All Seminars

Show:
Title: Mock theta functions and quantum modular forms
Seminar: Algebra
Speaker: Larry Rolen of University of Cologne
Contact: David Zureick-Brown, dzb@mathcs.emory.edu
Date: 2015-03-17 at 4:00PM
Venue: W304
Download Flyer
Abstract:
In this talk, I will describe several related recent results related to mock theta functions, which are functions described by the Indian mathematician Ramanujan shortly before his death in 1920. These functions have very recently been understood in a modern framework thanks to the work of Zwegers and Bruinier-Funke. Here, we will revisit the original writings of Ramanujan and look at his original conception of these functions, which gives rise to a surprising picture connecting important objects such as generating functions in combinatorics and quantum modular forms.
Title: Interactive Machine Learning Across Domains
Colloquium: N/A
Speaker: Lev Reyzin of University of Illinois at Chicago
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2015-03-16 at 4:00PM
Venue: W303
Download Flyer
Abstract:
Interactive learning algorithms have the power to engage with their data and can overcome many limitations of their passive counterparts. In this talk, I will present new algorithms and new models that I have developed for interactive learning. These results include the development of new pool-based, bandit, and query learners. I will also discuss applications and future research challenges for interactive machine learning settings, focusing on the life sciences.
Title: Algebraic Preconditioning of Symmetric Indefinite Systems
Seminar: Numerical Analysis and Scientific Computing
Speaker: Miroslav Tuma of Academy of Sciences of the Czech Republic
Contact: Michele Benzi, benzi@mathcs.emory.edu
Date: 2015-03-10 at 4:00PM
Venue: W301
Download Flyer
Abstract:
Sparse symmetric indefinite linear systems of equations arise in many practical applications. An iterative method is frequently the method of choice to solve such systems but a system transformation called preconditioning is often required for the solver to be effective. In the talk we will deal with development of incomplete factorization algorithms that can be used to compute high quality preconditioners. We will consider both general indefinite systems and saddle-point problems. Our approach is based on the recently adopted limited memory approach (based on the work of Tismenetsky, 1991) that generalizes recent work on incomplete Cholesky factorization preconditioners. A number of new ideas are proposed with the goal of improving the stability, robustness and efficiency of the resulting preconditioner. For general indefinite systems, these include the monitoring of stability as the factorization proceeds and the use of pivot modifications when potential instability is observed. Numerical experiments involving test problems arising from a range of real-world applications are used to demonstrate the effectiveness of our approach and comparisons are made with a state-of-the-art sparse direct solver. The talk will be based on joint work with Jennifer Scott, Rutherford Appleton Laboratory.
Title: Zero-cycles and rational points on rationally connected varieties, after Harpaz and Wittenberg
Seminar: Algebra
Speaker: Jean-Louis Colliot-Thelene of Universite Paris-Sud
Contact: David Zureick-Brown, dzb@mathcs.emory.edu
Date: 2015-03-03 at 4:00PM
Venue: W304
Download Flyer
Abstract:
Harpaz and Wittenberg have recently proved a general result on the local-global principle for zero-cycles on rationally connected varieties. There is also a conditional variant for rational points. I shall explain some of the ideas in their paper. Reference : http://arxiv.org/abs/1409.0993
Title: Extracting medically interpretable concepts from complex health data
Colloquium: N/A
Speaker: Joyce Ho of University of Texas at Austin
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2015-03-03 at 4:00PM
Venue: W303
Download Flyer
Abstract:
Electronic health records (EHRs) are an increasingly important source of patient information. Efficient analysis of EHRs can help address many healthcare problems by improving clinical decisions, facilitating knowledge discoveries, and enabling the development of cost-effective treatment and management programs. However, EHRs pose many formidable challenges for traditional analytics. The data are collected across diverse populations, consist of heterogeneous and noisy information, and have varying time resolutions. Moreover, healthcare professionals are unaccustomed to interpreting high-dimensional EHRs; they prefer concise medical concepts. Thus, a major question is how to transform EHR into meaningful concepts with modest levels of expert guidance.\\ \\ In this talk, I will discuss two approaches to extract concise, meaningful concepts from certain types of health datasets. First, I will describe a dynamic time series model that tracks a patient's cardiac arrest risk based on physiological measurements (i.e., heart rate, blood pressure, etc.) in an intensive care unit. Our algorithm is inspired by financial econometric and yields interpretability and predictability of a cardiac arrest event. Next, I will present a sparse, nonnegative tensor factorization model to obtain multiple medical concepts with minimal human supervision. Tensor factorization utilizes information in the multiway structure to derive concise latent factors even with limited observations. We applied tensor analysis to real EHRs from the Geisinger Health System to automatically identify relevant medical concepts. Both our models are powerful and data-driven approaches to extract medically interpretable concepts from complex health data.
Title: Understanding Information: From Bits to Brains
Colloquium: N/A
Speaker: Avani P. Wildani nee Gadani of The Salk Institute
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2015-03-02 at 4:00PM
Venue: W303
Download Flyer
Abstract:
Information is the currency of the modern era, and there are surprising similarities in data processing and representation between computer systems and neuroscience. In the first half of this talk, I will discuss how to dynamically identify related blocks or files in trace data and use the resulting data groups to make information storage more efficient and robust. From there, I will discuss how the classical systems metrics of reliability, performance, and availability apply to biologically plausible neural networks, including recent work exploring the balance between classification accuracy and robustness. Finally, I will show how computational models from vision can be applied to understand information flow in the visual cortex, and how algebraic topology is a promising method to classify neurons by network function and categorize visual stimuli.
Title: Novel Geometric Algorithms for Machine Learning Problems
Colloquium: N/A
Speaker: Hu Ding of State University of New York at Buffalo
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2015-02-27 at 3:00PM
Venue: W303
Download Flyer
Abstract:
Machine learning is a discipline that concerns the construction and study of algorithms for learning from data, and plays a critical role in many other fields, such as computer vision, speech recognition, social network, bioinformatics, etc. As the data scale increases dramatically in the big-data era, a number of new challenges arise, which require new ideas from other areas. In this talk, I will show that such challenges in a number of fundamental machine learning problems can be resolved by exploiting their geometric properties. Particularly, I will present three geometric-algorithm-based results for various machine learning problems: (1) a unified framework for a class of constrained clustering problems in high dimensional space; (2) a combinatorial algorithm for support vector machine (SVM) with outliers; and (3) algorithms for extracting chromosome association patterns from a population of cells. The first two results are for fundamental problems in machine learning, and the last one is for studying the organization and dynamics of the cell nucleus, an important problem in cell biology. Some geometric-algorithm-based future work in machine learning will also be discussed.
Title: The Foxby-morphism and derived equivalences
Seminar: Algebra
Speaker: Satya Mandal of University of Kansas
Contact: David Zureick-Brown, dzb@mathcs.emory.edu
Date: 2015-02-26 at 4:00PM
Venue: W306
Download Flyer
Abstract:
Suppose $X$ is a quasi-projective scheme over a noetherian (Cohen-Macaulay) affine scheme $Spec(A)$, with $dim X=d$. In $K$-theory and related areas (Witt theory, Grothendieck-Witt theory), bounded chain complexes $G_{\bullet}$ of Coherent sheaves or locally free sheaves play an important role. One considers the category $Ch^b(Coh(X))$ (resp. $Ch^b(V(X))$) of bounded chain complexes of coherent sheaves (resp. of locally free sheaves). One also considers, the corresponding derived categories $D^b(Coh(X)$, $D^b(V(X))$, which is obtained by inverting the quasi-isomorphisms in the chain complex categories. \vspace{4pt} Given a chain complex map $L_{\bullet}\to G_{\bullet}$, between two complexes $L_{\bullet}$, $G_{\bullet}$, with extra information on homologies, one complex can be viewed as \emph{an approximation to} the other. Given one such complex $G_{\bullet}$, constructing such a complex $L_{\bullet}$, with desired properties, and constructing a map $L_{\bullet}\to G_{\bullet}$ would be challenging. In the affine case $X=Spec(A)$, such a map was constructed by Hans-Bjorn Foxby (unpublished), several other versions of the same was given by others. In this lecture we implement the construction of Foxby to quasi-affine case and give applications. Intuitively, one can look at this implementation as a "graded" version of Foxby's construction.
Title: Scalable Big Graph Data Processing
Colloquium: N/A
Speaker: Kisung Lee of Georgia Institute of Technology
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2015-02-26 at 4:00PM
Venue: W303
Download Flyer
Abstract:
The application of graph data analytics is virtually unlimited because graph data are everywhere, from the friendship graphs of social networks to networks of the human brain. Even though graph data analytics is essential for gaining insight into big graphs, large-scale graph processing is complex because of its graph-specific challenges, including complicated correlations among data entities, highly skewed distribution, various graph operations, and the sheer enormity of graph data. This presentation will focus specifically on two new distributed systems for the scalable processing of big graph data. I will first present a graph system that efficiently supports graph pattern query processing (subgraph matching) by scalable graph partitioning and efficient distributed query processing. I will then describe a distributed system for iterative graph computations that can reduce memory requirements for running iterative graph algorithms while ensuring competitive performance. I will conclude the presentation by introducing a set of challenges for developing a general purpose graph analytics system that can support both efficient graph query processing and fast iterative graph computations under one unified system architecture.\\ \\ Bio: Kisung Lee is a Ph.D. candidate in the School of Computer Science at Georgia Tech. His research interests lie in the intersection of big data systems and distributed computing systems. Kisung has also worked on research problems in spatial data management and social network analytics. He has been a recipient of the best paper awards of IEEE Cloud 2012 and MobiQuitous 2014. Kisung received his B.S. and M.S. degrees in computer science from KAIST and has served as a reviewer for several conferences and journals.
Title: High Performance Spatial and Spatio-temporal Data Management
Colloquium: N/A
Speaker: Suprio Ray of University of Toronto
Contact: Vaidy Sunderam, vss@emory.edu
Date: 2015-02-24 at 4:15PM
Venue: W303
Download Flyer
Abstract:
The rapid growth of spatial data volume and technological trends in storage capacity and processing power are fuelling many emerging spatial and spatio-temporal applications from a wide range of domains. Spatial join is widely used by many of the emerging spatial data analysis applications. However, spatial join processing on even a moderate sized dataset is very time consuming. At the same time, there is a rapid expansion in available processing cores, through multicore machines and Cloud computing. The confluence of these trends points to a need for effective parallelization of spatial query processing. Unfortunately, traditional parallel spatial databases are ill-equipped to deal with the performance heterogeneity that is common in the Cloud. In this talk I present two systems that I developed to parallelize spatial join queries in the Cloud and in a large main memory multicore machine.\\ \\ With the proliferation of GPS-enabled mobile devices and sensors, Location Based Services (LBS) have become the most prominent among the spatio-temporal applications. These applications are characterized by high rate of location updates and many concurrent short running range queries. With a large number of devices, the rate of location update can easily surpass 1 million or more updates per second. Traditional relational databases may not be well-suited for this. As the era of "Internet of things" approaches, this issue is expected to get accentuated. In my talk I present a parallel in-memory spatio-temporal indexing technique to support the demands of LBS workloads. Our system achieves significantly better performance than existing approaches to indexing spatio-temporal data. I also present a parallel in-memory spatio-temporal topological join approach. Finally, I outline some of my ideas for future research.