective of computer vision research. Extensive work has been done in the past, but the resulting systems are still brittle. In this talk, I will review some of the computer vision projects that we have recently undertaken in those areas. I will show (hopefully) new approaches which enable higher performance than existing approaches or to address tasks that cannot be currently addressed.

Insights used in these approaches include the use of reasoning techniques as a throw back to earlier days of computer vision, better use of 3D geometry and shape instead or in addition to appearance, representation of structural information instead of statistical information, e.g., using geometric representation on tracked features for action recognition, and use of "first-person" visual data in which the images are acquired from a user's perspective rather than from an outside perspective, as is the case in surveillance applications, for example.

Thursday, April 29, 2010, 11:00AM



ACES 2.402

Robots as Social Learners

Cynthia Breazeal   [homepage]

MIT

As personal robots enter our workplaces and homes, it will be important for them to learn new tasks and abilities from a wide demographic of people. Ideally, people will be able to teach robots as naturally as one another. Consequently, robots should be socially competent enough to take advantage of the same sorts of interpersonal cues and skills that humans readily use to teach and learn. Our research seeks to identify simple, natural, and prevalent teaching cues and then program robots with social-affective mechanisms to enable them to learn efficiently and effectively from natural interactions. In this talk, I present several social skills implemented on our robots and discuss how they address the challenge of building robots that learn from people. These skills include the ability to direct attention, to understand affect and intent, to express its learning process to the human instructor, and to regulate its interaction with the instructor. Through these examples, we show how social, emotional, and expressive factors can be used in interesting ways to build robots that learn from people in a manner that is natural for people to teach.

About the speaker:

Dr. Cynthia Breazeal is an Associate Professor of Media Arts and Sciences at the Massachusetts Institute of Technology where she founded and directs the Personal Robots Group at the Media Lab and is director of the Center for Future Storytelling. She is a pioneer of Social Robotics and Human Robot Interaction (HRI). Her research program focuses on developing personal robots and interactive characters that engage humans in human-centric terms, work with humans as partners, and learn from people via tutelage. More recent work investigates the impact of long term HRI applied to entertainment, communication, quality of life, health, and educational goals. She has authored the book "Designing Sociable Robots" and has published over 100 peer-reviewed articles in journals and conferences in autonomous robotics, artificial intelligence, human robot interaction, and robot learning. She has been awarded an ONR Young Investigator Award, honored as finalist in the National Design Awards in Communication, and recognized as a prominent young innovator by the National Academy of Engineering's Gilbreth Lecture Award. She received her ScD in Electrical Engineering and Computer Science from MIT in 2000.

Friday, April 30, 2010, 11:00AM



ACES 2.402

Learning, Reasoning, and Action in the Open World

Eric Horvitz   [homepage]

Microsoft Research

Systems that sense, learn, and reason from streams of data promise to provide extraordinary value to people and society. Harnessing computational principles to build systems that operate in the open world can also teach us about the sufficiency of existing models, and frame new directions for research. I will discuss efforts on perception, learning, and inference in the open world, highlighting key ideas in the context of projects in transportation, energy, and healthcare. Finally, I will discuss opportunities for building systems with new kinds of open-world competencies by weaving together components that leverage advances from several research subdisciplines.

About the speaker:

Eric Horvitz is a Distinguished Scientist and Research Area Manager at Microsoft Research. His interests span theoretical and practical challenges with developing systems that perceive, learn, and reason. He is a Fellow of the AAAI and of the AAAS and is serving as the Immediate Past President of the AAAI. He has also served on the NSF Computer & Information Science & Engineering (CISE) Advisory Board, the Computing Community Consortium (CCC), the DARPA Information Science and Technology Study Group (ISAT), and the Naval Research Advisory Committee (NRAC). He received his PhD and MD degrees at Stanford University. More information can be found at research.microsoft.com/~horvitz.

Monday, May 24, 2010, 11:00AM



ACES 2.402

Computational Study Of Nonverbal Social Communication

Louis-Philippe Morency   [homepage]

USC

The goal of this emerging research field is to recognize, model and predict human nonverbal behavior in the context of interaction with virtual humans, robots and other human participants. At the core of this research field is the need for new computational models of human interaction emphasizing the multi-modal, multi-participant and multi-behavior aspects of human behavior. This multi-disciplinary research topic overlaps the fields of multi-modal interaction, social psychology, computer vision, machine learning and artificial intelligence, and has many applications in areas as diverse as medicine, robotics and education.

During my talk, I will focus on three novel approaches to achieve efficient and robust nonverbal behavior modeling and recognition: (1) a new visual tracking framework (GAVAM) with automatic initialization and bounded drift which acquires online the view-based appearance of the object, (2) the use of latent-state models in discriminative sequence classification (Latent-Dynamic CRF) to capture the influence of unobservable factors on nonverbal behavior and (3) the integration of contextual information (specifically dialogue context) to improve nonverbal prediction and recognition.

About the speaker:

Dr. Louis-Philippe Morency is currently research assistant professor at the University of Southern California (USC) and research scientist at USC Institute for Creative Technologies where he leads the Multimodal Communication and Computation Laboratory (MultiComp Lab). He received his Ph.D. from MIT Computer Science and Artificial Intelligence Laboratory in 2006. His main research interest is computational study of nonverbal social communication, a multi-disciplinary research topic that overlays the fields of multi-modal interaction, machine learning, computer vision, social psychology and artificial intelligence. He developed "Watson", a real-time library for nonverbal behavior recognition and which became the de-facto standard for adding perception to embodied agent interfaces. He received many awards for his work on nonverbal behavior computation including three best-paper awards in 2008 (at various IEEE and ACM conferences). He was recently selected by IEEE Intelligent Systems as one of the "Ten to Watch" for the future of AI research.

[ FAI Archives ]

Fall 2013 - Spring 2014

Fall 2012 - Spring 2013

Fall 2011 - Spring 2012

Fall 2010 - Spring 2011

Fall 2009 - Spring 2010

Fall 2008 - Spring 2009

Fall 2007 - Spring 2008

Fall 2006 - Spring 2007

Fall 2005 - Spring 2006

Spring 2005

Fall 2004

Spring 2004

Fall 2003

Spring 2003

Fall 2002

Spring 2002

Fall 2001

Spring 2001

Fall 2000

Spring 2000