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 Nick Jong, Lily Mihalkova, or Dan Tecuci . |
Upcoming talks
Past talks
Friday, September 7, 11:00am
|
Modeling the Reasoning of Agents in GamesDr. Avi Pfeffer [homepage]
[Sign-up schedule for individual meetings]Why do agents (people or computers) do things in strategic situations? Answering this question will impact how we build computer systems to assist, represent or interact with people in interactions with other agents such as negotiations and resource allocation. We identify four reasoning patterns that agents might use: choosing an action for its direct effect on the agent's utility, attempting to manipulate another agent, signalling information to another agent that the first agent knows, or revealing or hiding information from another agent that the first agent itself does not know. We present criteria that characterize each reasoning pattern as a pattern of paths in a multi-agent influence diagram, a graphical representation of games. We define a class of strategies in which agents do not make unmotivated distinctions, and show that if we assume all agents play these kinds of strategies, our categorization of reasoning patterns is complete and captures all situations in which an agent has reason to make a decision. We then study how people use two reasoning patterns in a particular negotiation game. We use machine learning to learn models of people's play, and embed our learned models in computer negotiators. We find that negotiators that use our learned model outperform classical game-theoretic agents and also outperform people. Finally, we learn models of the way people's behavior changes in ongoing interactions with the same agent, particularly the degree to which retrospective (rewarding or punishing past behavior) and prospective (attempting to induce future good behavior) reasoning play a role. About the speaker:Avi Pfeffer is Associate Professor at Computer Science at Harvard University. His research is directed towards achieving rational behavior in intelligent systems, based on the principles of probability theory, decision theory, Bayesian learning and game theory. He received his PhD in 2000 from Stanford University, where his dissertation on probabilistic reasoning received the Arthur Samuel Thesis Award. Dr Pfeffer has published technical papers on probabilistic reasoning, strategic reasoning, agent modeling, temporal reasoning, and database systems. He was awarded the NSF Career Award in 2001 for work on strategic reasoning, and the Alfred P. Sloan Foundation Research Fellowship in 2002. |
Friday, September 28, 11:00am
|
Regret-based Methods for Preference Elicitation and Mechanism DesignDr. Craig Boutilier [homepage]
[Sign-up schedule for individual meetings]Preference elicitation is generally required when making or recommending decisions on behalf of users whose utility function is not known with certainty. Although one can engage in elicitation until a utility function is perfectly known, in practice, this is infeasible. In this talk, I explore the use of minimax regret as (a) a distribution-free means for making decisions with imprecise utility information; and (b) a means for guiding elicitation in a way that focuses only on relevant aspects of a user's preferences. The talk will develop efficient integer programming approaches to this problem and heuristic techniques for elicitation. Preference elicitation is, of course, an important component of (economic) mechanism design as well. Classical approaches to mechanism design require participants to fully reveal their utility functions. Time permitting, I will sketch some recent results on the use of minimax regret in the automated design of partial revelation mechanisms. With only partial revelation of preferences, we provide bounds on both incentive and outcome quality by generalizing VCG. (This talk describes joint work with various collaborators.) About the speaker:Craig Boutilier received his Ph.D. in Computer Science (1992) from the University of Toronto, Canada. He is Professor and Chair of the Department of Computer Science at the University of Toronto. He was previously an Associate Professor at the University of British Columbia, a consulting professor at Stanford University, and has served on the Technical Advisory Board of CombineNet, Inc. since 2001. Dr. Boutilier's research interests span a wide range of topics, with a focus on decision making under uncertainty, including preference elicitation, mechanism design, game theory, Markov decision processes, and reinforcement learning. He is a Fellow of the American Association of Artificial Intelligence (AAAI) and the recipient of the Isaac Walton Killam Research Fellowship, an IBM Faculty Award and the Killam Teaching Award. |
Friday, October 26, 11:00am
|
Simulation Semantics: A Computational framework for exploring t he links between Language, Cognition and ActionDr. Srini Narayanan [homepage]
[Sign-up schedule for individual meetings]The UCB/ICSI NTL (http://www.icsi.berkeley.edu/NTL ) project is an ongoing attempt to model language behavior in a way that is both neurally plausible and computationally practical. Work within the NTL project coupled with a variety of converging evidence from Cognitive Linguistics, Psychology and Neuroscience suggests that language understanding involves embodied enactment which we call "simulation semantics".Simulation semantics hypothesizes the mind as "simulating" the external world while functioning in it. The "simulation" takes noisy linguistic input together with general knowledge and makes new inferences to figure out what the input means and to guide response. Monitoring the state of the external world, drawing inferences, and acting jointly constitute a dynamic ongoing interactive process. This talk reports on a computational realization of the simulation semantics hypothesis and preliminary results on applying the model to vexing problems in Natural Language Understanding. About the speaker:Srini Narayanan leads the strongly interdisciplinary ICSI (http://www.icsi.berkeley.edu) AI group comprising of computer scientists, linguists, and neuroscientists studying language and cognition. Srini Narayanan is a core faculty member in Cognitive Science (http://ls.berkeley.edu/ugis/cogsci) and an Adjunct Professor at the Institute for Cognitive and Brain Sciences (http://icbs.berkeley.edu) at the University of California, Berkeley. His current research interests include cognitive computation, computational biology, natural language understanding, and information technology for developing regions. Prof. Narayanan is the recipient (with Dan Jurafsky) of a David Marr distinguished paper award, a Google faculty research award, and a Fellowship at the Institute for Advanced Study, Berlin (http://www.wiko-berlin.de). |
Friday, November 16, 11:00am
|
Inducing Semantic Taxonomies for WordsDr. Dan Jurafsky [homepage]
[Sign-up schedule for individual meetings]Online resources for word meaning like dictionaries and thesauri are a useful resource for natural language processing. But English has a lot of words, and adds more every day. Hand-built resources can't keep up. We report on three studies on inducing the meaning of words from text on the Web in the context of augmenting WordNet, a large online thesaurus of English. We first describe a semi- supervised method for learning when a new word is a `hypernym' or in the 'is-a' relation with another word, based on combining `weak learner' hypernym detection patterns. We then show a new probabilistic algorithm for taxonomy induction which incorporates evidence from multiple relation detectors (for hyponymy and synonymy) to optimize the entire structure of the taxonomy. Finally we show how to improve the granularity of word senses in a thesaurus by clustering word senses. This talk describes joint work with Rion Snow and Andrew Ng.About the speaker:Dan Jurafsky is an associate professor in the Department of Linguistics, and by courtesy in the Department of Computer Science, at Stanford University. Before Stanford he taught for 8 years at the University of Colorado, Boulder. Dan studies statistical models of human and machine processing of text and speech, focusing lately on computational models of semantics and discourse, and on conversational speech. |
Friday, November 23, 2:00pm
|
ASIMO Revisited—The Quest for IntelligenceDr. Edgar Koerner
[Sign-up schedule for individual meetings]Having largely succeeded in the development of humanoid mobility and the adaptation of humanoid robots to the human environment, providing those artefacts with capabilities for autonomous interaction with its environment challenges our current understanding of intelligence. Still, the brain is the only one existing system that is capable of generating intelligent behaviour in real world and real time. Therefore, understanding how the brain works may provide us with an idea of what are the essential principles behind that marvellous performance we admire. In our approach to understand essential characteristics of processing in the brain we follow an “analysis by synthesis” philosophy. We believe not special local processing algorithms and dynamics, but how the brain organises processing is the key to enable that marvellous performance and robustness in sensory understanding and creating smart behaviour in an unpredictable environment, which is still unparalleled by any technical system. Therefore, the major focus of our approach is on the global and local control in the brain that enables autonomous learning and the self-organisation of knowledge representation. We target the control architecture of brain-like intelligent systems at microscopic, mesoscopic and macroscopic level in parallel, whereby special emphasis is put on the utilisation of insights from phylogenetic and ontogenetic development. Step by step we try to assemble systems that interact with the environment and that provide us with the necessary feedback on the role of architectural constraints and dynamics of the brain. Finally, to address the global control architecture, we integrate visual and auditory recognition with autonomous behaviour generation on ASIMO for a step by step development of individual cognitive capabilities based on the practical experience in interacting with the real world. |
Friday, November 30, 11:00am
|
Structured PredictionDr. Ben Taskar [homepage]
[Sign-up schedule for individual meetings]Structured prediction is a framework for solving problems of classification or regression in which the output variables are mutually dependent or constrained. These dependencies and constraints reflect sequential, spatial or combinatorial structure in the problem domain, and capturing such interactions is often as important as capturing input-output dependencies. Many such problems, including natural language parsing, machine translation, object segmentation, gene prediction, protein alignment and numerous other tasks in computational linguistics, speech, vision, biology, are not new. However, recent advances have brought about a unified view, efficient methodology and more importantly, significant accuracy improvements for both classical and novel problems. I will outline the fundamental computational and statistical methods and challenges in structured prediction.About the speaker:Ben Taskar is Assistant Professor at the University of Pennsylvania. His primary research interests are machine learning and applications to computational linguistics, computer vision, and computational biology. |
Friday, December 7, 11:00am
|
Scientific Data Mining: The Discovery and Use of Complex Networks in Neuroscience and GenomicsDr. Terran Lane [homepage]
[Sign-up schedule for individual meetings]Modern science is overwhelmed by a sea of data. Recent years have brought us sensor technologies that produce gigabytes to terabytes of information per experiment: functional neuroimaging technologies, genetic microarrays and high-throughput assays, digital telescopes, and environmental sensor networks, to name just a few. These technologies offer unprecedented opportunity for scientific discovery to the domain scientists. Yet at the same time, they present a daunting analysis task: to extract meaningful, substantiable patterns from this overwhelming mass of data. Further, the data are typically extremely noisy and the patterns of interest are often multivariate and nonlinear.To address these analysis problems, computer scientists in the machine learning and data mining communities have been developing the field of scientific data mining: using advanced computational and statistical tools to extract complex patterns from large, difficult, scientific data sets. In this talk, I will give an overview of my recent work on scientific data mining in two different domains: neuroscience and genomics. On the former front, I will discuss the problem of network identification: finding the network of functional activity interactions that underlies some behavioral pattern. The ability to find such networks is critical to neuroscientists who are working to understand mental illnesses such as dementia or schizophrenia. On the latter front, I will discuss the task of biological parameter estimation for RNA interference (RNAi). In this case, we use the structure of known activity networks to infer parameters of the biological process that produced it. These parameters, in turn, help biologists and pharmacists develop better RNAi-based genetic screens and pharmaceuticals. About the speaker:Terran Lane is Assistant Professor of computer science at the University of New Mexico. His primary academic interests are: machine learning; reinforcement learning, behavior, and control; and artificial intelligence in general. He is also interested in computer/information security/privacy and bioinformatics. |
Friday, January 18, 11:00am
|
Robust Computational SemanticsDr. Johan Bos [homepage]
[Sign-up schedule for individual meetings]Formal methods for the analysis of the meaning of natural language expressions have long been restricted to the ivory tower built by semanticists, logicians, and philosophers of language. It is only in exceptional cases that they make their way straight into natural language processing tools. Recently, this situation has changed. Thanks to the development of treebanks (large collections of texts annotated with syntactic structures), robust statistical parsers trained on such treebanks, and the development of large-scale semantic lexica, we now have at our disposal systems that are able to produce formal semantic representations achieving very high coverage. Even only a few years ago this was pure fantasy. This is an interesting development (and result) and shouldn't be left unnoticed. It makes formal semantics accessible to practical natural language processing, and opens the door to using inference tools developed in the area of automated deduction, such as theorem provers and model builders for first-order logic. Most of all, it identifies potential gaps between theory and practice, and forces the computational semanticist to transfer theoretical ideas developed in isolation into one unifying framework aiming at covering a wide variety of semantic phenomena. In this talk I present such a system developed by myself over the last three years. This system, Boxer, implements Discourse Representation Theory (DRT), a formal theory of meaning, with the help of Combinatory Categorial Grammar (CCG) for producing syntactic structure, and a typed lambda calculus to specify the syntax-semantics interface. In conjunction with a robust CCG parser, Boxer achieves very high coverage on newswire text producing first-order representations that can directly feed into standard automated theorem provers. The existence of Boxer is clear evidence that practicing semantics is not bound to pencil and paper exercises anymore, nor to implementations covering only "baby fragments" of English. Overall, from the perspective of computational semantics, this is good news. It demonstrates that we have made substantial progress. But Boxer, however impressive it may be, has many shortcomings too. Modelling all nuances of meaning is an immense task—perhaps even impossible—and what Boxer does, as any rival system, is only produce an approximation of the meaning of an input text. An interesting question to ask then is how good this approximation is. How do we access the semantic adaquacy of systems like Boxer that claim are able to compute meaning? This issue I will address in the second part of the talk. About the speaker:Johan Bos got his first degree in Computational Linguistics from the University of Groningen in 1993. He then moved on to Saarbruecken, Germany, and completed his PhD there in 2001. After that he took up a post-doc position at the University of Edinburgh at the Informatics Department. He is currently enjoying a fellowship funded by the Italian ministry and since 2005 based at the University of Rome "La Sapienza". His research interests comprise almost all aspects of natural language processing, in particular those that concern semantics. |
Tuesday, February 5, 2:00pm
|
Bayesian Methods for Unsupervised Language LearningDr. Sharon Goldwater [homepage]
Unsupervised learning of linguistic structure is a difficult task. Frequently, standard techniques such as maximum-likelihood estimation yield poor results or are simply inappropriate (as when the class of models under consideration includes models of varying complexity). In this talk, I discuss how Bayesian statistical methods can be applied to the problem of unsupervised language learning to develop principled model-based systems and improve results. I first present some work on word segmentation, the problem of identifying word boundaries in continuous text or speech. I show that maximum-likelihood estimation is inappropriate for this task and discussing a nonparametric Bayesian modeling solution. I then argue, using part-of-speech tagging as an example, that a Bayesian approach provides advantages even when maximum-likelihood (or maximum a posteriori) estimation is possible. I conclude by discussing some of the challenges that remain in pursuing a Bayesian approach to language learning.About the speaker:Sharon Goldwater is a postdoctoral scholar in the linguistics department at Stanford University, where she works with Dan Jurafsky, Chris Manning, and others in the Stanford natural language processing group. Her research focuses on unsupervised learning and computer modeling of language acquisition, particularly phonology and morphology. She completed her master's degree in computer science in 2005 and her Ph.D. in linguistics in 2006, both from Brown University. Prior to graduate school, she worked as a researcher in the Artificial Intelligence Laboratory at SRI International. |
Tuesday, February 19, 2:00pm
|
Methods for Predicting IDA*'s PerformanceDr. Robert Holte [homepage]
The goal of the research presented in this talk is to accurately predict the number of nodes that the heuristic search algorithm IDA* will expand for a given depth bound, heuristic, and start state or set of start states. The talk begins by describing the landmark paper by Mike Reid and Rich Korf in AAAI 1998. This paper presented a simple but very effective analysis framework and developed a formula that was shown experimentally to make almost perfect predictions. This prediction method has two shortcomings:
About the speaker:Professor Robert Holte is a well-known member of the international machine learning research community, former editor-in-chief of the leading international journal in this field (Machine Learning), and current director of the Alberta Ingenuity Centre for Machine Learning. His main scientific contributions are his seminal works on the problem of small disjuncts and the performance of very simple classification rules. His current machine learning research investigates cost-sensitive learning and learning in game-playing (for example: opponent modelling in poker, and the use of learning for gameplay analysis of commercial computer games). 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. He has over 55 scientific papers to his credit, covering both pure and applied research, and has served on the steering committee or program committee of numerous major international AI conferences. |
Friday, February 29, 11:00am
|
Tractable Decision Making in Some Partially Observable DomainsDr. Eyal Amir [homepage]
[Sign-up schedule for individual meetings]Many complex domains offer limited information about their exact state and the way actions affect them. There, autonomous agents need to make decisions at the same time that they learn action models and track the state of the domain. This combined problem can be represented within the framework of reinforcement learning in POMDPs, and is known to be computationally difficult. In this presentation I will describe a new framework for such decision making, learning, and tracking. This framework applies results that we achieved about updating logical formulas (belief states) after deterministic actions. It includes algorithms that represent and update the set of possible action models and world states compactly and tractably. It makes a decision with this set, and updates the set after taking the chosen action. Most importantly, and somewhat surprisingly, the number of actions that our framework takes to achieve a goal is bounded polynomially by the length of an optimal plan in a fully observable, fully known domain, under lax conditions. Finally, our framework leads to a new stochastic-filtering approach that has better accuracy than previous techniques.About the speaker:Eyal Amir is an Assistant Professor of Computer Science at the University of Illinois at Urbana-Champaign (UIUC) since January 2004. His research includes reasoning, learning, and decision making with logical and probabilistic knowledge, dynamic systems, and commonsense reasoning. Before UIUC he was a postdoctoral researcher at UC Berkeley (2001-2003) with Stuart Russell, and did his Ph.D. on logical reasoning in AI with John McCarthy. He received B.Sc. and M.Sc. degrees in mathematics and computer science from Bar-Ilan University, Israel in 1992 and 1994, respectively. Eyal is a Fellow of the Center for Advanced Studies and of the Beckman Institute at UIUC (2007-2008), was chosen by IEEE as one of the "10 to watch in AI" (2006), received the NSF CAREER award (2006), and awarded the Arthur L. Samuel award for best Computer Science Ph.D. thesis (2001-2002) at Stanford University. |
Wednesday, March 5, 11:00am
|
A Framework for View Learning in Statistical Relational LearningDr. Vítor Santos Costa [homepage]
[Sign-up schedule for individual meetings]Statistical relational learning (SRL) algorithms learn statistical models from relational data, such as that stored in a relational database. In prior work we introduced view learning for SRL, in which the view of a relational database can be automatically modified, yielding more accurate statistical models. In this talk I present SAYU-VISTA, an algorithm which advances beyond the initial view learning approach in three ways. First, it learns views that introduce new relational tables, rather than merely new fields for an existing table of the database. Second, new tables or new fields are not limited to being approximations to some target concept; instead, the new approach performs a type of predicate invention. The new approach avoids the classical problem with predicate invention, of learning many useless predicates, by keeping only new fields or tables (i.e., new predicates) that immediately improve the performance of the statistical model. Third, retained fields or tables can then be used in the definitions of further new fields or tables. We evaluate the new view learning approach on three relational classification tasks.About the speaker:Vítor Santos Costa is a Professor at the Universidade do Porto. He was previously a Professor at Universidade Federal do Rio de Janeiro and twice a visiting professor for a year at the University of Wisconsin-Madison. His research interests are on the area of Logic Programming implementation, where he maintains and develops the YAP Prolog system, and on Inductive Logic Programming and Statistical Relational Learning, where he participated in the development of CLP(BN), SAYU and VISTA. |
Wednesday, March 19, 1:30pm
|
How Analytics Is Revolutionizing E-CommerceDr. Joseph Sirosh
[Sign-up schedule for individual meetings]Machine learning and data mining applications are central to eCommerce. Recommendation systems, online risk management, behavioral ad targeting, targeted web and email marketing, online self-service systems, search relevance ranking etc. are just a few of the applications that make extensive use of data mining techniques to enable a unique customer experience on the web. This talk will explore the enormous wealth of opportunities for leveraging data to enhance the eCommerce experience. Using examples drawn from leading web sites, I will provide an overview of how some of these systems work and how they generate great value for customers. We'll then look to some emerging trends and what the future holds for data mining and machine learning on the web. About the speaker:Joseph Sirosh is currently Vice President of Business Intelligence and Transaction Risk Management at Amazon.com. Prior to joining Amazon.com he worked at Fair Isaac Corporation as VP of the Advanced Technology R&D group, exploring advanced analytic and data mining applications. At Fair Isaac and at HNC Software prior to that, he has led several significant R&D projects on security and fraud detection, predictive modeling, information retrieval, content management, intelligent agents and bioinformatics. He has made important contributions in the field of neural network algorithms and in understanding the fundamental principles by which information is organized and processed in the brain. Dr. Sirosh has published over 20 technical papers and one book, and has been a lead investigator of various research grants from DARPA and other Government agencies. He holds a PhD and Masters in Computer Science from the University of Texas at Austin, and a B. Tech. in Computer Science & Engineering, from the Indian Institute of Technology, Chennai. |
Monday, March 31, 3:00pm
|
Discriminative Synchronous Transduction for Statistical Machine TranslationDr. Miles Osborne [homepage]
[Sign-up schedule for individual meetings]Large-scale discriminative machine translation promises to further the state-of-the-art but has failed to deliver convincing gains over current (mainly) generative systems. We argue that a principle reason for this failure is not dealing with multiple, equivalent ways of producing the same translation for a given source sentence. We present a translation model based upon a synchronous context free grammar. By explicitly summing-out all possible derivations, we are able to improve translation results. Estimation is exact, since we can employ dynamic programming techniques used in the log-linear parsing literature when calculating expectations. Decoding however is only approximate, since no tractable algorithms exist for marginalising-out derivations.Joint work with Phil Blunsom and Trevor Cohn. About the speaker:Miles Osborne is on the faculty at Informatics, Edinburgh University. His main interests are in machine learning applied to natural language, statistical machine translation, and more recently, randomised algorithms. He has started looking at processing Blogs using Hadoop, a handful of derelict machines and some cardboard boxes. |
Tuesday, April 1, 2:30pm
|
Appearance Based Navigation and the FAB-MAP AlgorithmDr. Paul Newman [homepage]
[Sign-up schedule for individual meetings]This talk considers an appearance-based topological approach to mobile robotic navigation and mapping. We shall introduce a new algorithm - Fast Appearance Based Mapping (FAB-MAP) - which is capable of building large scale (>>km) topological maps and detecting loop closure with a cost linear in the size of the map. Loop closing is the problem of correctly asserting that a robot has returned to a previously visited area. It is a particularly hard but important component of the Simultaneous Localization and Mapping (SLAM) problem. Here a mobile robot explores an a-priori unknown environment performing on-the-fly mapping while the map is used to localize the vehicle. Many SLAM implementations look to internal map and vehicleestimates to make decisions about whether a vehicle is revisiting a previously mapped area or is exploring a new region of workspace. We suggest that one of the reasons loop closing is hard in SLAM is precisely because these internal estimates can, despite best efforts, be in gross error. FAB-MAP makes no recourse to the metric estimates of the SLAM system it supports and aids—it is entirely independent. We illustrate the effectiveness of the algorithm on several outdoor and indoor data sets producing both purely topological maps and by integrating topological constraints within large scale metric SLAM maps built with 3D laser data. About the speaker:Paul Newman obtained an M.Eng. in Engineering Science from Oxford University in 1995. He then undertook a Ph.D. in autonomous navigation at the Australian Center for Field Robotics, University of Sydney, Australia. In 1999 he returned to the United Kingdom to work in the commercial sub-sea navigation industry. In late 2000 he joined the Dept of Ocean Engineering at M.I.T. where as a post-doc and later a research scientist, he worked on algorithms and software for robust autonomous navigation for both land and sub-sea agents. In early 2003 he returned to Oxford as a Departmental Lecturer in Engineering Science before being appointed to a University Lectureship in Information Engineering and becoming a Fellow of New College in 2005. He heads the Oxford Mobile robotics Research group and has research interests in pretty much anything to do with autonomous navigation but particularly Simultaneous Localisation and Mapping. He is on the editorial board of the International Journal of Robotics Research and The Journal of Field Robotics and a IEEE R.A.S European Distinguished Lecturer for 2008 and 2009. |
Monday,
|
Using Echo State Networks for Anomaly Detection in Underground Coal MinesDr. Oliver Obst [homepage]
[Sign-up schedule for individual meetings]We investigated the problem of identifying anomalies in monitoring critical gas concentrations using a sensor network in an underground coal mine. In this domain, one of the main problems is a provision of mine specific anomaly detection, with cyclical (moving) instead of flatline (static) alarm threshold levels. An additional practical difficulty in modelling a specific mine is the lack of fully labelled data of normal and abnormal situations. In my talk, I'll present an approach addressing these difficulties based on echo state networks learning mine specific anomalies when only normal data is available. Echo state networks utilize incremental updates driven by new sensor readings, thus enabling a detection of anomalies at any time during the sensor network operation. We evaluate this approach against a benchmark—Bayesian network based anomaly detection, and observe that the quality of the overall predictions is comparable to the benchmark. However, the echo state networks maintain the same level of predictive accuracy for data from multiple sources. Therefore, the ability of echo state networks to model dynamical systems make this approach more suitable for anomaly detection and predictions in sensor networks. (This is joint work with X. Rosalind Wang and Mikhail Prokopenko.) About the speaker:Oliver Obst obtained a Diplom in Computer Science (Informatics) at the University of Koblenz, Germany, in 1999. At the beginning of 2006, he completed his Ph.D. in Artificial Intelligence at the same University. Before joining the Adaptive Systems Group at CSIRO ICT Centre, Sydney in 2007, he worked as a postdoctoral researcher at the Intelligent Systems Department, University of Bremen, Germany, and at the Interdisciplinary Machine Learning Research Group, University of Newcastle, Australia. His research interest include simulation, prediction and anomaly detection. |
Friday,
|
Processing Code-Switched TextDr. Thamar Solorio [homepage]
[Sign-up schedule for individual meetings]Code-switching is an interesting linguistic phenomenon commonly observed in highly bilingual communities. It consists of mixing languages in the same conversational event. Despite its popularity, this type of discourse has received very little attention from the natural language processing community. Most of the work in this area attempts to solve problems where the language samples, either spoken or written, are monolingual. We recently started working on developing a part-of-speech tagger for Spanish-English code-switched text. In the first half of this talk I will discuss results of different approaches to solve the tagging problem by taking advantage of existing resources for both languages. The long-term goal of this research is to develop a full syntactic parser for English-Spanish code-switched text, commonly known as Spanglish, that can be exploited to tackle higher-level tasks on mixed-language sources. Although the work is focused on English-Spanish bilingual discourse, the knowledge acquired from this project can later be extended to other language combinations. In the second half, I will discuss a related project aimed at exploiting our bilingual tagger to develop an automated screening tool for the early identification of Specific Language Impairment in Spanish-English bilingual children. About the speaker:Thamar Solorio is a postdoctoral scholar in the Human Language Technology Research Institute at the University of Texas at Dallas. Before joining UTD she was a Lecturer in the Computer Science department at the University of Texas at El Paso. She received her PhD in Computer Science in 2005 from the National Institute of Astrophysics, Optics and Electronics, in Mexico. She is interested in developing machine learning approaches for the syntactic analysis of interlanguages. |
Friday, May 23, 11:00am
|
Object Recognition by Scene AlignmentDr. Antonio Torralba [homepage]
[Sign-up schedule for individual meetings]Object detection and recognition is generally posed as a matching problem between the object representation and the image features (e.g., aligning pictorial cues, shape correspondence, constellations of parts, etc.) while rejecting the background features using an outlier process. In this work, we take a different approach: we formulate the object detection problem as a problem of aligning elements of the entire scene. The background, instead of being treated as a set of outliers, is used to guide the detection process. Our approach relies on the observation that when we have a big enough database then we can find with high probability some images in the database very close to a query image, as in similar scenes with similar objects arranged in similar spatial configurations. If the images in the retrieval set are partially labeled, then we can transfer the knowledge of the labeling to the query image, and the problem of object recognition becomes a problem of aligning scene regions. But, can we find a dataset large enough to cover a large number of scene configurations? Given an input image, how do we find a good retrieval set, and, finally, how we do transfer the labels to the input image? We will use two datasets; 1) the LabelMe dataset, which contains more than 10,000 labeled images with over 180,000 annotated objects. 2) The tiny images dataset: A dataset of weakly labeled images with more than 79,000,000 images. We use this database to perform object and scene classification, examining performance over a range of semantic levels. Work in collaboration with Rob Fergus, Bryan Russell, Ce Liu and William T. Freeman Additional information and links to relevant papers can be found at: URL: http://people.csail.mit.edu/torralba/tinyimages/ About the speaker:Antonio Torralba joined the Department of Electrical Engineering and Computer Science at MIT in August 2007 as an Assistant Professor and a member of the Computer Science and Artificial Intelligence Laboratory. He earned a degree in telecommunications engineering from the Technical University of Catalonia (Spain) in 1994. He received his Ph.D. in Image and Signal Processing from the Institut National Polytechnique, Grenoble, France, in 2000. Then, he moved to Boston as a postdoctoral fellow at MIT. In 2004, he became a research scientist at the Computer Science and Artificial Intelligence Laboratory. |
Wednesday, June 18, 11:00am
|
Synthesis of Strategies from Interaction Traces and Coping with Noise in Non-Zero-Sum GamesTsz-Chiu Au [homepage]
To create new and better agents in multi-agent environments, we may want to examine the strategies of several existing agents, in order to combine their best skills. One problem is that in general, we won't know what those strategies are; instead, we'll only have observations of the agents' interactions with other agents. In this talk, I describe how to take a set of interaction traces produced by different pairs of players in a two-player repeated game, and then find the best way to combine them into a composite strategy. I also describe how to incorporate the composite strategy into an existing agent, as an enhancement of the agent's original strategy. In cross-validated experiments involving 126 agents (most of which written by students as class projects) for the Iterated Prisoner's Dilemma, Iterated Chicken Game, and Iterated Battle of the Sexes, composite strategies produced from these agents were able to make improvement to the performance of nearly all of the agents. The speaker will also talk about a technique, Symbolic Noise Detection (SND), for detecting noise (i.e., mistakes or miscommunications) among agents in repeated games. The idea behind SND is that if we can build a model of the other agent's behavior, we can use this model to detect and correct actions that have been affected by noise. In the 20th Anniversary Iterated Prisoner's Dilemma competition, the SND agent placed third in the "noise" category, and was the best performer among programs that had no "slave" programs feeding points to them. About the speaker:Tsz-Chiu Au is a graduate student at Department of Computer Science, University of Maryland. (expected PhD in 2008). He received his B. Eng. degree from Hong Kong University of Science and Technology. His research interests lie in AI planning, multi-agent systems and problem solving by searching. He helped to develop SHOP2, a planning system based on hierarchical task network (HTN). His research accomplishments include his work on coping with noise in non zero-sum games, synthesis of strategies from interaction traces and managing volatile data for planning processes in semantic web service composition. |
Tuesday, June 24, 11:00am
|
From Experience to Reason: Dyna-Style Planning with Linear Function Approximation and Prioritized SweepingDr. Rich Sutton [homepage]
[Sign-up schedule for individual meetings]Understanding the world, representing its state and dynamics at multiple
levels of abstraction, and being able to use this knowledge flexibly to
achieve goals are key abilities sought in all approaches to artificial
intelligence. In this talk I approach them from the point of view of
reinforcement learning, which means an emphasis on learning and action, and
on how these interrelate with planning. In particular, I expand on the idea
that learning and planning can be done simultaneously and by the same
algorithm---operating either on real experience (learning) or imagined
experience (planning). This idea became popular in the 1990s, under the name
"Dyna architecture," in part because it was one of very few planning systems
that worked with a learned model of the world. However, a limitation of past
work with the Dyna architecture was that it used a table-lookup form for the
world model (in which every state was treated distinctly without
generalization) which does not scale to large problems. Scaling requires
replacing the tables with parameterized function approximators. However, we
now know that combining reinforcement learning with function approximation
can become unstable when trained "off-policy", i.e., with counterfactuals
such as are inherent in planning. Our main new result is to establish
conditions under which the stability of Dyna-style planning can be
proved. Given this, we can also immediately establish the soundness of
several natural generalizations of prioritized sweeping to linear function
approximation. Prioritized sweeping is a search control method for Dyna that
focuses planning effort where it has greatest effect, sometimes dramatically
increasing planning efficiency. The resulting system is probably the most
efficient online reinforcement-learning method known. I conclude by
discussing extensions of the Dyna idea to temporally abstract courses of
action (options) and "reasoning" -- planning about subgoals other than the
ultimate. About the speaker:Richard S. Sutton is a professor and iCORE chair in the department of computing science at the University of Alberta. He is a fellow of the American Association for Artificial Intelligence and co-author of the textbook Reinforcement Learning: An Introduction from MIT Press. Before joining the University of Alberta in 2003, he worked in industry at AT&T and GTE Labs, and in academia at the University of Massachusetts. He received a PhD in computer science from the University of Massachusetts in 1984 and a BA in psychology from Stanford University in 1978. Rich's research interests center on the learning problems facing a decision-maker interacting with its environment, which he sees as central to artificial intelligence. He is also interested in animal learning psychology, in connectionist networks, and generally in systems that continually improve their representations and models of the world. |
Monday, July 7, 2:00pm
|
Talking with Robots: A Case Study in Cognitive Architectures for RobotsDr. Jeremy Wyatt [homepage]
In what ways can we integrate multiple types of sensing and action in a robot? This question gets to the heart of deep issues in AI such as the nature and use of representations, and the control of the flow of information in a cognitive architecture. In this talk I will describe some work we are doing on architectures for cognitive robots. I will describe an architectural schema we have developed, a toolkit for developing robot systems using it, examples of robot systems that we have built, and some of the problems that arise if the schema is accepted. These include four research problems which we refer to as the problems of binding; filtering; processing management and action fusion. I will describe our current approaches to the first three, with a focus on using existing approaches POMDP planning to solving a simple processing management problem in vision. If I have time I will also say what I think is missing, what is wrong, and where it needs to go. About the speaker:Jeremy Wyatt (www.cs.bham.ac.uk/~jlw) is a senior lecturer in the School of Computer Science at the University of Birmingham. He is a co-director of the Intelligent Robotics Laboratory and a Leverhulme Trust Research Fellow. His research interests include: reinforcement learning, learning and planning in POMDPs, cognitive architectures for robots, robot learning, committee machines and ensemble learning, planning in underwater vehicles, and planning of visual processing. He has a PhD in Artificial Intelligence from the University of Edinburgh (1997), and supervised the 2004 winner of the British Computer Society's Distinguished Dissertation Award. Among other things he has worked on the CoSy project on cognitive systems, the CogX project on self-understanding and self-extension in cognitive systems, and a project on automated diagnosis for autonomous underwater vehicles. |
Friday, August 15, 11:00am
|
Probabilistic Reasoning about Large or Unknown WorldsDr. Brian Milch [homepage]
Intelligent systems must reason about the objects in their world, such as the vehicles on a road monitored by video cameras, or the people and events mentioned in a set of documents. In many cases, these objects are not known in advance; the system must infer their existence. Reasoning about such large and potentially unknown worlds calls for probabilistic techniques that generalize across objects rather than handling each one individually. I will begin this talk by introducing Bayesian logic, or BLOG, a language for defining distributions over possible worlds with varying sets of objects. Every well-formed BLOG model is guaranteed to fully define a distribution, even if it defines infinitely many random variables. I will then describe an approximate inference algorithm for BLOG that executes a Markov chain over "partial worlds", each of which instantiates only a relevant subset of the variables. This algorithm will be demonstrated on a citation-matching problem: identifying the distinct publications referred to by citation strings in online papers. I will also describe an exact inference algorithm that is applicable to models with large but known sets of objects. By exploiting two distinct forms of symmetry in such models, this algorithm can achieve exponential speed-ups even compared to previous "lifted" algorithms. About the speaker:Brian Milch recently completed a post-doc with Prof. Leslie Kaelbling in the Computer Science and Artificial Intelligence Lab at MIT; he will join the Search Quality group at Google in September. Joining Google will be a homecoming for Brian, since he also spent a year there after graduating from Stanford in 2000. He then entered the Ph.D. program at U.C. Berkeley, where he worked with Prof. Stuart Russell and received his Ph.D. in 2006. Brian is the recipient of an NSF Graduate Research Fellowship and a Siebel Scholarship, and was named one of the "Ten to Watch" in artificial intelligence by IEEE Intelligent Systems in 2008. |