Forum for Artificial Intelligence
Archive
|
This website is the archive for past Forum for Artificial Intelligence talks. Please click this link to navigate to the list of current talks. FAI 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 Catherine Andersson. |
[ Upcoming talks ]
Wednesday, August 31, 2011, 7:00PM
|
Driving and Developing AI Technology for the Mars Exploration Rover (MER) MissionTara Estlin [homepage]
The Mars Exploration Rover (MER) Mission Spirit and Opportunity rovers have demonstrated that mobile rovers are a valuable means of exploring the surface of other planets. Surface rovers offer scientists the ability to move around a planetary surface and explore different areas of interest. The MER rovers have traveled over many kilometers of terrain and survived harsh planetary conditions, including Martian winters and major dust storms, to continue collecting data. In this talk, Dr. Estlin will discuss her experiences as a rover driver for the MER Mission and describe some of the techniques used to move and gather science data on the Mars surface.
Advances in Mars rover mobility have also increased rover traverse range, and with it the opportunity for scientific discovery. Dr. Estlin will also introduce a new AI autonomy technology that was uploaded to the Opportunity rover in late 2009. The Autonomous Exploration for Gathering Increased Science (AEGIS) system enables automated data collection by rovers through the use of onboard data analysis techniques. AEGIS can automatically analyze images acquired by a rover and select science targets in those images based on pre-specified scientist criteria. If high-priority science targets are located, high-quality data can be autonomously acquired of the targets using the MER Panoramic Camera instrument. Dr. Estlin will provide brief overview of the AEGIS automated targeting capability and describe how it is currently being used onboard the MER mission Opportunity rover. About the speaker:Dr. Tara Estlin has over 10 years of experience in developing spacecraft autonomy software and is a senior member of the JPL Artificial Intelligence Group. Dr. Estlin is currently leading the AEGIS Project, which is providing new automated targeting technology for remote sensing instruments on the Mars Exploration Rover (MER) mission and was recently awarded the 2011 NASA Software of the Year award. For the past seven years, Dr. Estlin has also been a rover driver for the MER mission where she is responsible for sequencing drive and arm deployment commands for the MER Spirit and Opportunity rovers.Dr. Estlin is a senior member of the JPL Artificial Intelligence Group. She holds a B.S. in computer science from Tulane University and M.S. and Ph.D. degrees in computer science from the University of Texas at Austin. |
Thursday, September 15, 2011, 11:00AM
|
Evolution of Time in Neural Networks: Present to Past to FutureYoonsuck Choe [homepage]
What is time? Since the function of the brain is closely tied in with that
of time, investigating the origin of time in the brain can help shed light
on this question. In this paper, we propose to use simulated evolution of
artificial neural networks to investigate the relationship between time
and brain function, and the evolution of time in the brain. A large number
of neural network models are based on a feedforward topology (perceptrons,
backpropagation networks, radial basis functions, support vector machines,
etc.), thus lacking dynamics. In such networks, the order of input
presentation is meaningless since the behavior is largely reactive. That
is, such neural networks can only operate in the present, having no access
to the past or the future. However, biological neural networks are mostly
constructed with a recurrent topology, and recurrent (artificial) neural
network models are able to exhibit rich temporal dynamics, thus time
becomes an essential factor in their operation. In this talk, we will
investigate the emergence of recollection and prediction in evolving
neural networks. First, we will show how reactive, feedforward networks
can evolve a memory-like function (recollection) through utilizing
external markers dropped and detected in the environment. Second, we will
investigate how recurrent networks with more predictable internal state
trajectory can emerge as an eventual winner in evolutionary struggle when
competing networks with less predictable trajectory show the same level of
behavioral performance. We expect our results to help us better understand
the evolutionary origin of recollection and prediction in neuronal
networks, and better appreciate the role of time in brain function.
About the speaker:Yoonsuck Choe is an associate professor and director of the Brain Networks Laboratory in the Department of Computer Science and Engineering at Texas A&M University. He received his B.S. degree in Computer Science from Yonsei University (Korea) in 1993, and his M.S. and Ph.D. degrees in Computer Sciences from the University of Texas at Austin in 1995 and 2001. His research interest is broadly in modeling and understanding brain function, from the local circuit level up to the whole brain scale, with a focus on the temporal and sensorimotor aspects. |
Monday, September 19, 2011, 3:00PM
|
Yahoo! Data Mining Seminar: Scaling Machine Learning to the InternetAlex Smola [homepage]
In this talk I will give an overview over an array of highly scalable
techniques for both observed and latent variable models. This makes
them well suited for problems such as classification, recommendation
systems, topic modeling and user profiling. I will present algorithms
for batch and online distributed convex optimization to deal with
large amounts of data, and hashing to address the issue of parameter
storage for personalization and collaborative filtering. Furthermore,
to deal with latent variable models I will discuss distributed
sampling algorithms capable of dealing with tens of billions of latent
variables on a cluster of 1000 machines. The algorithms described are
used for personalization, spam filtering, recommendation, document
analysis, and advertising.
About the speaker:Alex Smola studied physics in Munich at the University of Technology, Munich, at the Universita degli Studi di Pavia and at AT&T Research in Holmdel. During this time he was at the Maximilianeum München and the Collegio Ghislieri in Pavia. In 1996 he received the Master degree at the University of Technology, Munich and in 1998 the Doctoral Degree in computer science at the University of Technology Berlin. Until 1999 he was a researcher at the IDA Group of the GMD Institute for Software Engineering and Computer Architecture in Berlin (now part of the Fraunhofer Geselschaft). After that, he worked as a Researcher and Group Leader at the Research School for Information Sciences and Engineering of the Australian National University. From 2004 onwards Alex worked as a Senior Principal Researcher and Program Leader at the Statistical Machine Learning Program at NICTA. He is currently working as Principal Research Scientist at Yahoo! Research. |
Monday, September 26, 2011, 12:00PM
|
Center for Perceptual Systems Seminar Series: Learning to MoveKaren E. Adolph [homepage]
A central issue for research in psychology, biology, robotics, and
computational modeling is how movements are generated and
controlled. I address this issue by asking how infants solve the
problem of moving. My research shows that basic motor skills such as
looking, reaching, and walking do not simply appear as the result of
maturation. Rather, motor skills are learned over months or years of
practice. Learning entails discovering new forms of movements, using
perceptual information to select movements appropriately and to modify
movements prospectively, and honing motor skills to make them more
fluent and efficient. How do babies do it? One clue is that infants
acquire immense amounts of variable, distributed practice with basic
motor skills. A second clue is that the learning process is geared
toward flexibility rather than rote performance: Infants are "learning
to learn" rather than acquiring fixed solutions. The process does not
end after infancy: Learning to learn is a life-long endeavor of
matching ongoing actions to the changing constraints of the body and
environment. |
Monday, October 3, 2011, 1:00PM
|
Vision: Sparsity via dense correspondences for videos and large-scale image databasesCe Liu [homepage]
We propose sparsity via dense correspondences as a general
principle for analyzing a number of images: an image can be well aligned to
its neighbors (both in videos and large-scale image databases), and
information can be propagated from neighbors for image synthesis and
analysis. We first present high-quality, adaptive video denoising and super
resolution systems based on reliable motion estimation. State-of-the-art
results were achieved through aggregating multiple frames via high-accuracy
optical flow fields. Correspondences between neighboring frames in a video
sequence may seem natural. How about arbitrary images? We show that
"neighboring frames" and "optical flow" can be generalized from videos to
large-scale image databases through dense scene alignment techniques such as
SIFT flow. Using SIFT flow, information is transferred and propagated in a
large-scale database for image parsing, recognition, segmentation, depth
estimation, and even image enhancement. We conclude that images are by
nature sparse via dense correspondences.
About the speaker:Ce Liu received the BS degree in automation and the ME degree in pattern recognition from the Department of Automation, Tsinghua University in 1999 and 2002, respectively. After receiving the PhD degree from the Massachusetts Institute of Technology in 2009, he now holds a researcher position at Microsoft Research New England. From 2002 to 2003, he worked at Microsoft Research Asia as an assistant researcher. His research interests include computer vision, computer graphics, and machine learning. He has published more than 30 papers in the top conferences and journals in these fields. He received the Outstanding Student Paper award at the Advances in Neural Information Processing Systems (NIPS) in 2006, and the Best Student Paper award at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2009. |
Thursday, October 6, 2011, 11:00AM
|
UTCS Distinguished Lecture Series - Planning, estimation, and execution in complex robot domainsLeslie Pack Kaelbling [homepage]
The fields of AI and robotics have made great improvements in many individual subfields, including in motion planning, symbolic planning, probabilistic reasoning, perception, and learning. Our goal is to develop an integrated approach to solving very large problems that are hopelessly intractable to solve optimally. We make a number of approximations during planning, including serializing subtasks, factoring distributions, and determinizing stochastic dynamics, but regain robustness and effectiveness through a continuous state-estimation and replanning process. This work is in early stages, but it has been demonstrated in simulation and on a real PR2 mobile manipulation problem.
About the speaker:Leslie Pack Kaelbling is Professor of Computer Science and Engineering and Research Director of the Computer Science and Artificial Intelligence Laboratory (CSAIL) at the Massachusetts Institute of Technology. She has previously held positions at Brown University, the Artificial Intelligence Center of SRI International, and at Teleos Research. She received an A.B. in Philosophy in 1983 and a Ph. D. in Computer Science in 1990, both from Stanford University. Prof. Kaelbling has done substantial research on designing situated agents, mobile robotics, reinforcement learning, and decision-theoretic planning. In 2000, she founded the Journal of Machine Learning Research, a high-quality journal that is both freely available electronically as well as published in archival form; she currently serves as editor-in-chief. Professor Kaelbling is an NSF Presidential Faculty Fellow, a former member of the AAAI Executive Council, the 1997 recipient of the IJCAI Computers and Thought Award, a trustee of IJCAII and a fellow of the AAAI. |
Friday, October 14, 2011, 11:00AM
|
In Search of Styles in Language: Identifying Deceptive Product Reviews, Wikipedia Vandalism, and the Gender of Authors via Statistical Stylometric AnalysisYejin Choi [homepage]
Language is a window into the mind. Stylometric analysis, the study of analyzing linguistic styles in language, can help uncovering the cognitive state and the personal identity of the writer. In this talk, I will present three case studies of Natural Language Processing (NLP) tasks that expand the scope of statistical stylometric analysis. First I will present the study of identifying deceptive product reviews, i.e., fake reviews that are written by people who are paid to fabricate positive reviews. As it turns out, it is surprisingly hard for human to distinguish fake reviews from truthful ones. Statistical analysis of language use on the other hand leads to nearly 90% accuracy, and provides us new clues in spotting suspicious reviews. Next I will introduce the study of detecting Wikipedia vandalism, where textual vandalism can be viewed as a unique genre in which a group of people with similar purpose share similar linguistic behavior. Finally, I will present the study of gender attribution, where we will examine whether there are gender-specific linguistic signals that go beyond the boundaries of topic and genre, and whether they are traceable even in modern and scientific literature.
About the speaker:Yejin Choi is an Assistant Professor in the Computer Science Department at Stony Brook University (SUNY Stony Brook). She received her Ph.D. in Computer Science from Cornell University in 2010 in the area of Natural Language Processing. Her research interests include stylometric analysis, natural language generation from images, and opinion & sentiment analysis in text. |
Friday, October 21, 2011, 11:00AM
|
Learning Compositional Semantics from Weak SupervisionPercy Liang [homepage]
What is the total population of the ten largest capitals in the US? Building a system to answer free-form questions such as this requires modeling the deep semantics of language. But to develop practical, scalable systems, we want to avoid the costly manual annotation of these deep semantic structures and instead learn from just surface-level supervision, e.g., question/answer pairs. To this end, we develop a new tree-based semantic representation which has favorable linguistic and computational properties, along with an algorithm that induces this hidden representation. Using our approach, we obtain significantly higher accuracy on the task of question answering compared to existing state-of-the-art methods, despite using less supervision.
About the speaker:Percy Liang is currently a post-doc at Google and will be starting as an assistant professor at Stanford next fall. He obtained his B.S./M.S. from MIT and Ph.D. from UC Berkeley. The general theme of his research, which spans machine learning and natural language processing, is learning richly-structured statistical models from limited supervision, most recently in the context of program induction and natural language semantics. He has won a best student paper at the International Conference on Machine Learning in 2008, received the NSF, GAANN, and NDSEG fellowships, and is also a 2010 Siebel Scholar. |
Friday, October 28, 2011, 11:00AM
|
Learning Dynamic Models with Non-sequenced DataJeff Schneider [homepage]
Virtually all methods of learning dynamic systems from data start with the
same basic assumption: the learning algorithm will be given a time sequence
of data generated from the dynamic system. We consider the case where the
training data comes from the system’s operation but with no temporal
ordering. The data are simply drawn as individual disconnected points.
While making this assumption may seem absurd at first glance, we observe
that many scientific modeling tasks have exactly this property.
We propose several methods for solving this problem. We write down an approximate likelihood function that may be optimized to learn dynamic models and show how kernel methods can be used to obtain non-linear models. We propose an alternative method that focuses on achieving temporal smoothness in the learned dynamics. Finally, we consider the case where a small amount of sequenced data is available along with a large amount of non-sequenced data. We propose the use of the Lyapunov equation and the non-sequenced data to provide regularization when performing regression on the sequenced data to learn a dynamic model. We demonstrate our methods on synthetic data and describe the results of our analysis of some bioinformatics data sets. About the speaker:Dr. Jeff Schneider is an associate research professor in the Carnegie Mellon University School of Computer Science. He received his PhD in Computer Science from the University of Rochester in 1995. He has over 15 years experience developing, publishing, and applying machine learning algorithms in government, science, and industry. He has dozens of publications and has given numerous invited talks and tutorials on the subject.In 1995 Dr. Schneider co-founded and became CEO of Schenley Park Research, a company dedicated to bringing new machine learning algorithms to industry. In 2004 he developed a new machine-learning based CNS drug discovery system and spent two years as the CIO of Psychogenics to commercialize the system. Through his academic, commercial, and consulting efforts, he has worked with several dozen companies and government agencies around the world. |
Friday, November 11, 2011, 11:00AM
|
How to learn from one sample? Statistical relational learning for single network domainsJennifer Neville [homepage]
Machine learning researchers focus on two distinct learning scenarios for structured data that can be represented as a graph (i.e., where there are dependencies among the class labels and attributes of linked nodes). In one scenario, the domain consists of a population of structured examples (e.g., chemical compounds). In this case, since the population is comprised of a set of independent graphs, we can reason about models and algorithms as the number of structured examples increases. In the other scenario, the domain consists of a single, potentially infinite-sized network (e.g., the Facebook friendship network). In this case, an increase in dataset size corresponds to acquiring a larger portion of the underlying network. In single network domains, even when there are multiple networks samples available for learning, they correspond to subnetworks drawn from the same underlying network and thus may be dependent.In our recent work, we have focused on the development and analysis of statistical relational learning (SRL) methods for single network domains, particularly social networks. Although SRL algorithms have been successfully applied for social network classification, the algorithmic foundations of SRL methods are based on an implicit assumption of an underlying population of networks---which does not hold for most social network datasets. In this talk, I will present our recent efforts to outline a more formal foundation for single network learning and discuss how the analysis has informed the development of more accurate estimation and evaluation methods. About the speaker:Jennifer Neville is an assistant professor at Purdue University with a joint appointment in the Departments of Computer Science and Statistics. She received her PhD from the University of Massachusetts Amherst in 2006. She received a DARPA IPTO Young Investigator Award in 2003 and was selected as a member of the DARPA Computer Science Study Group in 2007. In 2008, she was chosen by IEEE as one of "AI's 10 to watch." Her research focuses on developing data mining and machine learning techniques for relational domains, including citation analysis, fraud detection, and social network analysis. |
Friday, November 18, 2011, 11:00AM
|
Robotic Decision Making for Sensing in the Natural WorldGeoffrey A. Hollinger [homepage]
There is growing interest in the use of robots to gather information from natural environments. Examples include biological monitoring, mine sweeping, oil spill cleanup, and seismic event detection. The increasing capabilities of the robots themselves enable more sophisticated decision making techniques that optimize information gathered and adapt as new information is received. The question
becomes: how do we develop path planning algorithms for information gathering tasks that are capable of dealing with the communication limitations, noisy sensing, and mobility restrictions present in natural environments?
This talk considers two problems related to path planning for Autonomous Underwater Vehicles (AUVs): (1) data gathering from an underwater sensor network equipped with acoustic communication and (2) autonomous inspection of the submerged portion of a ship hull. For the first problem, I present path planning methods that extend algorithms for variants of the Traveling Salesperson Problem (TSP) and show how these algorithms can be integrated with realistic acoustic communication models. For the second problem, I discuss techniques for constructing watertight 3D meshes from sonar-derived point clouds and introduce uncertainty modeling through non-parametric Bayesian regression. Uncertainty modeling provides novel cost functions for planning the path of the robot that allow for formal analysis through connections to submodular optimization and active learning. Such theoretical analysis provides insight into the underlying structure of active sensing problems. Finally, I present experiments that demonstrate the high performance of the proposed solutions versus the state of the art in robot path planning.
About the speaker:Geoffrey A. Hollinger is a Postdoctoral Research Associate in the Robotic Embedded Systems Laboratory and Viterbi School of Engineering at the University of Southern California. He is currently interested in adaptive sensing and distributed coordination for robots operating with limited communication. He has also worked on multi-robot search at Carnegie Mellon University, personal robotics at Intel Research Pittsburgh, active estimation at the University of Pennsylvania's GRASP Laboratory, and miniature inspection robots for the Space Shuttle at NASA's Marshall Space Flight Center. He received his Ph.D. (2010) and M.S. (2007) in Robotics from Carnegie Mellon University and his B.S. in General Engineering along with his B.A. in Philosophy from Swarthmore College (2005). |
Monday, November 28, 2011, 11:00AM
|
Algorithms for L1 Regularized Reinforcement LearningRon Parr [homepage]
Regularization is an important tool in most areas of machine learning, yet its role and importance in reinforcement learning has only recently started to come into focus. Efficient algorithms for regularized reinforcement learning are an important step towards easing the feature selection challenge, one of the most significant practical obstacles to wider deployment of reinforcement learning algorithms. I will present two L1 regularized reinforcement learning algorithms. The first algorithm uses an approach based upon linear complementarity to solve for the L1 regularized linear fixed point problem introduced by Kolter and Ng in their LARS-TD work. The linear complementarity perspective permits more efficient and robust solution methods, as well as stronger theoretical guarantees. The second algorithm is based upon approximate linear programming (ALP), where regularization leads to a novel way of providing solution quality guarantees in the typical case where the approximate linear program has fewer constraints than states in the problem. Our experimental results demonstrate that these algorithms can efficiently find good solutions to reinforcement learning problems by automatically selecting a small set of features from hundreds or even thousands of automatically generated features. Joint work with Jeff Johns (Duke), Christopher Painter-Wakefield (Duke), Marek Petrik (IBM), Gavin Taylor (United States Naval Academy), and Shlomo Zilberstein (U Mass.) About the speaker:Ron Parr is an associate professor at the Duke University Department of Computer Science. He received his A.B. (Cum Laude) in Philosophy in 1990 from Princeton University, where he was advised by Gilbert Harman. In 1998, he received his Ph.D. in computer science from the University of California at Berkeley, under the supervision of Stuart Russell. After graduating from Berkeley, Ron spent two years as a postdoctoral research associate at Stanford University, where he worked with Daphne Koller. He was selected as a Sloan fellow in 2003. In 2006, he received the NSF CAREER ward and served on DARPA's computer science study group (CSSG). He was program co-chair for the 2007 UAI conference, and general chair of the 2008 UAI conference. Ron has served on the editorial boards of JAIR and MLJ, and is an action editor for JMLR. |
Friday, December 2, 2011, 11:00AM
|
Reading Minds?: Predicting Mental Function from Neuroimaging DataRussell Poldrack [homepage]
Neuroimaging tools have become increasingly powerful, to the degree
that some investigators have claimed to be able to "read minds" using
functional MRI. I will discuss our work which has examined the degree
to which mental functions can be predicted from neuroimaging data,
focusing particularly on generalization to new individuals. I will
show that we can accurately classify which of a large set of mental
functions a person is engaged in, using classifiers trained on other
individuals. I will also show that we can predict individual
variability in behavior using high-dimensional regression. Finally, I
will demonstrate how we can classify individuals using meta-analytic
data derived from literature mining.
About the speaker:Russ Poldrack received his Ph.D in Cognitive Psychology from the University of Illinois at Urbana-Champaign. He was a postdoctoral fellow at Stanford, and held faculty positions at Harvard Medical School and UCLA before becoming Director of the Imaging Research Center at the University of Texas at Austin. His research uses neuroimaging to understand the neural basis of decision making, executive function, and learning. He has also written extensively on conceptual and analytic issues regarding fMRI. In addition, he is deeply involved in the development of informatics resources for cognitive neuroscience, including the Cognitive Atlas project and the OpenFMRI project. His research has received awards from the Organization for Human Brain Mapping (OHBM) and American Psychological Association, and in 2009 he served as Chair of the OHBM. He is as Associate Editor for Frontiers in Human Neuroscience, and has served on the editorial boards for Trends in Cognitive Sciences, Cerebral Cortex, Human Brain Mapping, Cognitive Science, and Neuroimage. |
Friday, December 2, 2011, 3:00PM
|
Statistical Translation as Constrained OptimizationChris Dyer [homepage]
I discuss translation as an optimization problem subject to three
kinds of constraints: lexical, relational, and constraints enforcing
target-language wellformedness. Lexical constraints ensure that the
lexical choices in the output are meaning-preserving; relational
constraints ensure that the relationships between words and phrases
(e.g., semantic roles and modifier-head relationships) are properly
transformed; and target-language wellformedness constraints ensure the
grammaticality of the output. In terms of the traditional
source-channel model of Brown et al. (1993), the "translation model"
encodes lexical and relational constraints and the "language model"
encodes target language wellformedness constraints. This
constraint-based framework suggests a discriminative
(generate-and-test) model of translation in which constraints are
encoded as features sensitive to input and output elements, and the
feature weights are trained to maximize the (conditional) likelihood
of the parallel data.
To verify the usefulness of the constraint-based approach, I discuss the performance of two models: first, a lexical translation model evaluated by the word alignments it learns. Unlike previous unsupervised alignment models, the new model utilizes features that capture diverse lexical and alignment relationships, including morphological relatedness, orthographic similarity, and conventional co-occurrence statistics. Results from typologically diverse language pairs demonstrate that the feature-rich model provides substantial performance benefits compared to state-of-the-art generative models. Second, I discuss the results of an end-to-end translation system in which lexical, relational, and wellformedness constraints modeled independently. Because of the independence assumptions, the model is substantially more compact than state-of-the-art translation models, but still performs significantly better on languages where source-target word order differences are substantial. About the speaker:Chris Dyer is a postdoctoral researcher in Noah Smith's lab in the Language Technologies Institute at Carnegie Mellon University. He completed his PhD on statistical machine translation with Philip Resnik at the University of Maryland in 2010. Together with Jimmy Lin, he is author of "Data-Intensive Text Processing with MapReduce", published by Morgan & Claypool in 2010. Current research interests include machine translation, unsupervised learning, Bayesian techniques, and "big data" problems in NLP. |
Thursday, January 19, 2012, 11:00AM
|
Dr. Fill: Crosswords Aren't Just for Humans Any MoreMatthew L. Ginsberg [homepage]
We describe Dr.Fill, a program that solves American-style crossword
puzzles. From a technical perspective, Dr.Fill works by converting
crosswords to weighted CSPs, and then using a variety of novel
techniques to find a solution. These techniques include generally
applicable heuristics for variable and value selection, a variant of
limited discrepancy search, and postprocessing and partitioning
ideas. Branch and bound is not used, as it was incompatible with
postprocessing and was determined experimentally to be of little
practical value. Dr.Fill's performance on crosswords from the American
Crossword Puzzle Tournament suggests that it ranks among the top fifty
or so crossword solvers in the world.
About the speaker:Matthew L. Ginsberg received his doctorate in mathematics from Oxford in 1980 at the age of 24. He remained on the faculty in Oxford until 1983, doing research in mathematical physics and computer science; during this period, he wrote a program that was used successfully to trade stock and stock options on Wall Street. Ginsberg's continuing interest in artificial intelligence brought him to Stanford in late 1983, where he remained for nine years. He then went on to found CIRL, the computational intelligence research laboratory at the University of Oregon, which he directed until 1996. He remained at CIRL until 1998, when CIRL spun off On Time Systems, a commercial entity focusing on scheduling and routing technology. Ginsberg has been the CEO of the company since its formation and is currently its chairman as well. Ginsberg is also the chairman and CEO of Green Driver, Inc., a sister company to On Time Systems that focuses on using real-time traffic and signal information to provide more fuel-efficient routes to drivers. Ginsberg's present research interests focus on constraint satisfaction. He is the author of numerous publications in this areas, the editor of "Readings in Nonmonotonic Reasoning," and the author of "Essentials of Artificial Intelligence," both published by Morgan Kaufmann. He is also the author of the bridge-playing program GIB, which made international news by finishing 12th in the world bridge championships in Lille, France, and the author of Dr. Fill, a crossword-solving program that will be participating in the American Crossword Puzzle Tournament in March of 2012." |
Friday, January 20, 2012, 11:00AM
|
UT Austin Villa 2011: RoboCup 3D Simulation League ChampionPatrick MacAlpine [homepage]
The RoboCup 3D simulation league is an international competition in
which autonomous simulated humanoid robots play soccer against each
other in a physically realistic environment. This talk will present
the key components of the UT Austin Villa 2011 RoboCup 3D simulation
league team. These key components include an omnidirectional walk
engine and associated walk parameter optimization framework, an
inverse kinematics based kicking architecture, and a dynamic role and
formation positioning system. UT Austin Villa won the RoboCup 2011 3D
simulation competition, consisting of 22 teams from 12 different
countries, in convincing fashion by winning all 24 games it
played. During the course of the competition the team scored 136 goals
while conceding none.
About the speaker:Patrick MacAlpine is a third year Computer Science PhD student at the University of Texas at Austin where he is researching autonomous multiagent systems and machine learning. His current focus is on using reinforcement learning to develop locomotion skills and strategy for the UT Austin Villa RoboCup 3D Simulation League team. Additionally he is exploring learning methods for recurrent neural networks. Before coming to UT, Patrick worked as a software engineer at Green Hills Software and Acelot, Inc. in Santa Barbara, California. Prior to that he received a BS and MEE degree in Electrical Engineering from Rice University. |
Friday, February 10, 2012, 11:00AM
|
Efficient Sampling-based Inference in presence of Logical StructureVibhav Gogate [homepage]
The emerging field of statistical relational learning (SRL) seeks to marry logical and probabilistic representation and reasoning techniques. A good marriage is essential because many real world application domains have both rich relational (i.e., logical) structure and large amount of uncertainty. Although, great progress has been made in solving the representational issues in SRL, progress in inference has been lacking. In this talk, I'll describe our ongoing attempt at achieving this much needed progress. Specifically, I'll focus on a class of simulation-based approximate inference technique called importance sampling and show how to substantially improve its speed, scalability and accuracy by exploiting logical structure. I'll show that our new sampling algorithm is a special case of the standard DPLL algorithm for Satisfiability (SAT) testing and theorem proving. This enables us to leverage efficient and highly scalable algorithms and software from SAT and theorem proving communities for efficient inference in SRL models. I'll conclude by presenting results from the UAI 2010 approximate inference challenge where our importance sampling method won first place in many categories. (Joint work with Rina Dechter and Pedro Domingos)
About the speaker:Vibhav Gogate is an assistant professor in the computer science department at the University of Texas at Dallas. He got his Ph.D. from University of California, Irvine in 2009 and then did a two-year postdoc at University of Washington. His research interests are in machine learning and artificial intelligence with a focus on probabilistic graphical models and statistical relational learning. He has authored over 20 papers in top-tier conferences and journals such as UAI, NIPS, AAAI, AISTATS and the AI journal. He is a co-winner of the 2010 Uncertainty in Artificial Intelligence (UAI) approximate inference challenge. |
Friday, February 17, 2012, 11:00AM
|
Computer simulation of the cerebelllumMichael Mauk [homepage]
I will describe our ongoing efforts to understand the cerebellum well
enough to replicate its function with a bottom-up,
biologically-constrained computer simulation. I'll describe the
factors that make the cerebellum especially amenable to this approach.
These factors enable an especially tight integration of simulation and
experiment. I'll use select experiments to illustrate what the
cerebellum computes and some of what we know about how it computes.
About the speaker:Mike Mauk is a Professor in the Center for Learning and Memory and the Section of Neurobiology at the University of Texas at Austin. He received a BS in psychology in 1979 and a PhD in 1985 from Stanford University based on work identifying brain pathways involved in learning. After postdoctoral training at Stanford Medical School, he joined the faculty of the UT-Houston Medical School in 1988. At UT-Houston Mike was promoted to associate professor in 1996, to full professor in 2001 and in 2004 was named the William M. Wheless III Professor of Biomedical Sciences. He came to UT-Austin in 2007. Mike is internationally known for his work on the brain mechanisms of learning, particularly for pioneering use of computer simulations to study the brain. |
Tuesday, February 21, 2012, 2:00PM
|
Reusable Teamwork for Multi-Agent, Multi-Robot TeamsGal Kaminka [homepage]
For many years, multi-robot researchers have focused on specific application-inspired basic tasks (e.g., coverage, moving in formation, foraging, patrolling) as a way of studying cooperation between robots. But users want to see increasingly complex missions being tackled, which challenge this methodology: first, some missions cannot be easily decomposed into the familiar basic tasks, making previous knowledge non-reusable; second, the target operating environments challenge the typically sterile settings assumed in many previous works (such challenges include adversaries, 100s-1000s of robots and software agents, multiple concurrent goals, human operators and users, and more).
In this talk, I will argue that the reusable components in complex missions are often found not in the tasks, but in the interactions between robots, i.e., that while taskwork varies significantly, teamwork is largely generic. And while many multi-robot researchers have begun exploring generic task-allocation methods, I will report on my group's work over the last decade, identifying and developing general mechanisms for teamwork, and integrating them at the architecture level to facilitate development of robust teams at reduced programming effort. I will sample some of our results in developing robots for missions ranging from robust formation maintenance, through patrolling, to soccer and urban search-and-rescue. About the speaker:Gal A. Kaminka is an associate professor at the computer science department, and the brain sciences research center, at Bar Ilan University (Israel). His research expertise includes multi-agent and multi-robot systems, teamwork and coordination, behavior and plan recognition, and modeling social behavior. He has received his PhD from the University of Southern California (2000), and spent two years as a post-doctorate fellow at Carnegie Mellon University. Today, Prof. Kaminka leads the MAVERICK research group at Bar Ilan, supervising over a dozen MSc and PhD students. He was awarded an IBM faculty award and top places at international robotics competitions. He served as the program chair of the 2008 Israeli Conference on Robotics, and the program co-chair of the 2010 Int'l Joint Conference on Autonomous Agents and Multi-Agent Systems (AAMAS). He has served on the international executive bodies of IFAAMAS (International Foundation of Autonomous Agents and Multi-Agent Systems) and AAAI (Association for Advancement of Artificial Intelligence). Currently, he is spending his sabbatical as Radcliffe Fellow, at Harvard University's Radcliffe Institute for Advanced Study. |
Thursday, February 23, 2012, 11:00AM
|
Knuth-Chen Meets Heuristic SearchRobert Holte [homepage]
In 1975 Knuth showed that a simple random-walk procedure gave an
unbiased estimate of the size of a depth-first backtrack search
tree. His student P-C Chen later improved Knuth's method by
partitioning the nodes in the search tree into types and broadening
the random walk so that at each level of search, for every type
reached at that level one node of that type would be expanded. In
this talk I describe our efforts to extend Chen's method to work in
the context of heuristic search. First, we consider the problem of
predicting the number of nodes expanded by IDA* for a given cost
bound. Chen's method is almost immediately applicable to this
problem, but it is outperformed by CDP, the current state-of-the-art
system for this prediction problem. However, when Chen's method is
used in conjunction with CDP's type system, new state-of-the-art
results are produced. Second, we consider the problem of predicting
the optimal solution cost of a search problem without actually finding
a solution. For this we devised a bidirectional variation of Chen's
method which we call BiSS. We show that BiSS makes more accurate
predictions than SCP, the current state-of-the-art, and scales to
larger state spaces than SCP can handle. If time permits I will also
discuss the application of BiSS to the problem of learning search
heuristics, and show that it reduces learning time for large state
spaces from days to minutes.
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. |
Friday, March 2, 2012, 11:00AM
|
The FO(.) Knowledge Base System project: an integration projectMarc Denecker [homepage]
The term FO(.) is used here as a generic term to denote extensions of classical first order logic FO. On the logical level, the goal of this project is to achieve a conceptual clean -non-hybrid- integration of logic constructs from different computational logics (logic programming extensions such as datalog, Abductive Logic Programming, Answer Set Programming, fixpoints logics, constraint programming) in the context of classical logic. On the computational level, the long term goal is to integrate (and extend) technologies developed in the respective fields to build a Knowledge Base System that supports various forms of inference. I will explain some motivations, principles and research questions raised by such a project. I will give an overview of the current system and some applications. One application for interactive configuration will serve to highlight a principle that separates declarative modelling languages from (procedural or declarative) programming languages: the reuse of a modelling to solve different computational tasks by applying different forms of inference.
About the speaker:Prof. Marc Denecker studied at the Katholic University Leuven (KUL) in Belgium, where he also did his PhD and has worked till now, with exception of a two year period at the University Libre de Bruxelles (ULB). His current interests range from theoretical topics such as foundations of knowledge representation, nonmonotonic reasoning, logic programming, classical logic, fixpoint and modal logics to building inference systems for integrations of these logics and the development of applications. |
Friday, March 23, 2012, 11:00AM
|
What our most forgettable words say about usJames W. Pennebaker [homepage]
Wait! Don't toss out those pronouns, prepositions, articles, and
auxiliary verbs. These seemingly insignificant words say the more
about people than any other word categories you usually analyze. In
this talk, a large number of studies will be described that point to
the links between function words and social and psychological states,
including personality, honesty, status, gender, social relationships,
and emotional state. Implications for psychology, computational
linguistics, business, and others parts of academia will be addressed.
About the speaker:James W. Pennebaker is the Regents Centennial Professor of Liberal Arts and the Departmental Chair in the Psychology Department at the University of Texas at Austin, where he received his Ph.D. in 1977. He has been on the faculty at the University of Virginia, Southern Methodist University, and, since 1997, The University of Texas. He and his students are exploring the links between traumatic experiences, expressive writing, natural language use, and physical and mental health. His studies find that physical health and work performance can improve by simple writing and/or talking exercises. His most recent research focuses on the nature of language and emotion in the real world. The words people use serve as powerful reflections of their personality and social worlds. Author or editor of 9 books and over 250 articles, Pennebaker has received numerous awards and honors. |
Friday, March 30, 2012, 11:00AM
|
Parsing Combinatory Categorial Grammar via Planning in Answer Set ProgrammingYulia Lierler [homepage]
The task of parsing -- recovering the internal structure of sentences
-- is an important task in natural language processing. Combinatory
categorial grammar (CCG) is one of the grammar formalisms that are
used for natural language parsing. We propose and implement a new
approach to CCG parsing that relies on a prominent knowledge
representation formalism, answer set programming (ASP) -- a
declarative programming paradigm. We formulate the task of CCG
parsing as a planning problem, cast it as an answer set program and
use an ASP computational tool to compute solutions that correspond to
valid parses. As a result we obtain what is the only CCG-based
wide-coverage parser capable of producing multiple parses for a
sentence. In addition, compared to other approaches, there is no need
to implement a specific parsing algorithm when this declarative
method is used. Interestingly, this project "marries" several
fundamental AI areas: natural language processing, planning, and
knowledge representation.
About the speaker:Yuliya Lierler is a Computing Innovation Fellow Postdoc at the Computer Science Department at the University of Kentucky. She completed her Ph.D. in Computer Science at the University of Texas at Austin in 2010. She is interested in knowledge representation, automated reasoning, declarative problem solving, and natural language understanding. |
Monday, April 2, 2012, 11:00AM
|
Better Business School Course AllocationAbraham Othman [homepage]
In the combinatorial allocation problem, agents have combinatorial preferences over bundles of possible allocations. The example I will discuss in detail is business school course allocation, where students may have complex preferences over schedules of courses (e.g., two desirable courses could meet at the same time). Other examples of combinatorial allocation include scheduling workers to shifts and scheduling airlines to landing slots. I will begin by giving the intuition why conventional approaches to this problem fail. I will then introduce a mechanism that implements a fair allocation while being approximately truthful and efficient. Unfortunately, implementing the mechanism requires minimizing a highly discontinuous function in ~100-dimensional space. I will detail the two-stage search technique we used to solve this difficult problem. At the master level, the center uses tabu search over the union of two distinct neighborhoods to suggest prices to the agents; at the agent level, we use MIPs to solve for student demands in parallel at the current prices. Our method scales near-optimally in the number of processors used and is able to solve realistic-size problems fast enough to be used in practice. This approach recently won a competition at the Wharton School to serve as their new course selection mechanism starting in Spring 2013. The practical success of the mechanism is exciting for two reasons beyond business schools. First, it offers a wide-ranging solution for repugnant market design with complex preferences. Second, it suggests a highly parallelizable approach to solving large-scale optimization problems. Joint work with Eric Budish (Chicago Booth). About the speaker:Bio TBA |
Friday, April 6, 2012, 11:00AM
|
Representing and Inferring the 3D Layout of RoomsDerek Hoiem [homepage]
We humans spend much of our time working, playing, and sleeping in rooms and can operate effectively in them. But computers have difficulty interpreting rooms because many important surfaces are hidden and nearby objects exhibit confusing perspective effects. For computers to interpret, navigate, or interact in rooms, they need better representations of space.
I will describe our efforts to represent and infer 3D layout of rooms. I will show how we can use structural priors to detect hidden boundaries and to create perspective-robust models of object appearance. I'll present results for inferring the 3D layout of walls and furniture from one image. I'll also present an application to inserting 3D objects into an image and relighting them (including cool videos). Finally, I'll touch on our most recent work to organize an RGB-depth image into surfaces and objects with physical support relations and will discuss the most pressing future directions.
About the speaker:Derek Hoiem is an assistant professor in Computer Science at the University of Illinois at Urbana-Champaign (UIUC). Before joining the UIUC faculty in 2009, Derek completed his Ph.D. in Robotics at Carnegie Mellon University in 2007 and was a postdoctoral fellow at the Beckman Institute from 2007-2008. Derek’s research interests focus on 3D scene interpretation and object recognition. His work has been recognized with a CVPR 2006 Best Paper award, a 2008 ACM Doctoral Dissertation Award honorable mention, and a 2011 NSF CAREER award. |
Friday, April 13, 2012, 11:00AM
|
Designing Learning Interactions for RobotsAndrea Thomaz [homepage]
In this talk I present recent work from the Socially Intelligent
Machines Lab at Georgia Tech. One of the focuses of our lab is on
Socially Guided Machine Learning, building robot systems that can
learn from everyday human teachers. We look at standard Machine
Learning interactions and redesign interfaces and algorithms to
support the collection of learning input from naive humans. This talk
covers results on high-level task goal learning, low-level skill
learning, and active learning interactions using several humanoid
robot platforms.
About the speaker:Andrea L. Thomaz is an Assistant Professor of Interactive Computing at the Georgia Institute of Technology. She directs the Socially Intelligent Machines lab, which is affiliated with the Robotics and Intelligent Machines (RIM) Center and with the Graphics Visualization and Usability (GVU) Center. She earned a B.S. in Electrical and Computer Engineering from the University of Texas at Austin in 1999, and Sc.M. and Ph.D. degrees from MIT in 2002 and 2006. Dr. Thomaz is published in the areas of Artificial Intelligence, Robotics, Human-Robot Interaction, and Human-Computer Interaction. She received an ONR Young Investigator Award in 2008, and an NSF CAREER award in 2010. Her work has been featured on the front page of the New York Times, and in 2009 she was named one of MIT Technology Review’s Top 35 under 35. |
Thursday, April 19, 2012, 1:00PM
|
Going Beyond NP: New Challenges in Inference TechnologyBart Selman [homepage]
In recent years, we have seen tremendous progress in inference
technologies. For example, in the area of Boolean satisfiability (SAT)
and Mixed Integer Programming (MIP) solvers now enable us to tackle
significant practical problem instances from applications in hardware
and software verification, planning, and scheduling. Key to this
success is the ability to strike the right balance between the
expressiveness of the underlying representation formalism and the
efficiency of the solvers. The next challenge is to extend the reach
of these solvers to more complex tasks that lie beyond NP.
About the speaker:Bart Selman is a professor of computer science at Cornell University. His research interests include efficient reasoning procedures, planning, knowledge representation, and connections between computer science and statistical physics. He has (co-)authored over 150 papers, which have appeared in venues spanning Nature, Science, Proc. Natl. Acad. of Sci., and a variety of conferences and journals in AI and Computer Science. He has received six Best Paper Awards, and is an Alfred P. Sloan Research Fellowship recipient, a Fellow of AAAI, and a Fellow of AAAS. |
Friday, April 20, 2012, 11:00AM
|
Cognitive Cooperative RobotsDaniele Nardi [homepage]
The talk aims at providing an overview of the recent research
achievements by our group at Sapienza University, Our work is
motivated by RoboCup related research problems, as well as by
applications in real world domains, such as service robotics and
disaster response robotics. Our aim is to address the challenge of
making robots intelligent, and our approach tries to combine in
various forms a symbolic representation of the robot's knowledge and
specialized techniques for processing sensor data and controlling the
motion of the robot. In particular, I will present our context-based
architectures, an action representation based on Petri Nets and our
recent work on Human Robot Interaction aiming at improving the
knowledge acquisition capabilities of the robot. |
Friday, May 4, 2012, 11:00AM
|
Combining data and networks to unravel the genetics of complex traitsAlexis Battle [homepage]
Complex traits, including many human diseases, are affected by multiple genetic elements, often working together in intricate networks and pathways. Recent technologies have allowed us to collect detailed genetic profiles on a large scale, opening up the possibility of identifying the genes and pathways disrupted in many diseases. However, untangling genetic factors from such data has presented significant statistical challenges. In particular, available studies are often underpowered, with up to millions of candidate genetic elements, but only hundreds or thousands of individuals. In this talk, I will discuss machine learning methods for inferring the effects of genetic variation on complex traits, based on structured probabilistic models. We model the effects of multiple genetic elements jointly on a given trait, and leverage known relationships among genes, incorporating gene networks as structured prior distributions. Thus, our models will preferentially identify multiple genetic elements which are known to be functionally connected, reflecting a more biologically plausible model of many complex traits, and providing improved statistical power. I will discuss the use of these methods in identifying genetic variants relevant to autoimmune diseases, and in identifying important interactions between genes in cancer survival. About the speaker:Alexis Battle is a PhD candidate in Computer Science at Stanford University. Her research in computational biology focuses on machine learning and probabilistic models for the genetics of complex traits. Alexis received her BS in Symbolic Systems from Stanford, and spent four years as a member of the technical staff at Google. She is the recipient of an NSF Graduate Research Fellowship and an NIH NIGMS training award. |