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 ]
Friday, September 19, 2014, 11:00AM
|
Behavior-Grounded Multisensory Object Perception and Exploration by a Humanoid RobotJivko Sinapov [homepage]
Infants use exploratory behaviors to learn about the objects around
them. Psychologists have theorized that behaviors such as touching,
pressing, lifting, and dropping enable infants to form grounded object
representations. For example, scratching an object can provide
information about its roughness, while lifting it can provide
information about its weight. In a sense, the exploratory behavior
acts as a ``question'' to the object, which is subsequently
``answered" by the sensory stimuli produced during the execution of
the behavior. In contrast, most object representations used by robots
today rely solely on computer vision or laser scan data, gathered
through passive observation. Such disembodied approaches to robotic
perception may be useful for recognizing an object using a 3D model
database, but nevertheless, will fail to infer object properties that
cannot be detected using vision alone. To bridge this gap, my research has pursued a framework for object perception and exploration in which the robot's representation of objects is grounded in its own sensorimotor experience with them. In this framework, an object is represented by sensorimotor contingencies that span a diverse set of exploratory behaviors and sensory modalities. Results from several large-scale experimental studies show that the behavior-grounded object representation enables a robot to solve a wide variety of tasks including recognition of objects based on the stimuli that they produce, object grouping and sorting, and learning category labels that describe objects and their properties. About the speaker:Jivko Sinapov received the Ph.D. degree in computer science and human-computer interaction from Iowa State University (ISU) in 2013. While working towards his Ph.D. at ISU’s Developmental Robotics Lab, he developed novel methods for behavioral object exploration and multi-modal perception. He is currently a Postdoctoral Fellow working with Peter Stone at the Artificial Intelligence lab. His research interests include developmental robotics, computational perception, autonomous manipulation, and human-robot interaction. |
Friday, October 3, 2014, 11:00AM
|
Robots, Skills, and SymbolsGeorge Konidaris [homepage]
Robots are increasingly becoming a part of our daily lives, from the automated vacuum cleaners in our homes to the rovers exploring Mars. However, while recent years have seen dramatic progress in the development of affordable, general-purpose robot hardware, the capabilities of that hardware far exceed our ability to write software to adequately control. The key challenge here is one of abstraction. Generally capable behavior requires high-level reasoning and planning, but perception and actuation must ultimately be performed using noisy, high-bandwidth, low-level sensors and effectors. My research uses methods from hierarchical reinforcement learning as a basis for constructing robot control hierarchies through the use of learned motor controllers, or skills. The first part of my talk will present work on autonomous robot skill acquisition. I will demonstrate a robot system that learns to complete a task, and then extracts components of its solution as reusable skills, which it deploys to quickly solve a second task. The second part will briefly focus on practical methods for acquiring skill control policies, through the use human demonstration and active learning. Finally, I will present my recent work on establishing a link between the skills available to a robot and the abstract representations it should use to plan with them. I will discuss the implications of these results for building true action hierarchies for reinforcement learning problems. About the speaker:George Konidaris is an Assistant Professor of Computer Science and Electrical and Computer Engineering at Duke University. He holds a BScHons from the University of the Witwatersrand, an MSc from the University of Edinburgh, and a PhD from the University of Massachusetts Amherst, having completed his thesis under the supervision of Professor Andy Barto. Prior to joining Duke, he was a postdoctoral researcher at MIT with Professors Leslie Kaelbling and Tomas Lozano-Perez. |
Friday, November 7, 2014, 11:00AM
|
Automating evidence synthesis via machine learning and natural language processingByron Wallace [homepage]
Evidence-based medicine (EBM) looks to inform patient care with the totality of available relevant evidence. Systematic reviews are the cornerstone of EBM and are critical to modern healthcare, informing everything from national health policy to bedside decision-making. But conducting systematic reviews is extremely laborious (and hence expensive): producing a single review requires thousands of person-hours. Moreover, the exponential expansion of the biomedical literature base has imposed an unprecedented burden on reviewers, thus multiplying costs. Researchers can no longer keep up with the primary literature, and this hinders the practice of evidence-based care. To mitigate this issue, I will discuss past and recent advances in machine learning and natural language processing methods that look to optimize the practice of EBM. These include methods for semi-automating evidence identification (i.e., citation screening) and more recent work on automating the extraction of structured data from full-text published articles describing clinical trials. As I will discuss, these problems pose challenging problems from a machine learning vantage point, and hence motivate the development of novel approaches. I will present evaluations of these methods in the context of EBM. About the speaker:Byron Wallace is an assistant professor in the School of Information at the University of Texas at Austin. He holds a PhD in Computer Science from Tufts University, where he was advised by Carla Brodley. Prior to joining UT, he was research faculty at Brown University, where he was part of the Center for Evidence-Based Medicine and also affiliated with the Brown Laboratory for Linguistic Information Processing. His primary research is in applications of machine learning and natural language processing to problems in health -- particularly in evidence-based medicine.Wallace's work is supported by grants from the NSF and the ARO. He was recognized as The Runner Up for the 2013 ACM Special Interest Group on Knowledge Discovery and Data Mining (SIG KDD) for his thesis work. |
Friday, November 21, 2014, 11:00AM
|
What Google Glass means for the future of photographyMarc Levoy [homepage]
Although head-mounted cameras (and displays) are not new, Google Glass has
the
potential to make these devices commonplace. This has implications for the
practice, art, and uses of photography. So what's different about doing
photography with Glass? First, Glass doesn't work like a conventional
camera;
it's hands-free, point-of-view, always available, and instantly triggerable.
Second, Glass facilitates different uses than a conventional camera:
recording
documents, making visual todo lists, logging your life, and swapping eyes
with
other Glass users. Third, Glass will be an open platform, unlike most
cameras.
This is not easy, because Glass is a heterogeneous computing platform, with
multiple processors having different performance, efficiency, and
programmability. The challenge is to invent software abstractions that
allow
control over the camera as well as access to these specialized processors.
Finally, devices like Glass that are head-mounted and perform computational
photography in real time have the potential to give wearers "superhero
vision",
like seeing in the dark, or magnifying subtle motion or changes. If such
devices can also perform computer vision in real time and are connected to
the
cloud, then they can do face recognition, live language translation, and
information recall. The hard part is not imagining these capabilities, but
deciding which ones are feasible, useful, and socially acceptable.
About the speaker:Marc Levoy is the VMware Founders Professor of Computer Science and Electrical Engineering, Emeritus, at Stanford University. He received a Bachelor's and Master's in Architecture from Cornell University in 1976 and 1978, and a PhD in Computer Science from the University of North Carolina at Chapel Hill in 1989. In the 1970's Levoy worked on computer animation, developing a cartoon animation system that was used by Hanna-Barbera Productions to make The Flintstones, Scooby Doo, and other shows. In the 1980's Levoy worked on volume rendering, a technique for displaying three-dimensional functions such as computed tomography (CT) or magnetic resonance (MR) data. In the 1990's he worked on 3D laser scanning, culminating in the Digital Michelangelo Project, in which he and his students spent a year in Italy digitizing the statues of Michelangelo. In the 2000's he worked on computational photography and microscopy, including light field imaging as commercialized by Lytro and other companies. At Stanford he taught computer graphics and the science of art, and still teaches digital photography. Outside of academia, Levoy co-designed the Google book scanner, launched Google's Street View project, and currently leads a team in GoogleX that has worked on Project Glass and the Nexus 6 HDR+ mode. Awards: Charles Goodwin Sands Medal for best undergraduate thesis (1976), National Science Foundation Presidential Young Investigator (1991), ACM SIGGRAPH Computer Graphics Achievement Award (1996), ACM Fellow (2007). |
Friday, January 23, 2015, 11:00AM
|
Relative Upper Confidence Bound for the K-Armed Dueling Bandit ProblemShimon Whiteson [homepage]
In this talk, I will propose a new method for the K-armed dueling bandit problem, a variation on the regular K-armed bandit problem that offers only relative feedback about pairs of arms. Our approach extends the Upper Confidence Bound algorithm to the relative setting by using estimates of the pairwise probabilities to select a promising arm and applying Upper Confidence Bound with the winner as a benchmark. We prove a sharp finite-time regret bound of order O(K log T) on a very general class of dueling bandit problems that matches a lower bound proven by Yue et al. In addition, our empirical results using real data from an information retrieval application show that it greatly outperforms the state of the art.
About the speaker:Shimon Whiteson is an Associate Professor and the head of the Autonomous Agents section at the Informatics Institute of the University of Amsterdam. He was recently awarded an ERC Starting Grant from the European Research Council as well as a VIDI grant for mid-career researchers from the Dutch national science foundation. He received his PhD in 2007 from the University of Texas at Austin, under the supervision of Peter Stone. His research focuses on decision-theoretic planning and learning, with applications to robotics, multi-camera tracking systems, and information retrieval. |
Friday, January 23, 2015, 2:00PM
|
Modern Exact and Approximate Combinatorial Optimization Algorithms for Graphical ModelsRina Dechter [homepage]
In this talk I will present several principles behind state of the art algorithms for solving combinatorial optimization tasks defined over graphical models (Bayesian networks, Markov networks, constraint networks, satisfiability) and demonstrate their performance on some benchmarks. Specifically I will present branch and bound search algorithms which explore the AND/OR search space over graphical models and thus exploit problem's decomposition (using AND nodes), equivalence (by caching) and pruning irrelevant subspaces via the power of bounding heuristics. In particular I will show how the two ideas of mini-bucket partitioning which relaxes the input problem using node duplication only, combined with linear programming relaxations ideas which optimize cost-shifting/re-parameterization schemes, can yield tight bounding heuristic information within systematic, anytime, search. Notably, a solver for finding the most probable explanation (MPE or MAP), embedding these principles has won first place in all time categories in the 2012 PASCAL2 approximate inference challenge, and first or second place in the UAI-2014 competitions. Recent work on parallel/distributed schemes and on m-best anytime solutions may be mentioned, as time permits. Parts of this work were done jointly with: Lars Otten, Alex Ihler, Radu Marinescu, Natasha Flerova, Kalev Kask About the speaker:Rina Dechter is a professor of Computer Science at the University of California, Irvine. She received her PhD in Computer Science at UCLA in 1985, an MS degree in Applied Mathematic from the Weizmann Institute and a B.S in Mathematics and Statistics from the Hebrew University, Jerusalem. Her research centers on computational aspects of automated reasoning and knowledge representation including search, constraint processing and probabilistic reasoning. Professor Dechter is an author of Constraint Processing published by Morgan Kaufmann, 2003, has authored over 150 research papers, and has served on the editorial boards of: Artificial Intelligence, the Constraint Journal, Journal of Artificial Intelligence Research and journal of Machine Learning (JLMR). She was awarded the Presidential Young investigator award in 1991, is a fellow of the American association of Artificial Intelligence since 1994, was a Radcliffe Fellowship 2005-2006, received the 2007 Association of Constraint Programming (ACP) research excellence award and is a 2013 Fellow of the ACM. She has been Co-Editor-in-Chief of Artificial Intelligence, since 2011. |
Friday, January 30, 2015, 12:00PM
|
Deeply Integrating Human and Machine Intelligence To Power Deployable SystemsJeffrey Bigham [homepage]
Over the past few years, I have been developing and deploying interactive crowd-powered systems that solve characteristic ?hard? problems to help people get things done in their everyday lives. For instance, VizWiz answers visual questions for blind people in seconds, Legion drives robots in response to natural language commands, Chorus holds helpful general conversations with human partners, and Scribe converts streaming speech to text in less than five seconds. My research envisions a future in which the intelligent systems that we have dreamed about for decades, which have inspired generations of computer scientists from its beginning, are brought about for the benefit of people. My work illustrates a path for achieving this vision by leveraging the on-demand labor of people to fill in for components that we cannot currently automate, by building frameworks that allow groups to do together what even expert individuals cannot do alone, and by gradually allowing machines to take over in a data-driven way. A crowd-powered world may seem counter to the goals of computer science, but I believe that it is precisely by creating and deploying the systems of our dreams that will learn how to advance computer science to create the machines that will someday realize them. [Talk hosted by UT School of Information; Talk info here] About the speaker:Jeffrey P. Bigham is an Associate Professor in the Human-Computer Interaction and Language Technologies Institutes in the School of Computer Science at Carnegie Mellon University. He uses clever combinations of crowds and computation to build truly intelligent systems, often with a focus on systems supporting people with disabilities. Dr. Bigham received his B.S.E degree in Computer Science from Princeton University in 2003, and received his Ph.D. in Computer Science and Engineering from the University of Washington in 2009. From 2009 to 2013, he was an Assistant Professor at the University of Rochester, where he founded the ROC HCI human-computer interaction research group. He has been a Visiting Researcher at MIT CSAIL, Microsoft Research, and Google[x]. He has received a number of awards for his work, including the MIT Technology Review Top 35 Innovators Under 35 Award, the Alfred P. Sloan Fellowship, and the National Science Foundation CAREER Award. |
Monday, February 2, 2015, 11:00AM
|
Modeling Lexically Divergent Paraphrases in Twitter (and Shakespeare!)Wei Xu [homepage]
Paraphrases are alternative linguistic expressions of the same meaning. Identifying paraphrases is fundamental to many natural language processing tasks and has been extensively studied for the standard contemporary English. In this talk I will present MULTIP (Multi-instance Learning Paraphrase Model), a joint word-sentence alignment model suited to identify paraphrases within the noisy user-generated texts on Twitter. The model infers latent word-level paraphrase anchors from only sentence-level annotations during learning. This is a major departure from previous approaches that rely on lexical or distributional similarities over sentence pairs. By reducing the dependence on word overlap as evidence of paraphrase, our approach identifies more lexically divergent expressions with equivalent meaning. For experiments, we constructed a Twitter Paraphrase Corpus using a novel and efficient crowdsourcing methodology. Our new approach improves the state-of-the-art performance of a method that combines a latent space model with a feature-based supervised classifier. I will also present findings on paraphrasing between standard English and Shakespearean styles. Joint work with Chris Callison-Burch (UPenn), Bill Dolan (MSR), Alan Ritter (OSU), Yangfeng Ji (GaTech), Colin Cherry (NRC) and Ralph Grishman (NYU). About the speaker:Wei Xu is a postdoc in Computer and Information Science Department at University of Pennsylvania, working with Chris Callison-Burch. Her research focuses on paraphrases, social media and information extraction. She received her PhD in Computer Science from New York University. She is organizing the SemEval-2015 shared task on Paraphrase and Semantic Similarity in Twitter, and the Workshop on Noisy User-generated Text at ACL-2015 (http://noisy-text.github.io/). During her PhD, she visited University of Washington for two years and interned at Microsoft Research, ETS and Amazon.com. |
Friday, February 13, 2015, 11:00AM
|
Keyphrase Extraction in Citation Networks: How Do Citation Contexts Help?Cornelia Caragea [homepage]
Keyphrase extraction is defined as the problem of automatically extracting descriptive phrases or concepts from documents. Keyphrases for a document act as a concise summary of the document and have been successfully used in many applications such as query formulation, document clustering, classification, recommendation, indexing, and summarization. Previous approaches to keyphrase extraction generally used the textual content of a target document or a local neighborhood that consists of textually-similar documents. We posit that, in a scholarly domain, in addition to a document’s textual content and textually-similar neighbors, other informative neighborhoods exist that have the potential to improve keyphrase extraction. In a scholarly domain, research papers are not isolated. Rather, they are highly inter-connected in giant citation networks, in which papers cite or are cited by other papers in appropriate citation contexts, i.e., short text segments surrounding a citation’s mention. These contexts are not arbitrary, but they serve as brief summaries of a cited paper. We effectively exploit citation context information for keyphrase extraction and show remarkable improvements in performance over strong baselines in both supervised and unsupervised settings.
About the speaker:Cornelia Caragea is an Assistant Professor at the University of North Texas, where she directs the Machine Learning group. Her research interests lie at the intersection of machine learning, information retrieval, and natural language processing, with applications to scholarly digital libraries. She has published research papers in prestigious venues such as AAAI, IJCAI, WWW, and ICDM. Cornelia reviewed for many journals including Nature, ACM Transactions on Intelligent Systems and Technology, and IEEE Transactions on Knowledge and Data Engineering, served on several NSF panels, and was a program committee member for top conferences such as ACL, IJCAI, Coling, and CIKM. She also helped organize several workshops on scholarly big data in conferences such as AAAI, IEEE BigData, and CIKM.Cornelia earned a Bachelor of Science degree in Computer Science and Mathematics from the University of Bucharest, and a Ph.D. in Computer Science from the Iowa State University. Prior to joining UNT in Fall 2012, she was a post-doctoral researcher at the Pennsylvania State University. Her appointment at UNT marks one of the University’s unique approaches to faculty hires: she is part of the Knowledge Discovery from Digital Information research cluster. |
Friday, February 20, 2015, 11:00AM
|
Human-Robot CollaborationStefanie Tellex [homepage]
In order for robots to collaborate with humans,
they must infer
helpful actions in the physical world by observing the human's
language, gesture,
and actions. A particular challenge for robots
that operate under uncertainty is identifying and
recovering from
failures in perception, actuation, and language interpretation. I
will describe our approaches to automatic failure recovery using
language and probabilistic methods. First we describe how a robot can
use a probabilistic language grounding framework to employ
information-theoretic dialog strategies, asking targeted questions to
reduce uncertainty about different parts of a natural language
command. Second, I will show how to invert a model for interpreting
language to generate targeted natural language requests for help from
a human partner, enabling a robot team to actively solicit help from a
person when they encounter problems. And third, I will describe steps
toward incremental interpretation of language and gesture as an
enabling technology for making robots that use coordination actions to
establish common ground with their human partner. This approach
points the way toward more general models of human-robot collaboration
building world models from both linguistic and non-linguistic input,
following complex grounded natural language commands, and engaging in
fluid, flexible
collaboration with their human partners.
About the speaker:Stefanie Tellex is an Assistant Professor of Computer Science and Assistant Professor of Engineering at Brown University. Her group, the Humans To Robots Lab, creates robots that seamlessly collaborate with people to meet their needs using language, gesture, and probabilistic inference, aiming to empower every person with a collaborative robot. She completed her Ph.D. at the MIT Media Lab in 2010, where she developed models for the meanings of spatial prepositions and motion verbs. Her postdoctoral work at MIT CSAIL focused on creating robots that understand natural language. She has published at SIGIR, HRI, RSS, AAAI, IROS, ICAPs and ICMI, winning Best Student Paper at SIGIR and ICMI, and Best Paper at RSS. She was named one of IEEE Spectrum's AI's 10 to Watch and won the Richard B. Salomon Faculty Research Award at Brown University. Her research interests include probabilistic graphical models, human-robot interaction, and grounded language understanding. |
Friday, April 3, 2015, 11:00AM
|
Machine Learning about People from their LanguageNoah Smith [homepage]
This talk describes new analysis algorithms for text data aimed at understanding the social world from which the data emerged. The political world offers some excellent questions to explore: Do US presidential candidates "move to the political center" after winning a primary election? Are Supreme Court justices swayed by amicus curiae briefs, documents crafted at great expense? I'll show how our computational models capture theoretical commitments and uncertainty, offering new tools for exploring these kinds of questions and more. Time permitting, we'll close with an analysis of a quarter million biographies, discussing what can be discovered about human lives as well as those who write about them. The primary collaborators on this research are my Ph.D. students David Bamman and Yanchuan Sim; collaborators from the Political Science Department at UNC Chapel Hill, Brice Acree, and Justin Gross; and Bryan Routledge from the Tepper School of Business at CMU. About the speaker:Noah Smith is Associate Professor of Language Technologies and Machine Learning in the School of Computer Science at Carnegie Mellon University. In fall 2015, he will join the University of Washington as Associate Professor of Computer Science & Engineering. He received his Ph.D. in Computer Science from Johns Hopkins University in 2006 and his B.S. in Computer Science and B.A. in Linguistics from the University of Maryland in 2001. His research interests include statistical natural language processing, especially unsupervised methods, machine learning for structured data, and applications of natural language processing. His book, Linguistic Structure Prediction, covers many of these topics. He has served on the editorial board of the journals Computational Linguistics (2009–2011), Journal of Artificial Intelligence Research (2011–present), and Transactions of the Association for Computational Linguistics (2012–present), and as the secretary-treasurer of SIGDAT (2012–present). His research group, Noah's ARK, is currently supported by the NSF, DARPA, IARPA, ARO, and gifts from Amazon and Google. Smith's work has been recognized with a Finmeccanica career development chair at CMU (2011–2014), an NSF CAREER award (2011–2016), a Hertz Foundation graduate fellowship (2001–2006), numerous best paper nominations and awards, and coverage by NPR, BBC, CBC, New York Times, Washington Post, and Time. |
Friday, April 10, 2015, 11:00AM
|
Faster Learning for Better DecisionsEmma Brunskill [homepage]
A fundamental goal of artificial intelligence is to create agents that learn to make good decisions as they interact with a stochastic environment. Some of the most exciting and valuable potential applications involve systems that interact directly with humans, such as intelligent tutoring systems or medical interfaces. In these cases, sample efficiency is highly important, as each decision, good or bad, is impacting a real person. I will describe our research on tackling this challenge, including transfer learning across sequential decision making tasks, as well as its relevance to improving educational tools.
About the speaker:Emma Brunskill is an assistant professor in the computer science department at Carnegie Mellon University. She is also affiliated with the machine learning department at CMU. She works on reinforcement learning, focusing on applications that involve artificial agents interacting with people, such as intelligent tutoring systems. She is a Rhodes Scholar, Microsoft Faculty Fellow and NSF CAREER award recipient, and her work has received best paper nominations in Education Data Mining (2012, 2013) and CHI (2014). |