Artificial Intelligence Seminar

Speaker: Chenhao Tan, Cornell University

Host: Lillian Lee

Title: Online social interactions: a lens on humans and a world for humans

Abstract: Online social interactions have become an integral part of people's lives, e.g., presidential candidates use Facebook and Twitter to engage with the public, programmers rely on Stackoverflow to write code, and various communities have been forming online. This unprecedented amount of social interaction offers tremendous opportunities to understand human behavior. Such an understanding can induce significant social impact, ranging from influencing election outcomes to better communication for everyone. 

My research leverages newly available massive datasets of social interactions to understand human behavior and predict human decisions. These results can be used to build or improve socio-technical systems. In this talk, I will explain my research at both micro and macro levels. At the micro level, I investigate the effect of wording in message sharing via natural experiments. I develop a classifier that outperforms humans in predicting which tweet will be retweeted more. At the macro level, I examine how users engage with multiple communities and find that surprisingly, users continually explore new communities on Reddit. Moreover, their exploration patterns in their early ``life'' can predict whether they will eventually abandon Reddit. I will finish with some discussion of future research directions in understanding human behavior.

Bio: Chenhao Tan is a Ph.D. Candidate in the Department of Computer Science at Cornell University. He earned Bachelor degrees in Computer Science and in Economics from Tsinghua University. His research spans a wide range of topics in social computing. He has published papers primarily at ACL and WWW, and also at KDD, WSDM, ICWSM, etc. His work has been covered by many news media outlets, such as the New York Times and the Washington Post. He also won a Facebook fellowship and a Yahoo! Key Scientific Challenges award.

“The AI-Seminar is sponsored by Yahoo!”

Speaker: Ashesh Jain, Cornell University

Host: Ross Knepper

Title: : Deep Learning on Spatio-Temporal Graphs

Abstract: Deep learning applications as we see today, such as image labeling, captioning, machine translation etc., are largely motivated by the 'web' or the 'internet' use cases. Modeling high-level reasoning's with deep neural networks is still something we don't understand quite well. On the other hand, the real-world around us involves complex interactions that span over space and time, and efficiently modeling them is core to enable robots, self-driving cars in traffic conditions, and to handle a variety of spatio-temporal applications. In this talk, I will first address a spatio-temporal application for assistive cars, and will present a sensory fusion deep learning architecture. I will then present a generic framework for combining the high-level reasoning's expressed over spatio-temporal graphs with the learning success of recurrent neural networks. I will show that the same framework can be applied to motion capture data, to assistive cars, and to robots for understanding human-object interactions. 

Bio:

“The AI-Seminar is sponsored by Yahoo!”

Speaker: Kilian Quirin Weinberger, Cornell University

Title: “Deep Manifold Traversal”

Abstract: Machine learning is increasingly used in high impact applications such as prediction of hospital re-admission, cancer screening or bio-medical research applications. As predictions become increasingly accurate, practitioners may be interested in identifying actionable changes to inputs in order to alter their class membership. For example, a doctor might want to know what changes to a patient's status would predict him/her to not be re-admitted to the hospital soon. Szegedy et al.
demonstrated in 2013 that identifying such changes can be very hard in image classification tasks. In fact, tiny, imperceptible changes can result in completely different predictions without any change to the true class label of the input. In this paper we ask the question if we can make small but meaningful changes in order to truly alter the class membership of images from a source class to a target class. To this end we propose deep manifold traversal, a method that learns the manifold of natural images and provides an effective mechanism to move images from one area (dominated by the source class) to another (dominated by the target class).The resulting algorithm is surprisingly effective and versatile. It allows unrestricted movements along the image manifold and only requires few images from source and target to identify meaningful changes. We demonstrate that the exact same procedure can be used to change an individual's appearance of age, facial expressions or even recolor black and white images.

“The AI-Seminar is sponsored by Yahoo!”

February 26th

Speaker: Ves Stoyanov, Facebook

Host: Claire Cardie

Title: Query Understanding for Facebook Search

Abstract: At Facebook, we strive to build a Search product that allows you to draw on the wisdom of a billion people. Producing high quality search results for multiple intents requires deeper understanding of the query. Our query understanding effort consists of several tasks such as segmenting the query, predicting its intent, recognizing entities mentioned in the query and their relations and producing a semantic parse for structured queries. While none of these problems is completely new, the characteristics of Search on Facebook add some original flavors: our searches are highly personalized and our query distribution is tail-heavy. I will describe some of our recent work on query understanding that improves accuracy by performing joint segmentation and intent prediction. 

“The AI-Seminar is sponsored by Yahoo!”

Speaker: Yexiang Xue, Cornell University

Host: Carla Gomes

Title: Combining Learning and Reasoning for Decision Making: Integrating Concepts from AI, Economics, Crowd-sourcing, and Sustainability

Abstract: Systems for decision-making under uncertainty generally require a tight integration of learning and reasoning techniques. In my research, I'm interested in developing strategies for such an integration in the context of a range of applications in computational sustainability. As an example, I will discuss​ an application for  data collection by citizen scientists to study bird migration patterns. Citizen science projects have been very successful at collecting rich datasets across different domains. However, the data collected by the citizen scientists are often biased, aligned more directly with the participants’ preferences rather than scientific objectives. We introduce a general methodology to improve the scientific quality of the data collected by citizen scientists. Our approach uses incentives to shift ​the interests of citizen scientists to be more aligned with the goal of obtaining unbiased samples from the field, thus improving the quality of the data collected. We formulate the problem as a two-stage game, which requires an integration of learning, to obtain the parameters that govern the individual behavior of the citizen scientists (the agents), with reasoning, to search for an optimal incentive allocation to achieve the goal of the organizer of the citizen science program. We apply our methodology to eBird, a well-established citizen science program of the Cornell Lab of Ornithology for the collection of  bird observations, as a gamified web application, called Avicaching. Our field results show that our Avicaching incentives are remarkably effective at steering the bird watchers' efforts to explore under-sampled areas and hence alleviate the data bias problem in eBird.

This is joint work with Ian Davies, Daniel Fink, Christopher Wood, Ronan Le Bras, Carla P. Gomes and Bart Selman.

“The AI-Seminar is sponsored by Yahoo!”

Speaker: Milind Tambe, USC

Host: Carla Gomes

Title: Security Games: Key Algorithmic Principles, Deployed Applications and Research Challenges

Abstract: Security is a critical concern around the world, whether it is the challenge of protecting ports, airports and other critical infrastructure, protecting endangered wildlife, forests and fisheries, suppressing urban crime or security in cyberspace. Unfortunately, limited security resources prevent full security coverage at all times; instead, we must optimize the use of limited security resources. To that end, our "security games" framework -- based on computational game theory, while also incorporating elements of human behavior modeling, AI planning under uncertainty and machine learning -- has led to building and deployment of decision aids for security agencies in the US and around the world.  These decision aids are in use by agencies such as the US Coast Guard,  the Federal Air Marshals Service and  by various police agencies at university campuses, airports and metro trains. Moreover, recent work on "green security games" has led  our decision aids to begin assisting NGOs in protection of wildlife; and "opportunistic crime security games" have focused on suppressing urban crime. I will discuss our use-inspired research in security games that is leading to new research challenges, including algorithms for scaling up security games as well as for handling significant adversarial uncertainty and learning models of human adversary behaviors.

(*) Joint work with a number of current and former PhD students, postdocs all listed at teamcore.usc.edu/security

Bio: Milind Tambe is Helen N. and Emmett H. Jones Professor in Engineering at the University of Southern California(USC). He is a fellow of AAAI and ACM, as well as recipient of the ACM/SIGART Autonomous Agents Research Award, Christopher Columbus Fellowship Foundation Homeland security award, INFORMS Wagner prize for excellence in Operations Research practice, Rist Prize of the Military Operations Research Society, IBM Faculty Award, Okawa foundation faculty research award, RoboCup scientific challenge award, and other local awards such as the Orange County Engineering Council Outstanding Project Achievement Award, USC Associates award for creativity in research and USC Viterbi use-inspired research award. Prof. Tambe has contributed several foundational papers in AI in areas such as multiagent teamwork, distributed constraint optimization (DCOP) and security games. For this research, he has received the influential paper award and a dozen best paper awards at conferences such as AAMAS, IJCAI, IAAI and IVA. In addition, Prof. Tambe pioneering real-world deployments of ''security games'' has led him and his team to receive the US Coast Guard Meritorious Team Commendation from the Commandant, US Coast Guard First District's Operational Excellence Award, Certificate of Appreciation from the US Federal Air Marshals Service and special commendation given by LA Airport police from the city of Los Angeles. For his teaching and service, Prof. Tambe has received the USC Steven B. Sample Teaching and Mentoring award and the ACM recognition of service award. He has also co-founded a company based on his research, ARMORWAY, where he serves as the director of research. Prof. Tambe received his Ph.D. from the School of Computer Science at Carnegie Mellon University.

“The AI-Seminar is sponsored by Yahoo!”

Speaker: Joe Halpern, Cornell University

Host: Ross Knepper

Title: Actual Causality: A Survey

Abstract: What does it mean that an event C ``actually caused'' event E? The problem of defining actual causation goes beyond mere philosophical speculation.  For example, in many legal arguments, it is precisely what needs to be established in order to determine responsibility.   (What exactly was the actual cause of the car accident or the medical problem?) The philosophy literature has been struggling with the problem of defining causality since the days of Hume, in the 1700s. Many of the definitions have been couched in terms of unterfactuals. (C is a cause of E if, had C not happened, then E would not have happened.) In 2001, Judea Pearl and I introduced a new definition of actual cause, using Pearl's notion of structural equations to model counterfactuals.  The definition has been revised twice since then, extended to deal with notions like "responsibility" and "blame", and applied in atabases and program verification.  I survey the last 15 years of work here, including joint work with Judea Pearl, Hana Chockler, and Chris Hitchcock.  The talk will be completely self-contained.

“The AI-Seminar is sponsored by Yahoo!”

Speaker: Malte Jung, Cornell University

Host: Ross Knepper

Title: Robots and the Dynamics of Emotions in Teams

Abstract: Over the last decade the idea that robots could become an integral part of teamwork developed from a promising vision into a reality.

Robots support teamwork across a wide range of settings covering search and rescue missions, minimally invasive surgeries, space exploration missions, and manufacturing. Scholars have increasingly explored the ways in which robots influence how work in teams is performed, but that work has primarily focused on task specific aspects of team functioning such as the development of situational awareness, common ground, and task coordination. Robots, however, can affect teamwork not only through the task-specific functions they have been developed to serve but also by affecting a team’s regulation of emotion. In this talk I present empirical findings from several studies that show how theory and methods that were originally developed to understand the role of emotions in marital interactions can help us to not only further our understanding of te!amwork but also to inform how we design robots to improve teamwork through their emotion regulatory behavior.

Bio: Malte Jung is an Assistant Professor in Information Science at Cornell University and the Nancy H. ’62 and Philip M. ’62 Young Sesquicentennial Faculty Fellow. His research focuses on the intersections of teamwork, robots, and emotion. The goal of his research is to inform our basic understanding of robots in work teams as well as to inform how we design technology to support teamwork across a wide range of settings. Malte Jung received his Ph.D. in Mechanical Engineering and his Ph.D. minor in Psychology from Stanford University. Prior to joining Cornell, Malte Jung completed a postdoc at the Center for Work, Technology, and Organization at Stanford University. He holds a Diploma in Mechanical Engineering from the Technical University of Munich and an M.S. degree in Mechanical Engineering from Stanford University.

“The AI-Seminar is sponsored by Yahoo!”

Speaker: SPRING BREAK- NO SEMINAR

Host:

Title:

Abstract:

Bio:

Speaker: ACSU LUNCH- NO SEMINAR

Host:

Title:

Abstract:

Bio:

"The AI-Seminar is sponsored by Yahoo!"

Speaker: Adrian Boteanu, Cornell University

Host: Ross Knepper

Title: Using Analogy to Explain Answers to Questions and Object Substitution

Abstract: Reasoning through analogy is a fundamental human reasoning pattern that relies on relational similarity. Understanding how analogies are formed facilitates inference and knowledge transfer. We present a model for relational similarity, the Semantic Similarity Engine (SSE). With a potential proportional analogy as input (A:B::C:D), SSE extracts and compares context-graphs that reveal the relational parallelism that analogies are based on. Its key features are that it provides both numeric and qualitative answers regarding analogical similarity, that it requires no encoding of the inputs, and that it mitigates uncertainty in the semantic network. We show how modeling analogies via SSE can be used in two distinct domains: analogy question interpretation and object substitution in tasks. Proportional analogy questions have been used in standardized tests, answering them requires profound language comprehension and identifying relational correspondence. We combine SSE together with a answer ranking and natural language explanation generation module and evaluate the system on two data-sets totaling 600 analogy questions. Our results show reliable performance and low false-positive rate in question answering; crowdsourced human evaluators agreed with 96\% of our analogy explanations.
Relational correspondence is important not only in linguistic domains, but also in physical environments such as task execution. The relations of an object within a task are multi-faceted, and analogical similarity is important when evaluating object equivalence. Object substitution is the problem of identifying equivalent objects, which complements robot plan repair by extending the planning domain before repair.
We show that integrating SSE in object substitution results in a substitution accuracy of 81\% across nine tasks, and we investigate how analogical similarity can abstract relational invariants for a task. We also demonstrate our work by implementing object substitutions on a physical robot.

“The AI-Seminar is sponsored by Yahoo!”

Speaker: Bistra Dilkina, Georgia Tech

Host: Carla Gomes

Title: Machine Learning for Branch and Bound Search  

Abstract: Branch-and-bound is probably the most widely used method in combinatorial optimization, including mixed integer programming, MAP inference and structured prediction. A lot of research work has focused on improving different components of the branch and bound method, yet often there is no systematic way to effectively make important design choices. We propose the use of machine learning to guide branch and bound search as a novel research direction in order to advance the state of the art in scalable problem solving. As a first step, we develop a novel framework for data-driven, on-the-fly design of variable selection strategies, a key component of the branch and bound method. Using a computationally expensive but effective branching heuristic at the start of the search, the framework fits a variable ranking model based on dynamic variable features that is then applied in the remaining search exploration as a fast approximation. We show an instantiation of this framework using CPLEX, a state-of-the-art MIP solver, and evaluate performance the MIPLIB benchmark. 

In the second part of my talk, I will also discuss my work on optimization approaches to network design for biodiversity conservation planning and for stochastic diffusion processes. While many have focused on optimizing a seed set of nodes from where to start a diffusion cascade, we instead ask the question of how to choose the network structure (add/delete nodes/edges) in order to facilitate or impede a given stochastic process that will affect the designed network. I will briefly present results for both the Independent Cascade and the Linear Threshold Models.

Bio: Bistra Dilkina is an assistant professor in the College of Computing at the Georgia Tech, and a Fellow at the Brook Byers Institute for Sustainable Systems. She received her PhD from Cornell University in 2012, and was a post-doctoral associate at the Institute for Computational Sustainability until 2013. Her research focuses on advancing the state of the art in combinatorial optimization techniques for solving real-world large-scale problems, particularly ones that arise in sustainability areas such as biodiversity conservation planning and urban planning. Her work spans discrete optimization, network design, stochastic optimization, satisfiability, and machine learning. Bistra has (co-)authored over 30 publications, and has won several awards, including Best Paper of the INFORMS ENRE Section, Lockheed Inspirational Young Faculty Award, Raytheon Faculty Fellowship, and Georgia Power Professor of Excellence Award. She is also the co-director of the Atlanta Data Science for Social Good (DSSG) program, which partners multi-disciplinary student teams with government and non-profit organizations to solve some of their most difficult problems using data science techniques such as analytics, modeling, and prediction. 

"If you would like to meet with the speaker, please contact Christianne McMillan White at .... Bistra is available for meetings 9-11:30am and 2:30-6pm on Friday."

“The AI-Seminar is sponsored by Yahoo!”

Speaker: Allison Chaney

Host: David Mimno

Title: Who, What, When, Where, and Why? A Computational Approach to Understanding Historical Events Using State Department Cables

Abstract: We develop computational methods for analyzing historical documents to identify events of potential historical significance. Significant events are characterized by interactions between entities (e.g., countries, organizations, individuals) that deviate from typical interaction patterns. When studying historical events, historians and political scientists commonly read large quantities of text to construct an accurate picture of who, what, when, and where—a necessary precursor to answering the more nuanced question, “Why?” Our methods help historians identify possible events from the texts of historical communication. Specifically, we build on topic modeling to distinguish between topics that describe “business-as-usual” and topics that deviate from these patterns, where deviations are also indicated by particular entities interacting during particular periods of time. To demonstrate our methods, we analyze a corpus of 2 million State Department cables from 1973 to 1977. Joint work with Hanna Wallach and David Blei

“The AI-Seminar is sponsored by Yahoo!”

Speaker: Rahmtin Rotabi

Title: The Status Gradient of Trends in Social Media

Abstract: An active line of research has studied the detection and representation of trends in social media content. There is still relatively little understanding, however, of methods to characterize the early adopters of these trends: who picks up on these trends at different points in time, and what is their role in the system?  We develop a framework for analyzing the population of users who participate in trending topics over the course of these topics' lifecycles.  Central to our analysis is the notion of a status gradient, describing how users of different activity levels adopt a trend at different points in time. Across multiple datasets, we find that this methodology reveals key differences in the nature of the early adopters in different domains.

------------------------------------------------------------------------

Speaker: Jack Hessel

Title: Science, AskScience, and BadScience: On the Coexistence of Highly Related Communities

Abstract: When large social-media platforms allow users to easily form and self-organize into interest groups, highly related communities can arise. For example, the Reddit site hosts not just a group called “food”, but also “HealthyFood”, “foodhacks”, “foodporn”, and “cooking”, among others. Are these highly related communities created for similar classes of reasons (e.g. “true” to distinguish one as a better community and “advice” to focus on helping fellow members)? How do users allocate attention between such close alternatives when they are available or emerge over time? Are there different types of relations between close alternatives such as sharing many users vs. a new community drawing away members of an older one vs. a splinter group failing to cohere into a viable separate community? We investigate the interactions between highly related communities using data from reddit.com consisting of 975M posts and comments spanning an 8-year period. We identify a set of typical affixes that users adopt to create highly related communities and build a taxonomy of affixes. One interesting finding regarding users’ behavior is: after a newer community is created, for several types of highly-related community pairs, users that engage in a newer community tend to be more active in their original community than users that do not explore, even when controlling for previous level of engagement.

-------------------------------------------------------------------------

Speaker: Chenhao Tan

Title: Lost in Propagation? Unfolding News Cycles from the Source

Abstract: The news media play an important role in informing the pub- lic on current events. Yet it has been difficult to understand the comprehensiveness of news media coverage on an event and how the reactions that the coverage evokes may diverge, because this requires identifying the origin of an event and tracing the information all the way to individuals who consume the news. In this work, we pinpoint the information source of an event in the form of a press release and investigate how its news cycle unfolds. We follow the news through three layers of propagation: the news articles covering the press release, shares of those articles in social media, and comments on the shares. We find that a news cycle typically lasts two days. Although news media in aggregate cover the information contained in the source, a single news article will typically only provide partial coverage. Sentiment, while dampened in news coverage relative to the source, again rises in social media shares and comments. As the information propagates through the layers, it tends to diverge from the source: while some ideas emphasized in the source fade, others emerge or gain in importance. We also discover how far the news article is from the information source in terms of sentiment or language does not help predict its popularity.

“The AI-Seminar is sponsored by Yahoo!”