All Seminars

Title: Scalable Probabilistic Inference for Complex Dynamical Models
Colloquium: N/A
Speaker: Lei Li of University of California, Berkeley
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2014-03-07 at 3:00PM
Venue: MSC W303
Download Flyer
Abstract:
Time series data arise in numerous applications, such as data center monitoring, tracking web user activities, health care, etc. Detecting patterns and learning features in collections of data sequences are crucial to solve real-world, domain specific problems, for example, to track moving objects in videos, to spot nefarious online activities, and to forecast patients' health states.\\ \\ In this talk, I define a new tensor dynamical model for multivariate data as well as an efficient algorithm for learning such models from data. In addition, I will present an efficient approach to jointly estimate parameters and latent states for a large class of models including nonlinear dynamical systems. Finally, I will present my work on efficient inference for a probabilistic declarative programming language, which aims to democratize machine learning and to enable practitioners to solve their domain specific problems.\\ \\ Bio: Dr. Lei Li is a Post-Doctoral researcher at EECS department of UC Berkeley. His research interest lies in the intersection of machine learning, statistical inference and database systems. Specifically, he has been working on Bayesian inference in open universe probabilistic models, probabilistic programming language, large-scale learning, time series, communication and social networks. He has served in the Program Committee for ICML 2014, SDM 2013/2014, and IJCAI 2011/2013. He has been invited as reviewer for TOMCCAP, DAMI, TKDE, TOSN, Neurocomputing, KDD, SIGMOD, VLDB, PKDD and WWW. He has been invited to review NSF proposal in 2010 and to DARPA's Information Science and Technology (ISAT) probabilistic programming workshop in 2013.\\ \\ Lei received his B.S. in Computer Science and Engineering from Shanghai Jiao Tong University in 2006 and Ph.D. in Computer Science from Carnegie Mellon University in 2011, respectively. His dissertation work on fast algorithms for mining co-evolving time series was awarded ACM KDD best dissertation (runner up).
Title: Finding a Happy Medium between Accuracy and Speed for Dependency Parsing
Colloquium: N/A
Speaker: Jinho Choi of University of Massachusetts Amherst
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2014-03-06 at 4:00PM
Venue: W306
Download Flyer
Abstract:
Why is Natural Language Processing interesting? What makes NLP hard? How can we bring NLP research to practice? These are all open-ended questions. In this talk, I present a novel approach called selectional branching, which optimizes both accuracy and speed for one of core NLP tasks, dependency parsing. Our approach uses confidence estimates to decide when to employ a beam, providing the accuracy of beam search at speeds close to a greedy dependency parsing approach. Selectional branching is guaranteed to perform faster than beam search yet performs as accurately. With the benchmark setup in English, our parser shows an accuracy of 92.96% and a speed of 9 milliseconds per sentence, which is faster and more accurate than the previous state-of-the-art transition-based parser using beam search. It also outperforms other dependency parsers using beam search, dynamic programming, integer linear programming, etc. for languages such as Danish, Dutch, Slovene, and Swedish.
Title: Towards Large Scale Open Domain Natural Language Processing
Colloquium: N/A
Speaker: Gourab Kundu of University of Illinois
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2014-03-05 at 3:00PM
Venue: MSC W201
Download Flyer
Abstract:
Machine Learning and Inference methods are becoming ubiquitous ñ a broad range of scientific advances and technologies rely on machine learning techniques. In particular, the big data revolution heavily depends on our ability to use statistical machine learning methods to make sense of the large amounts of data we have. Research in Natural Language Processing has both benefited and contributed to the advancement of machine learning and inference methods. However multiple problems still hinder the broad application of some of these methods. Performance Degradation of machine learning based systems in domains other than the training domain is one of the key problems hindering widespread deployment of these systems.\\ \\ In this talk, I will present techniques for domain adaptation "on the fly", that allows adaptation to test domains using the same model from training domain. This is accomplished by transforming text from the test domain to look more like the training domain and running the same model from the training domain. This process of text adaptation treats the model as black box, thus makes the adaptation of complex pipelines of models easy and flexible. The next key challenge for machine learning is the processing of vast amounts of data in an efficient manner. Prediction problems for tools are often complicated, for natural language processing and other disciplines, making application of these tools to big data infeasible. The later part of the talk will focus on improving the scalability of machine learning tools with complex prediction component to meet the challenges of big data. I will show how it is possible to amortize the cost of prediction over the lifetime of any machine learning tool. Particularly, I will focus on amortizing integer linear programs which can represent a wide variety of prediction problems. I will present exact and approximate theorems for speeding up the solution time of new integer programs by reusing solutions of previously solved integer programs.\\ \\ Gourab Kundu is a doctoral candidate in Computer Science Department of University of Illinois at Urbana-Champaign, supervised by Prof. Dan Roth. He has also worked in IBM research and Google for summer internships. He is broadly interested in all aspects of machine learning and natural language processing. He has publications in top tier natural language processing conferences along with a best student paper in CoNLL 2011.
Title: Bounded gaps between primes in Chebotarev sets
Seminar: Algebra
Speaker: Jesse Thorner of Emory University
Contact: David Zureick-Brown, dzb@mathcs.emory.edu
Date: 2014-03-04 at 4:00PM
Venue: W302
Download Flyer
Abstract:
A new and exciting breakthrough due to Maynard establishes that there exist infinitely many pairs of primes $p_1,p_2$ with $|p_1-p_2|\leq 600$ as a consequence of the Bombieri-Vinogradov Theorem. In this paper, we apply their general method to the setting of Chebotarev sets of primes. We study applications of these bounded gaps with an emphasis on ranks of prime quadratic twists of elliptic curves over $\mathbb{Q}$.
Title: Weights and Measures: Fast Prediction in an Era of Big-Data
Colloquium: N/A
Speaker: Lev Reyzin of University of Illinois at Chicago
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2014-03-04 at 4:00PM
Venue: MSC W303
Download Flyer
Abstract:
In this talk I will discuss algorithms I have developed for learning in a world where data is abundant and predicting quickly and accurately is essential. In particular, I will focus on some recent work on modern variants of supervised and bandit learning. One common element of the algorithms I will present is that they nontrivially improve upon classical weighing and sampling methods to produce provable and practical improvements over traditional approaches.
Title: Two-Hilbert spaces Mourre theory for the Laplace-Beltrami operator on manifolds with asymptotically cylindrical ends
Seminar: Analysis and Differential Geometry
Speaker: Rafael Tiedra de Aldecoa of Catholic University of Chile
Contact: David Borthwick, davidb@mathcs.emory.edu
Date: 2014-03-04 at 4:00PM
Venue: MSC W301
Download Flyer
Abstract:
We review some aspects of Mourre theory in a two-Hilbert spaces setting. Then we apply this theory to the spectral analysis for the Laplace-Beltrami operator on manifolds with asymptotically cylindrical ends. This is a joint work with Serge Richard (University of Nagoya).
Title: When Big Data Meets BRAIN Initiative: Large-Scale Structured Sparse Learning with Applications in Imaging Genomics
Colloquium: N/A
Speaker: Heng Huang of University of Texas at Arlington
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2014-03-03 at 4:00PM
Venue: MSC W303
Download Flyer
Abstract:
Sparsity is one of the intrinsic properties of real-world data, thus sparse machine learning has recently emerged as powerful tool to obtain models of high-dimensional data with high degree of interpretability at low computational cost, and provide great opportunities to analyze the big, complex, and diverse datasets. By enforcing properly designed structured sparsity, we can integrate the specific data structures and domain knowledge into the machine learning models to simplify data models and discover predictive patterns in big data analytics. Big Data research is accelerating the translation of biological and biomedical data to advance the detection, diagnosis, treatment and prevention of diseases, including the recently announced BRAIN (Brain Research through Advancing Innovative Neurotechnologies) Initiative. To address the challenging problems in current big data mining, we proposed several novel large-scale structured sparse learning models for multi-dimensional data integration, heterogeneous multi-task learning, group/graph structured data analysis, and longitudinal feature learning. We applied our new structured sparse learning models to analyze the multi-modal neuroimaging and genome-wide array data in Imaging Genomics and discover the phenotypic and genotypic biomarkers to characterize the neurodegenerative process in the progression of Alzheimer’s disease and other complex brain disorders. We also utilized our new machine learning models to analyze the Electronic Medical Records for predicting the heart failure patients’ readmission and drug side effects, detect the multi-dimensional biomarkers in The Cancer Genome Atlas (TCGA) research, and identify the brain circuitry patterns in Human Connectome.
Title: An Integrated Human Decision Making Model under Extended Belief-Desire-Intention Framework: Emergency Evacuation Applications
Seminar: Computer Science
Speaker: Young-Jun Son of The University of Arizona
Contact: TBA
Date: 2014-02-28 at 11:00AM
Venue: MSC W303
Download Flyer
Abstract:
In this talk, we discuss an integrated Belief-Desire-Intention (BDI) modeling framework for human decision making, whose sub-modules are based on Bayesian belief network, Decision-Field-Theory, and probabilistic depth first search technique. A key novelty of the proposed model is its ability to represent both the human decision-making and decision-planning functions in a unified framework. In this talk, the proposed modeling framework is demonstrated for human’s evacuation behaviors under a terrorist bomb attack situation. To mimic realistic human behaviors, attributes of the BDI framework are reverse-engineered from the human-in-the-loop experiments conducted in the Cave Automatic Virtual Environment (CAVE) available at The University of Arizona. A crowd simulation is then constructed, where individual human behaviors are based on what was learned from the CAVE experiments. In this work, the simulated environment and humans conforming to the proposed BDI framework are implemented in AnyLogic® agent-based simulation software, where each human entity calls external Netica BBN software to perform its perceptual processing function and Soar software to perform its real-time planning and decision-execution functions. The constructed crowd simulation is then used to test impact of several factors (e.g. demographics of people, number of policemen, information sharing via speakers) on evacuation performance (e.g. average evacuation time, percentage of casualties). Finally, we discuss other emergency evacuation applications (e.g. evacuation behaviors under fire in a factory) and research extensions for the proposed BDI framework.
Title: Assured Information Distillation in Social Sensing
Colloquium: N/A
Speaker: Dong Wang of University of Illinois at Urbana-Champaign
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2014-02-27 at 4:00PM
Venue: W306
Download Flyer
Abstract:
The advent of sensors and online social broadcast media (e.g., Twitter and Flickr) create a deluge of unfiltered, unstructured, and unvetted data about the physical environment. This opens up unprecedented challenges and opportunities in social sensing, where the goal is to distill assured information from social sources and devices in their possession. This talk will present a new analytical framework and theories to obtain reliable information with quality guarantees from large amounts of unreliable social sensing data. Noticeably, our analytical framework is the first to jointly model the complex interactions among three deeply coupled networks underlying the data; namely, the information, social and physical networks. The talk will also introduce a new information distillation system we built, called Apollo, which has been applied in a wide range of social sensing scenarios such as real event/disaster tracking, geo-tagging, smart road applications, and language/dialect classification. Apollo is now used by different branches at Army Research Lab (ARL).
Title: Solved and unsolved problems in elementary number theory
Seminar: Joint Athens-Atlanta Number Theory Seminar
Speaker: Paul Pollack of UGA
Contact: David Zureick-Brown, dzb@mathcs.emory.edu
Date: 2014-02-25 at 4:00PM
Venue: W302
Download Flyer
Abstract:
This will be a survey of certain easy-to-understand problems in elementary number theory about which "not enough" is known. We will start with a discussion of the infinitude of primes, then discuss the ancient concept of perfect numbers (and related notions), and then branch off into other realms as the spirit of Paul Erd\H{o}s leads us.