Artificial Intelligence Seminar

Spring 2015
Friday 12:00-1:15
Gates Hall 122

 

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.

 

February 6th

Speaker: Josh Moore, Cornell University

Host: Thorsten Joachims

Bio: Joshua Moore is a fifth-year PhD student in the Department of Computer Science at Cornell University, advised by Thorsten Joachims. He obtained his BS in Computer Science and BS in Applied Mathematics from the Georgia Institute of Technology, and he is a National Science Foundation Graduate Research Fellow. His work centers around the use of embedding methods for modeling and data analysis tasks, especially in the area of Music Information Retrieval and the domain of user listening behavior. He was the recipient of the Best Student Paper Award at ISMIR 2014 for his work on analysis of geography and user listening behavior.

Title: Embedding Methods for Generative Modeling and Visual Data Analysis

Abstract: In this talk, we develop probabilistic embedding models for generative modeling and visual data analysis tasks. Embedding models are a class of models that assume the existence of a latent vector space representation for each of the modeled objects (where an object could be a word, a user, or a song or movie, depending on the task). In our models, we assume that this vector space defines a probability distribution over user choices or recommendations given a context of recent behavior. Furthermore, the scoring function which links the space to the goodness of fit between a choice and a context is chosen to focus on Euclidean distances in the embedding space. This allows us to build modular, interpretable models for a number of tasks. Although these models are generally applicable to many kinds of data, we mainly explore our applications of the models to tasks in the domain of Music Information Retrieval (MIR). First, we consider a playlist generation task which uses historical playlist logs to produce a model of good playlists, as well as a semantic genre space of the modeled songs. Then, we explore adding side information to the model by incorporating social tags to generalize to songs not seen in training, or out-of-vocabulary songs. In more recent work, we combine the interpretability of the semantic spaces of objects that result from our models with an extension to incorporate time dynamics. By training this temporal model on eight years of user listening histories from Last.fm and inspecting the resulting model, we can perform visual and spatial analysis of temporal data to detect influential trends and events in the data. Finally, in work that won the best student paper award at ISMIR 2014, we apply our model to geo-tagged tweets about song plays, creating a space where cities are near their preferred artists. As a result, we can analyze the space of cities in our model to detect geographical, cultural, and linguistic patterns in music listening behavior in cities around the world.

“The AI-Seminar is sponsored by Yahoo!”

February 13th

Speaker: Karthik Raman, Cornell University

Host: Thorsten Joachims

Bio:

Title: Exploiting the Complementary Strengths of Man and Machine: Learning from Human (Rational) Behavior

Abstract: Intelligent systems, ranging from internet search engines and online retailers to personal robots and MOOCs, live in a symbiotic relationship with their users - or at least they should. On the one hand, users greatly benefit from the services provided by these systems. On the other hand, these systems can greatly benefit from the world knowledge that users communicate through their interactions with the system. These interactions -- queries, clicks, votes, purchases, answers, demonstrations, etc. -- provide enormous potential for economically and autonomously optimizing these systems and for gaining unprecedented amounts of world knowledge required to solve some of the hardest AI problems.

In this talk I discuss the challenges of learning from data that results from human behavior. I will present new machine learning models and algorithms that explicitly account for the human decision making process and underlying factors such as human expertise, skills and needs. The talk will also explore how we can look to optimize human interactions to build robust learning systems with provable performance guarantees. I will also present examples, from the domains of search, recommendation and educational analytics, where we have successfully deployed systems for cost-effectively learning with humans in the loop.

“The AI-Seminar is sponsored by Yahoo!”

February 20th

Speaker: Lu Wang, Cornell University

Host: Thorsten Joachims

Bio:

Title: Natural Language Processing for Understanding Socially-Generated Content

Abstract: From simple logs to sophisticated proceedings, text has long been integral to knowledge sharing and discovery. Over the past few decades, technology advances have enabled an explosive growth of text data, far outpacing human beings' speed of understanding text. In order to save human's effort and time, there exists a strong need for computer algorithms that improve our ability to find, absorb, and extract information as we need.

In this talk, I will present my Ph.D. work on how to use Natural Language Processing (NLP) techniques to (1) address users' information needs, and (2) analyze online social interactions. I will first show an abstract generation system that can automatically produce summaries of the essential output from multi-party meetings. I will also present a socially-informed timeline generation system, which can automatically connect the relevant entities and stories for complex events. It produces coherent summaries that consist of information from both traditional news media and social contexts such as user comments. Finally, I will describe future directions of my work, which include developing robust NLP algorithms for large-scale data from domains of social importance. In particular, I am interested in domain-specific language understanding and generation techniques for computational social science. I also plan to discuss the potential of NLP in health informatics by bridging the gap between sophisticated medical knowledge and ordinary users' information need.

“The AI-Seminar is sponsored by Yahoo!”

February 27th

Speaker: Vasumathi Raman

Host:

Bio: Vasu Raman is a postdoctoral scholar in the Department of Computing and Mathematical Sciences at the California Institute of Technology, Pasadena, CA. Her research explores algorithmic methods for designing and controlling autonomous cyber-physical systems, guaranteeing correctness with respect to user-defined specifications. She earned a PhD in Computer Science from Cornell University and a BA in Computer Science and Mathematics from Wellesley College.

Title: Synthesis for Autonomy: Connecting the Physical, the Logical and the Human

Abstract:
Autonomous cyber-physical systems like robots and self-driving cars stretch the limits of engineering, and reasoning precisely about their behavior is increasingly difficult. Moreover, a correct implementation is often elusive even when the desired outcome is obvious. Algorithmic synthesis makes the formal specification of desired behavior an integral part of the controller design process -- the resulting controllers are correct by construction. Synthesis enables non-expert human users to provide complex system specifications, and be assured of their fulfillment. However, this goal of automatically generating autonomous control from user-defined specifications poses a host of unprecedented challenges.

In this talk, I will present two frameworks for automated synthesis from high-level specifications. I will first describe a system for generating, troubleshooting, and executing controllers for autonomous robots from natural language specifications, including analytical tools for automatically determining causes of failure to synthesize. I will then introduce a novel approach to synthesis for high-dimensional systems operating in uncertain environments, bypassing the need for discretizing the state space and admitting specifications that govern not just the order of events, but also their timing. Finally, I will present new research directions that push the envelope of synthesis for autonomous systems.

“The AI-Seminar is sponsored by Yahoo!”

March 6th

Speaker: Dan Weld, University of Washington

Host: Bart Selman

Bio: Daniel S. Weld is Thomas J. Cable / WRF Professor of Computer Science and Engineering at the University of Washington. Weld received a BS from Yale in 1982, a PhD from MIT in 1988, a Presidential Young Investigator’s award in 1989, an Office of Naval Research Young Investigator’s award in 1990, was named AAAI Fellow in 1999, and ACM Fellow in 2006. Weld is also an active entrepreneur with several patents and technology licenses. In May 1996, he co-founded Netbot, creator of Jango Shopping Search, later acquired by Excite. In October 1998, Weld co-founded AdRelevance, a monitoring service for internet advertising, which was acquired by Media Metrix and is now operated by Nielsen NetRatings. In June 1999, Weld co-founded data integration company Nimble Technology which was acquired by the Actuate Corporation. In January 2001, Weld joined the Madrona Venture Group as a Venture Partner and member of the Technical Advisory Board.

Title: Planning to Control Crowd-Sourced Workflows

Abstract: Crowd-sourcing labor markets (e.g., Amazon Mechanical Turk) are booming, because they enable rapid construction of complex workflows that seamlessly mix human computation with computer automation. Example applications range from photo tagging to audio-visual transcription and interlingual translation. Unfortunately, constructing a good workflow is difficult, because the quality of the work performed by humans is highly variable. Typically, a task designer will experiment with several alternative workflows to accomplish a task, varying the amount of redundant labor, until she devises a control strategy which delivers acceptable performance. Fortunately, this control challenge can often be formulated as an automated planning problem… ripe for algorithms from the probabilistic planning and reinforcement learning literature. This talk describes our recent work on the decision-theoretic control of crowd sourcing and suggests open problems for future research.

“The AI-Seminar is sponsored by Yahoo!”

March 13th

Speaker: Karthik Sridharan, Cornell

Host:

Bio:

Title: Hierarchies of relaxations for online prediction problems with evolving constraints

Abstract: We study online prediction where regret of the algorithm is measured against a benchmark defined via evolving constraints. This framework captures online prediction on graphs, as well as other prediction problems with combinatorial structure. A key aspect here is that finding the optimal benchmark predictor (even in hindsight, given all the data) might be computationally hard due to the combinatorial nature of the constraints. Despite this, we provide polynomial-time prediction algorithms that achieve low regret against combinatorial benchmark sets. We do so by building improper learning algorithms based on two ideas that work together. The first is to alleviate part of the computational burden through random playout, and the second is to employ Lasserre semidefinite hierarchies to approximate the resulting integer program. Interestingly, for our prediction algorithms, we only need to compute the values of the semidefinite programs and not the rounded solutions. However, the integrality gap for Lasserre hierarchy does enter the generic regret bound in terms of Rademacher complexity of the benchmark set. This establishes a trade-off between the computation time and the regret bound of the algorithm.

“The AI-Seminar is sponsored by Yahoo!”

March 20th

Speaker: Gerard Biau

Host: Giles Hooker

Bio:

Title: Distributed statistical algorithms

Abstract: Distributed computing offers a high degree of  flexibility to accommodate modern learning constraints and the ever increasing size of datasets  involved in massive data issues.  Drawing inspiration from the theory of distributed computation models developed in the context of gradient-type optimization algorithms, I will present a consensus-based asynchronous distributed approach for nonparametric online regression and analyze some of its asymptotic properties. Substantial numerical evidence involving up to 28 parallel processors is provided on synthetic datasets to assess the excellent performance of the method, both in terms of computation time and prediction accuracy.

“The AI-Seminar is sponsored by Yahoo!”

March 27th

Speaker: Bishan Yang

Host:

Bio: Bishan Yang is a Ph.D. candidate at Cornell University. Her research interests lie in natural language processing and machine learning. She has been working on developing effective machine learning methods for information extraction problems. She earned her M.S. and B.S. degrees in Computer Science from Peking University. During her graduate studies, she has interned in the machine learning group at Microsoft Research, the NLP group at Google Research, and eBay Research Labs.

Title: Extracting Information from Text by Exploiting Semantic Structure

Abstract: A lot of information that is useful to people in their daily life is in the form of text, for example, emails, news, books and social media. In the age of information overload, we need computers that can automatically extract important information from large amounts of text according to our needs. My research aims at developing statistical machine learning methods to tackle the key problems in information extraction: How to accurately identify text descriptions that carry the information of interest? How to interpret their meanings in context? How to aggregate information from a large number of text documents?

In this talk I will focus on information extraction of opinions and events. The richness and complexity of natural language make the task difficult. Existing machine learning approaches usually provide limited performance on the task due to shallow natural language understanding. To address this problem, I develop learning and inference models that can exploit rich semantic structure of text, e.g. predicate-argument structure, discourse structure, and coreference structure, leading to promising improvements of the information extraction system. I will present three approaches: a joint inference approach that can simultaneously extract opinion expression, sources, targets and their relations; a structured learning approach that is capable of inferring sentiment based on context; and a distance-dependent Bayesian model for aggregating event information in news articles.

“The AI-Seminar is sponsored by Yahoo!”

April 3rd

Speaker: Brad Hayes, Yale

Host: Ross Knepper

Bio: Brad Hayes is a PhD candidate and a member of the Social Robotics Laboratory in the Department of Computer Science at Yale University, advised by Professor Brian Scassellati. Brad is interested in building robotic systems that are capable of safe, complex collaboration with humans who may not have programming or robotics expertise. His work involves a combination of learning from demonstration, human teaming psychology, intention recognition and projection, and human-robot interaction.

Title: Supportive Behaviors for Human-Robot Teaming

Abstract: Robots capable of collaborating with people provide tremendous value, bringing with them the potential to revolutionize a wide array of industries ranging from healthcare to education to manufacturing. Particularly in domains where modern robots are ineffective, human-robot teaming can be leveraged to increase the efficiency, capability, and safety of people. Central to building these autonomous systems are the problems of teammate goal inference and multi-agent coordination, both of which can be extremely challenging without a priori task knowledge or behavioral models. In this talk I will cover my recent work in developing robots that can learn and generate what we term 'supportive behaviors': off-goal actions a teammate can perform that facilitate task completion. 

“The AI-Seminar is sponsored by Yahoo!”

April 10th

Speaker: NO SEMINAR- ACSU Lunch

Host:

Bio:

Title:

Abstract:

“The AI-Seminar is sponsored by Yahoo!”

April 17th

Speaker: Ryan Adams, Harvard

Host: Peter Frazier

Bio: Ryan Adams is an Assistant Professor of Computer Science at Harvard.

Title: Implementing Probabilistic Graphical Models with Chemical Reaction Networks

Abstract: Recent work on molecular programming has explored new possibilities for computational abstractions with biomolecules, including logic gates, neural networks, and linear systems.  In the future such abstractions might enable nanoscale devices that can sense and control the world at a molecular scale.  Just as in macroscale robotics, it is critical that such devices can learn about their environment and reason under uncertainty. At this small scale, systems are often modeled as chemical reaction networks.  I will describe a procedure by which arbitrary probabilistic graphical models, represented as factor graphs over discrete random variables, can be compiled into chemical reaction networks that implement inference.  I will show how marginalization based on sum-product message passing can be implemented in terms of reactions between chemical species whose concentrations represent probabilities.  The steady state concentrations of these species correspond to the marginal distributions of the random variables in the graph.  As with standard sum-product inference, this procedure yields exact results for tree-structured graphs, and approximate solutions for loopy graphs.

This is joint work with Nils Napp.

“The AI-Seminar is sponsored by Yahoo!”

April 24th

Speaker: Edo Liberty, Yahoo

NOTE: Slides for Edo Liberty's talk are here.

http://www.cs.yale.edu/homes/el327/papers/lowRankMatrixApproximation.pdf

Host: Thorsten Joachims

Bio: Edo Liberty is a Research Director at Yahoo Labs and manages Yahoo's Scalable Machine Learning group. He received his BSc in Physics and in Computer Science from Tel Aviv university and his PhD in Computer Science from Yale university, where he also held a postdoctoral position in the Program in Applied Mathematics. Edo was also a co-founder and the CTO of Cognitive-Networks prior to joining Yahoo in 2010.

Title: Low rank matrix approximation offline, in streams, and online.

Abstract: Computing low rank approximations of data matrices is useful for signal processing, regression, clustering, dimension reduction and many other computational tasks. Yet, classical algorithms for doing that often fall short in modern settings. Modern data matrices are often too large to store and may only become available over time. This talk will describe new techniques for computing low rank approximations to matrices in the offline, streaming and online computational models.

“The AI-Seminar is sponsored by Yahoo!”

April 24th
(additional seminar at 2pm)

Speaker: Henny Admoni, Yale

Host: Ross Knepper

Bio: Henny Admoni is a final-year PhD candidate in Computer Science at the Social Robotics Laboratory at Yale University. Henny creates and studies intelligent, autonomous robots that help make people’s lives better by assisting them in social environments like homes and offices. Her research is about how to use nonverbal communication, such as eye gaze and pointing, to make human-robot interactions more natural and effective for people. Henny holds an MS in Computer Science from Yale University, and a BA/MA joint degree in Computer Science from Wesleyan University. Henny's scholarship has been recognized with awards such as the NSF Graduate Research Fellowship, the Google Anita Borg Memorial Scholarship, and the Palantir Women in Technology Scholarship.

Title:Nonverbal Communication for Assistive Human-Robot Interaction

Abstract: Autonomous robots are already integrated into human environments: manufacturing robots weld cars together in factories, and robot vacuum cleaners keep millions of homes clean. But these autonomous robots are restricted to acting alone, primarily without human intervention. The inability to interact with people limits robots' usefulness in more personal day-to-day tasks. However, as robot costs decrease, there is increasing potential for robots to enter homes and workspaces as personal tools for individuals, improving people's lives with natural, social interactions. For example, these robots may be tutors for children, collaborative assistants for factory workers, or personal aides for people with motor disabilities. Human environments are inherently social, so to succeed in their tasks, these robots must understand and use the rich social communication structures that exist in these environments. One important structure is nonverbal communication. This encompasses a range of behavior from facial expression to whole-body proxemics. My work is focused on eye gaze and deixis (pointing) as it pertains to communicating attention and enabling collaboration. In this talk, I describe the development of a computational model for nonverbal communication, and the development of a robot behavior controller based on this model which uses eye gaze and pointing to help a user with a collaborative construction task. I also describe the human-robot interaction studies that support this work: what people can teach us about building better robots, and what robots can teach us about people and communication.

“The AI-Seminar is sponsored by Yahoo!”

May 1st

Speaker: Ariel Procaccia, Computer Science Dept., Carnegie Mellon University

Host: Eva Tardos

Bio: Ariel Procaccia is an assistant professor in the Computer Science Department at Carnegie Mellon University.  He usually works on problems at the interface of computer science and economics. Besides computational fair division, some of his current research interests include computational social choice, and its applications to crowdsourcing and human computation; computational game theory, and its applications to physical security and cybersecurity; and computational mechanism design, and its applications to kidney exchange. His distinctions include the IJCAI Computers and Thought Award (2015), the Sloan Research Fellowship (2015), the NSF Early Career Development Award (2014), and the IFAAMAS Victor Lesser Distinguished Dissertation Award (2009).

Title: Computational Fair Division

Abstract: I will present an exciting new interaction between AI, theoretical computer science, and fair division theory, which is leading to some of the first-ever applied fair division methods. In particular, I will explain how computational thinking provides a novel perspective on the classic problem of allocating indivisible goods, and how these ideas are integrated into
Spliddit (http://www.spliddit.org), a not-for-profit fair division website that aims to make the world a bit fairer. I will also describe our ongoing work with a California school district to develop a practicable mechanism for fairly allocating classrooms to charter schools, which has given rise to novel theoretical questions as well as nontrivial computational challenges.

“The AI-Seminar is sponsored by Yahoo!”

May 8th

Speakers: Vlad Niculae & Chenhao Tan

Host:

Bio: TBA

Title:The Structure of Political Media Coverage as Revealed by Quoting Patterns - Speaker: Vlad Niculae

Abstract:Given the extremely large pool of events and stories available, media outlets need to focus on a subset of issues and aspects to convey to their audience. Outlets are often accused of exhibiting a systematic bias in this selection process, with different outlets portraying different versions of reality. However, in the absence of objective measures and empirical evidence, the direction and extent of systematicity remains widely disputed.

We propose a framework based on quoting patterns for quantifying and characterizing the degree to which media outlets exhibit systematic bias.  We apply this framework to a massive dataset of news articles spanning the six years of Obama's presidency and all of his speeches, and reveal that a systematic pattern indeed emerges from the outlet's quoting behavior.  Moreover, we  show that this pattern can be successfully exploited in an unsupervised prediction setting, to determine which new quotes an outlet will select to broadcast.  By encoding bias patterns in a low-rank space we provide an analysis of the structure of political discourse. This reveals a latent media bias space that aligns surprisingly well with political ideology and outlet type.  A linguistic analysis exposes striking differences across these latent dimensions, showing how the different types of media outlets portray different realities even when reporting on the same events.  For example, the outlets mapped to the mainstream conservative side of the latent space focus on president's quotes that portray a presidential persona disproportionately characterized by negativity.

More details are available at http://snap.stanford.edu/quotus/.

This is joint work with Caroline Suen (Stanford), Justine Zhang (Stanford, joining Cornell), Cristian Danescu-Niculescu-Mizil (Cornell) and Jure Leskovec (Stanford).

“The AI-Seminar is sponsored by Yahoo!”

May 8th

Speakers: Vlad Niculae & Chenhao Tan

Host:

Bio: TBA

Title: All Who Wander: On the Prevalence and Characteristics of Multi-community Engagement - Speaker: Chenhao Tan

Abstract: Although analyzing user behavior within individual communities is an active and rich research domain, people usually interact with multiple communities both on-and off-line. How do users act in such multi-community environments? Although there are a host of intriguing aspects to this question, it has received much less attention in the research community, in comparison to the intra-community case. In this paper, we examine three aspects of multi-community engagement: the sequence of communities that users post to, the language that users employ in those communities, and the feedback that users receive, using longitudinal posting behavior on Reddit as our main data source, and DBLP for auxiliary
experiments. We also demonstrate the effectiveness of features drawn from these aspects in predicting users’ future level of activity.

One might expect that a user’s trajectory mimics the “settling-down” process in real life: an initial exploration of sub-communities before settling down into a few niches. However, we find that the users in our data continually post in new
communities; moreover, as time goes on, they post increasingly evenly among a more diverse set of smaller communities. Interestingly, it seems that users that eventually leave the community are “destined” to do so from the very beginning,
in the sense of showing significantly different “wandering” patterns very early on in their trajectories; this finding has potentially important design implications for community maintainers. Our multi-community perspective also allows us to investigate the “situation vs. personality” debate from language usage across different communities.

https://chenhaot.com/pages/multi-community.html

This is joint work with Lillian Lee.

“The AI-Seminar is sponsored by Yahoo!”

June 8th

Speaker: Paul N. Bennett, Microsoft Research

Host: Thorsten Joachims

Bio: Paul Bennett is a Senior 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.  His recent work has been recognized with a SIGIR 2012 Best Paper Honorable Mention and a SIGIR 2013 Best Student Paper award. 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:Events and Controversies: Influences of a Shocking News Event on Information Seeking

Abstract:It has been suggested that online search and retrieval contributes to the intellectual isolation of users within their preexisting ideologies, where people’s prior views are strengthened and alternative viewpoints are infrequently encountered. This so-called “filter bubble” phenomenon has been called out as especially detrimental when it comes to dialog among people on controversial, emotionally charged topics, such as the labeling of genetically modified food, the right to bear arms, the death penalty, and online privacy. We seek to identify and study information-seeking behavior and access to alternative versus reinforcing viewpoints following shocking, emotional, and large-scale news events. We choose for a case study to analyze search and browsing on gun control/rights, a strongly polarizing topic for both citizens and leaders of the United States. We study the period of time preceding and following a mass shooting to understand how its occurrence, follow-on discussions, and debate may have been linked to changes in the patterns of searching and browsing. We employ information-theoretic measures to quantify the diversity of Web domains of interest to users and understand the browsing patterns of users. We use these measures to characterize the influence of news events on these web search and browsing patterns.

This is joint work with Danai Koutra and Eric Horvitz

“The AI-Seminar is sponsored by Yahoo!”

See also the AI graduate study brochure.

Please contact any of the faculty below if you'd like to give a talk this semester. We especially encourage graduate students to sign up!

Sponsored by


CS7790, Spring '15

Erik Andersen
Serge Belongie
Claire Cardie
Tanzeem Choudhury
Cristian Danescu-Niculescu-Mizil
Shimon Edelman
Carla Gomes
Joe Halpern
Haym Hirsh
Dan Huttenlocher
Ross Knepper
Thorsten Joachims
Lillian Lee
David Mimno
Ashutosh Saxena
Bart Selman
Karthik Sridharan
Ramin Zabih

Back to CS course websites