Artificial Intelligence Seminar

Fall 2013 Friday 12:00-1:15 Upson 5130

The AI seminar will meet weekly for lectures by graduate students, faculty, and researchers emphasizing work-in-progress and recent results in AI research. Lunch will be served starting at noon, with the talks running between 12:15 and 1:15. The new format is designed to allow AI chit-chat before the talks begin. Also, we're trying to make some of the presentations less formal so that students and faculty will feel comfortable using the seminar to give presentations about work in progress or practice talks for conferences.

 August 21st Speakers: Ankur Handa Host: Ashutosh Saxena Abstract: "Higher frame-rates offers a natural advantage to track rapid motion, but advanced real-time vision systems rarely exceed the standard 10-60Hz range, arguing that the computation required would be too great. However, increasing frame-rate is mitigated by reduced computational cost per frame in trackers which take advantage of prediction; the linearisations that are commonly used in the standard cost functions become increasingly more valid when the frame-rate is increased. Additionally, when we consider the physics of image formation, high frame-rate implies that the upper bound on shutter time is reduced, leading to less motion blur but more noise. So, putting these factors together, how are application-dependent performance requirements of accuracy, robustness and computational cost optimised as frame-rate varies? Using 3D camera tracking as our test problem, and analysing a fundamental dense whole image alignment approach, a systematic investigation can be carried out via the careful synthesis of photorealistic video using ray-tracing of a detailed 3D scene, experimentally obtained photometric response and noise models, and rapid camera motions. The multi-frame-rate, multi-resolution, multi-light-level dataset that is used for exhaustive analysis of the experiment over different frame-rates is based on tens of thousands of hours of CPU rendering time. We will look at the quantitative conclusions that emerge out of these experiments about frame-rate selection and highlight the crucial role of full consideration of physical image formation in pushing tracking performance. " “The AI-Seminar is sponsored by Yahoo!”
 August 30th Speakers: Yael Moses, The Interdisciplinary Center, Israel Host: Noah Snavely Bio: I joined  the school of computer science at the Interdisciplinary Center, Herzliya in 1999. I am Associated Professor in the Efi Arazi School of Computer Science.  I finished my Ph.D. in 1994 at the Department of Computer Science at the Weizmann Institute of Science, Rehovot. During the years 1993-1994 I was a Post-doctoral fellow, at the Robotics group of Oxford University. From 1997-1998 I was a  post-doctoral fellow at the Weizmann Institute of Science. I spent the years 2004-2007 on sabbatical in Sydney at NICTA and UNSW. My research interest is in the field of computer vision, in which images or videos are analyzed to understand the scene viewed by the cameras. My early work concentrated on theoretical limitations of object recognition based on a single image. Later, my interest shifted to deal with multi-camera systems, which extend the single-camera setup in a nontrivial way, and significantly increase the capabilities of computer vision systems. My main passion is in developing efficient solutions for applications that are too challenging to solve using a single camera by using a multi-camera system. I am especially interested in efficient fusion of the available information from a set of cameras, and the scalability of computer vision algorithms for large systems of cameras. Title: Multi-Camera Systems: Synchronization, Color-matching, and Applications Abstract: The use of multi-camera systems allows to overcome limitations of a single camera, but it also poses new challenges. This talk will consider three aspects of multi-camera settings.  I start by presenting an application of a multi-camera system, a method to track people in a dense crowd in scenes that are considered very challenging to a single camera system. I will then present a method for online synchronization that relies only on the video sequences in challenging settings for which features or object tracking are too hard and frame dropping may present.  Time synchronization of video sequences is often a must for the analysis of visual information from a multi-camera system.  Finally, I will propose a method for computing a piecewise consistent color mappings between pairs of imaged. Our method allows to overcome the variations between colors of corresponding regions in different images taken under  various acquisition conditions (e.g., different viewpoint and different illumination). Under these conditions, commonly used color mappings fails. Our method can be used to overcome the limitations of methods that assumes corresponding regions have similar colors, in cases where this assumption does not hold.  This talk will be based on  joint works with Ran Eshel, Dima Pundik, Sefy Kagarlitsky and Yacov Hel-Or. “The AI-Seminar is sponsored by Yahoo!”
 September 6th Speakers: Dr. Alejandro (Alex) Jaimes, Director of Research at Yahoo Host: Prof. Tsuhan Chen Bio: Dr. Alejandro (Alex) Jaimes is Director of Research at Yahoo! where he is in charge of the Social Media Engagement (SOMER) and Learning for Multimedia and Vision (LMV) groups in Barcelona and Bangalore. In the Spring of 2013 he was a visiting Professor at KAIST’s (South Korea) Web Science Department under the WCU program. His research focuses on Human-Centered Computing, particularly in the areas of social media and Multimedia. The output of his teams’ research has been included in several products at Yahoo! and he led the launch of Yahoo! Clues, a product created in 2010. Dr. Jaimes is general chair of ACM Multimedia 2013, Developers Track Chair for WWW 2014, Practice and Experience track chair for WWW 2013, the founder of the ACM Multimedia Interactive Art program, and Industry Track chair for ACM RecSys 2010 and UMAP 2013, among others. His work has led to over 80 technical publications in international conferences and journals. He has been an invited speaker at the Big Data & Analytics Innovation Summit (2013), Practitioner Web Analytics (2010), CIVR 2010, ECML-PKDD 2010 and KDD 2009 and (Industry tracks), ACM Recommender Systems 2008 (panel), DAGM 2008 (keynote), and several others. Before joining Yahoo! Dr. Jaimes was a visiting professor at U. Carlos III in Madrid and founded and managed the User Modeling and Data Mining group at Telefónica Research. Prior to that Dr. Jaimes was Scientific Manager at IDIAP-EPFL (Switzerland), and was previously at Fuji Xerox (Japan), IBM TJ Watson (USA), IBM Tokyo Research Laboratory (Japan), Siemens Corporate Research (USA), and AT&T Bell Laboratories (USA). Dr. Jaimes received a Ph.D. in Electrical Engineering (2003) and a M.S. in Computer Science from Columbia U. (1997) in NYC. Title: Insights from Big Data: Interaction, Design, and Innovation   Abstract: In recent years, our ability to process large amounts of data has increased significantly, creating many opportunities for innovation. Having large quantities of data, however, does not necessarily turn into actionable insights that make a difference for users in consumer applications. In this talk I will give a quick overview of some ways in which “big data” can be used in industry, with a particular focus on Human-Centered approaches to innovation. In particular, I will discuss how the combination of qualitative and quantitative methods can be of benefit, giving examples around social media and giving an overview of some of the areas of research I am currently focusing on at Yahoo!. Within this context, I will outline a blueprint for a research framework as it applies to innovation, and discuss specific technical approaches within that framework. I will argue on the importance of taking a human-centered view and highlight what I consider the most fundamental problems in computer science today from that perspective. “The AI-Seminar is sponsored by Yahoo!”
 September 13th Speakers: Brad Gulko, Cornell University Host: Ashutosh Saxena Title: Maximin Safety: When Failing to Lose is Preferable to Trying to Win. Abstract: Finally, an AI seminar that may HELP you get a date! In games, social robotics, and behavioral economics, the ability to predict human behavior can be very desirable. While humans are often modeled as expected-utility maximizing actors, observed behaviors like the Ellsberg paradox show that humans regularly make choices inconsistent with expected utility under ANY probability distribution. One such behavior is the decoy effect, in which a dominated option drives preferences toward dominating alternatives. Decoy parables shows up in popular media and have been empirically demonstrated in human preferences for consumer goods, general physical attractiveness as well as among the birds and bees. Unfortunately, commonly used decision models based on Expected Utility, Maximin Utility, Minimax Regret and even Pareto Optimality all require insensitivity of preferences to the introduction of dominated alternatives, and thus are inconsistent with decoy behavior. To accommodate decoy, we introduce "safety", a quantity describing the distance to worst outcomes, much the way regret describes distance to best ones. Safety intuitively blends loss-avoidance with preference for relative performance over absolute into "loser-avoidance", a conceptual dual to regret's "winner-attraction". We show how safety-seeking behavior can lead to the decoy effect and provide the key element to the axiomatization of the Maximin Safety decision rule, based on on Stoye's axiomatization of Regret. This talk is based on a 30 minute talk I gave at the 2013 ECSQARU conference. It is notation-light and there should be lots of time for questions, beverages & extra pizza. “The AI-Seminar is sponsored by Yahoo!”
 September 20th Speakers: Paul Bennett, CMU Host: Thorsten Joachims Bio: Paul Bennett is a Researcher in the Context, Learning & User Experience for Search (CLUES) group at Microsoft Research where he focuses on the development, improvement, and analysis of machine learning and data mining methods as components of real-world, large-scale adaptive systems. His research has advanced techniques for ensemble methods and the combination of information sources, calibration, consensus methods for noisy supervision labels, active learning and evaluation, supervised classification (with an emphasis on hierarchical classification) and ranking with applications to information retrieval, crowdsourcing, behavioral modeling and analysis, and personalization. He completed his dissertation on combining text classifiers using reliability indicators in 2006 at Carnegie Mellon where he was advised by Profs. Jaime Carbonell and John Lafferty. Title: Mining and Learning from Context in Web Search Abstract: User and behavioral modeling can play a critical role in a variety of online services such as web search, advertising, e-commerce, and news recommendation. With regard to web search, while information retrieval has made significant progress in returning relevant results for a single query, much search activity is conducted within a richer context of a current task focus, recent search activities as well as longer-term preferences. For example, our ability to accurately interpret the current query can be informed by knowledge of the web pages a searcher was viewing when initiating the search or recent actions of the searcher such as queries issued, results clicked, and pages viewed. We develop a framework that enables representation of a broad variety of context including the searcher's long-term interests, recent activity, current focus, and other user characteristics. We then demonstrate how that can be used to improve the quality of search results. We describe recent progress on three key challenges in this domain: enriching information retrieval via automatically generated metadata; mining contextual signals from large scale logs; and using contextual epresentations in learning to improve both standard ad hoc and personalized retrieval. This talk will present joint work with Fedor Borisyuk, Jinyoung Kim, Nam Nguyen, Lidan Wang, Filip Radlinski, Ryen White, Kevyn Collins-Thompson, Wei Chu, Susan Dumais, Peter Bailey, Emine Yilmaz, Xiaoyuan Cui, David Sontag, Sebastian de la Chica, and Bodo von Billerbeck. “The AI-Seminar is sponsored by Yahoo!”
 September 27th Speakers: Kevin Leyton-Brown, University of British Columbia Host: Eva Tardos Bio: Kevin Leyton-Brown is an associate professor of computer science at the University of British Columbia. He holds a PhD and M.Sc. from Stanford University (2003; 2001) and a B.Sc. from McMaster University (1998). He studies the intersection of computer science and microeconomics, addressing computational problems in economic contexts and incentive issues in multiagent systems. He also applies machine learning to the automated design and analysis of algorithms for solving hard computational problems. He has co-written two books, "Multiagent Systems" and "Essentials of Game Theory," and over ninety peer-refereed technical articles. He and his coauthors have received paper awards from JAIR, ACM-EC, AAMAS and LION, and numerous medals for the portfolio-based SAT solver SATzilla at international SAT competitions (2003-12). He was program chair for the ACM Conference on Electronic Commerce (ACM-EC) in 2012, and serves as an associate editor for the Journal of Artificial Intelligence Research (JAIR), the Artificial Intelligence Journal (AIJ), and ACM Transactions on Economics and Computation. He co-taught the Coursera course "Game Theory" to over 130,000 students, and has received awards for his teaching at UBC. He split his 2010-11 sabbatical between Makerere University in Kampala, Uganda, and the Institute for Advanced Studies at Hebrew University of Jerusalem, Israel. He has served as a consultant for Trading Dynamics Inc., Ariba Inc., Cariocas Inc., Auctionomics, Inc., and kudu.ug, and was scientific advisor to Vancouver-based Zite Inc. until it was acquired by CNN in 2011. Title: “Beyond Equilibrium: Predicting Human Behavior in Normal Form Games” Abstract: It is common to assume that agents will adopt Nash equilibrium strategies; however, experimental studies have demonstrated that Nash equilibrium is often a poor description of human players' behavior in unrepeated normal-form games. We analyzed four widely studied models (QRE, Lk, Cognitive Hierarchy, QLk) that aim to describe actual, rather than idealized, human behavior in such games.  We performed a meta-analysis of these models, leveraging nine different data sets from the literature recording human play of two-player games.  We began by evaluating the models' generalization or predictive performance, asking how well a model fits unseen test data after having had its parameters calibrated based on separate training data. Surprisingly, we found that (what we dub) the QLk model of Stahl and Wilson [1994] consistently achieved the best performance. Motivated by this finding, we describe methods for analyzing the posterior distributions over a model's parameters. We found that QLk's parameters were being set to values that were not consistent with their intended economic interpretation. We thus explored variations of QLk, ultimately identifying a new model family that has fewer parameters, gives rise to more parsimonious parameter values, and achieves better predictive performance. “The AI-Seminar is sponsored by Yahoo!”
 October 4th Speakers: David Mimno (Cornell) Host: Lillian Lee Title: Scalable Inference for Latent Dirichlet Allocation Abstract: Topic models are a means for approaching unstructured text collections, by representing documents as combinations of automatically discovered themes. These models are most useful when trained with hundreds or even thousands of topics on massive corpora, but exact inference is intractable even for simple models. In this talk I will present work on scalable inference for topic models, including sparse Gibbs sampling, stochastic variational inference, and recent work in spectral methods. “The AI-Seminar is sponsored by Yahoo!”
 October 11th Speakers: No Seminar- FALL BREAK Host: Bio: Title: Abstract: “The AI-Seminar is sponsored by Yahoo!”
 October 18th Speakers: Honglak Lee, University of Michigan Host: Ashutosh Saxena Bio: Honglak Lee is an Assistant Professor of Computer Science and Engineering at the University of Michigan, Ann Arbor. He received his Ph.D. from Computer Science Department at Stanford University in 2010, advised by Andrew Ng. His primary research interests lie in machine learning, which spans over deep learning, unsupervised and semi-supervised learning, transfer learning, graphical models, and optimization. He also works on application problems in computer vision, audio recognition, robot perception, and text processing. His work received best paper awards at ICML and CEAS. He received a Google Faculty Research Award, and he has served as a guest editor of IEEE TPAMI Special Issue on Learning Deep Architectures.  He was selected by IEEE Intelligent Systems as one of AI's 10 to Watch. Title: Representation Learning: Progress, Challenges, and New Directions Abstract: Although machine learning is a powerful tool for artificial intelligence and data mining problems, the quality of the feature representations has been a critical limiting factor in success of machine learning systems. To address this problem, representation learning algorithms have recently emerged as ways to learn feature hierarchies from unlabeled and labeled data. In this talk, I will present my perspectives on the progress and challenges, as well as my recent related work on some new problems: (1) output representation learning for structured output prediction, (2) weakly supervised representation learning, (3) disentangling factors of variation with deep generative models. “The AI-Seminar is sponsored by Yahoo!”
 October 25th Speakers: Bharat R Rao, Ph.D. Deloitte Financial Advisory Services Host: Thorsten Joachims Bio: Dr. Rao is a Director in the Advanced Analytics practice of Deloitte Financial Advisory Services LLP, focusing on Health Care and Life Sciences Analytics, and reducing Fraud, Waste and Abuse in healthcare.  He is recognized as a leading expert in big data, text mining, health care analytics and predictive modeling. He has received many awards, including the        ACM SIGKDD lifetime Service Award, 2011 for “contributions to society for pioneering applications of data mining,” is the only two-time winner of Data Mining Practice Prize for the best industrial and government data analytics solution deployment, and was awarded the Siemens Inventor of the Year. He is a frequent keynote speaker at conferences, has published 100 scholarly articles, one book, and has 60 patents. Dr. Rao has over 25 years of data analytics expertise, 15 of which have been focused on healthcare and life sciences. Prior to joining Deloitte FAS, Dr. Rao held various leadership positions at Siemens Healthcare and Siemens Corporation.  At Siemens, he was responsible for the innovations pipeline for a multi-billion dollar Healthcare IT business, led the development of data analytics and cloud solutions that analyzed millions of EHR’s and large clinical, claims, text, and image databases, and founded the Siemens Data Mining R&D program. Dr. Rao’s innovations include automated analytics solutions for (i) detecting potentially harmful events with low false alarm rates; (ii) determining compliance with and deviations from policies; (iii) extracting relevant information and documents from multi-source data (free text, images, web); (iv) predicting risk of failure, poor outcomes and suggesting interventions by proactively monitoring data streams (transactions, service logs); (v) securely sharing, sampling, and mining personally identifiable information (PII) and other protected data across different institutions. Dr. Rao received his MS and PhD in Electrical & Computer Engineering from the University of Illinois, Urbana-Champaign, and his Bachelor of Technology in Electronics Engineering from the Indian Institute of Technology, Madras, India.  Outside the sphere of technology, his interests include cricket (the game, not the insect), classic rock, the history of innovation, and coaching soccer. Title: Rapid Learning Systems for Electronic Health Records Abstract: Consider the following healthcare trends: There is a tremendous increase in the amount of patient, life sciences and process data in electronic form, fueled by advances in healthcare IT technology, and health reform legislation. The amount of medical information (e.g., evidence-based knowledge) and published knowledge is said to be doubling every few years. There is an explosion in the number of available therapies and diagnostic options for patient care, often enabling precise targeting of therapy to disease conditions. Despite these advances, healthcare is facing a crisis: namely, there is a steady unsustainable increase in medical costs without a corresponding improvement of patient outcomes.  We believe that direct analysis of electronic health records can play a key role in overcoming these critical problems that have massive impact on society at large. We begin by list some of the machine learning and data mining challenges in directly learning from electronic health records (EHRs).  We will delve into some of these challenges, a) maintaining patient privacy while learning predictive models for individualized therapy selection from EHRs, and b) a Bayesian framework for dealing with highly “noisy” patient data (Variation in the amount / type / reliability of data for different patients, and the lack of reliable ground truth).  We present an emerging application in healthcare – the use of machine learning to proactively detect and prevent healthcare fraud, a fundamental paradigm shift with the potential to reduce health costs by hundreds of \$B.  We conclude the talk with a glimpse of a future where medical systems could be continually analyzed for optimizing healthcare costs and outcomes. “The AI-Seminar is sponsored by Yahoo!”
 November 1st Speaker: Yejin Choi, Stony Brook University Host: Claire Cardie Bio: Yejin Choi received her Ph.D. in Computer Science at Cornell University, and BS in Computer Science and Engineering at Seoul National University. She spent the summer of 2009 as a research intern at Yahoo! Research and joined the faculty of Computer Science Department at Stony Brook University in Sep 2010. http://www.cs.stonybrook.edu/~ychoi/ Title: Learning to Describe the Visual World: Language Meets Vision Abstract: The web is increasingly visual with billions of images interwoven with text. This sheer abundance of web data with mixed modality raises new challenges and opportunities for integrative models connecting computer vision with natural language processing. In this talk, I will present a series of our work for learning to describe the visual world, as a means to tap into the visual knowledge shared between natural language text and images. I will first describe our explorations to automatically composing image descriptions, ranging from highly formulaic, robotic ones to more creative and human-like descriptions. I will then discuss the challenge of information misalignment between naturally existing images and their descriptions, and present our attempts to improve the alignment. Finally, linking to the Prototype Theory developed in Psychology, I will briefly introduce our work that learns to map encyclopedic categories of objects into entry-level categories, essentially resulting in a large-scale, data-driven emulation of what have been laboratory experiments in Psychology.   “The AI-Seminar is sponsored by Yahoo!”
 November 8th Speakers: Bruno Abrahao Host: Bobby Kleinberg Title: Extracting hidden structures in social and information networks Abstract: The accelerated evolution of online social interactions and information systems has brought a marked growth in their complexity. The dynamics of these systems are not entirely regulated or engineered, but are instead governed by hidden structures that form as a result of organic growth, but are neither directly observable nor predictable by design. Understanding these structures is crucial for harnessing the benefits of social and information networks while supporting their health and growth. Leveraging the vast amounts of data generated by the Web and the sciences, Network Science has achieved remarkable progress at identifying and modeling certain hidden structures. Approaching these problems in a rigorous way will be crucial to enable further discoveries. This talk surveys the insights gained from applying a novel, principled approaches to three existing modeling and learning tasks: community detection, network inference, and Internet modeling. I will conclude the talk by illustrating the application of sociological theories to guide empirical analysis of online social network data, revealing hidden social structures that enable a deeper understanding of our own behavior. Joint work with  Flavio Chierichetti (Sapienza), Karen Cook (Stanford), John Hopcroft (Cornell), Robert Kleinberg (Cornell), Alessandro Panconesi (Sapienza),  Sucheta Soundarajan (Rutgers), Bogdan State (Stanford) “The AI-Seminar is sponsored by Yahoo!”
 November 15th Speaker: Ashesh Jain, Cornell University Host: Ashutosh Saxena Title: Beyond geometric path planning: Learning context-driven trajectory preferences via sub-optimal feedback Abstract: We consider the problem of learning preferences over trajectories for mobile manipulators such as personal robots and assembly line robots. The preferences we learn are more intricate than those arising from simple geometric constraints on robot's trajectory, such as distance of the robot from human etc. Our preferences are rather governed by the surrounding context of various objects and human interactions in the environment. Such preferences makes the problem challenging because the criterion of defining a good trajectory now varies with the task, with the environment and across the users. Furthermore, demonstrating optimal trajectories (e.g., learning from expert's demonstrations) is often challenging and non-intuitive on high degrees of freedom manipulators. In this work, we propose an approach that requires a \textit{non-expert} user to only incrementally improve the trajectory currently proposed by the robot. We implement our algorithm on two high degree-of-freedom robots, PR2 and Baxter, and present three intuitive mechanisms for providing such incremental feedback. In our experimental evaluation we consider two context rich settings -- household chores and grocery store checkout -- and show that users are able to train the robot with just a few feedback (taking only a few minutes). Despite receiving sub-optimal feedback from non-expert users, our algorithm enjoys theoretical bounds on regret that match the asymptotic rates of optimal trajectory algorithms. Speakers: Joshua Moore, Cornell University Host: Thorsten Joachims Title: Taste over Time: the Temporal Dynamics of User Preferences Abstract: We develop temporal embedding models for exploring how listening preferences of a population develop over time. In particular, we propose time-dynamic probabilistic embedding models that incorporate users and songs in a joint Euclidian space in which they gradually change position over time. Using large-scale Scrobbler data from Last.fm spanning a period of 8 years, our models generate trajectories of how user tastes changed over time, how artists developed, and how songs move in the embedded space. This ability to visualize and quantify listening preferences of a large population of people over a multi-year time period provides exciting opportunities for data-driven exploration of musicological trends and patterns. “The AI-Seminar is sponsored by Yahoo!”
 November 22th Speakers: NO SEMINAR- ACSU LUNCH Host: Bio: Title: Abstract: “The AI-Seminar is sponsored by Yahoo!”
 November 29th Speakers: NO SEMINAR- THANKSGIVING BREAK Host: Bio: Title: Abstract: “The AI-Seminar is sponsored by Yahoo!”
 December 6th Speakers: Adam Bjorndahl, Cornell University Host: Joe Halpern Title: Language-based games Abstract: We introduce a generalization of classical game theory wherein each player has a fixed “language of preference”: a player can prefer one state of the world to another iff they can describe the difference between the two in this language. The expressiveness of the language therefore plays a crucial role in determining the parameters of the game. By choosing appropriately rich languages, this framework can capture classical games as well as various generalizations thereof (e.g. psychological games, reference-dependent preferences, and Bayesian games). On the other hand, coarseness in the language---cases where there are fewer descriptions than there are actual differences to describe---offers insight into some long-standing puzzles of human decision-making. The Allais paradox, for instance, can be resolved simply and intuitively using a language with coarse beliefs: that is, by assuming that probabilities are represented not on a continuum, but discretely, using finitely-many “levels” of likelihood (e.g. ”no chance”, “slight chance”, “unlikely”, “likely”, etc.). Many standard solution concepts from classical game theory can be imported into the language-based framework by taking their /epistemic characterizations/ as definitional. In this way, we obtain natural generalizations of Nash equilibrium, correlated equilibrium, and rationalizability. We show that there are language-based games that admit no Nash equilibria using a simple example where one player wishes to surprise her opponent. By contrast, the existence of rationalizable strategies can be proved under mild conditions. This is joint work with Joe Halpern and Rafael Pass. “The AI-Seminar is sponsored by Yahoo!”

CS7790, Fall '13
Claire Cardie
Carla Gomes
Joe Halpern
Dan Huttenlocher
Thorsten Joachims
Lillian Lee
Ashutosh Saxena
Bart Selman
Ramin Zabih

Back to CS course websites