All Seminars
| Title: Supporting Information Access through Text Quality Prediction and Automatic Summarization |
|---|
| Seminar: Computer Science |
| Speaker: Annie Louis of University of Edinburgh |
| Contact: Vaidy Sunderam, VSS@emory.edu |
| Date: 2014-03-28 at 3:00PM |
| Venue: MSC W303 |
| Download Flyer |
| Abstract: When users look for information on the world wide web, not only do they seek relevant documents, but also desire information that is of high quality, and is organized, and summarized in an easy to digest format. In this talk, I will overview my work in two areas aiming to create richer search and browsing experiences for users. First, I will describe work on automatically assessing the writing quality of documents. Well-written documents are obviously more valuable to users. While spelling and grammar errors are detected easily by current systems, there are numerous other aspects of writing quality for which computational methods do not exist. I will outline some models I have built to predict coherent, concise and popular writing style. Second I will describe some ongoing work to automatically uncover the structure of and summarize conversations from web discussion forums. Often forums are organized as threads with posts appearing in a simple chronological order. I will describe two models which add more structure to these conversations: one which groups forum participants based on the content and reply patterns of their conversations, and a second model to identify easy and difficult instructions suggested in computer troubleshooting forums. |
| Title: Multilevel Monte Carlo simulations with algebraically constructed coarse spaces |
|---|
| Seminar: Numerical Analysis and Scientific Computing |
| Speaker: U. Villa, P. Vassilevski of Center for Applied Scientific Computing (CASC), Lawrence Livermore National Laboratory (LLNL) |
| Contact: Veneziani, |
| Date: 2014-03-28 at 4:00PM |
| Venue: W306 |
| Download Flyer |
| Abstract: We consider the numerical simulation of multiscale multiphysics phenomena with uncertain input data in a Multilevel Monte Carlo (MLMC) framework. Multilevel Monte Carlo techniques typically rely on the existence of hierarchies of computational meshes obtained by successive refinement. We apply MLMC to unstructured meshes by using specialized element-based agglomeration techniques that allow us to construct hierarchies of coarse spaces that possess stability and approximation properties for wide classes of PDEs. An application to subsurface flow simulation in mixed finite element setting illustrates our approach. This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344 |
| Title: Optimal Central Bank Interventions in the presence of Regime Switching |
|---|
| Seminar: General Colloquium |
| Speaker: Luz Rocio Sotomayor of Georgia State University |
| Contact: Steve Batterson, sb@mathcs.emory.edu |
| Date: 2014-03-27 at 4:00PM |
| Venue: W306 |
| Download Flyer |
| Abstract: In a foreign exchange market that is affected by the conditions of the monetary fundamentals (unemployment, interest rate, inflation rate, trade deficit and GDP, among others), we model the foreign exchange rate as a process with parameters modulated by an observable continuous-time Markov chain. Under this setup, we consider the problem of the domestic Central Bank that has to choose the optimal intervention strategy that minimizes the total intervention cost of keeping the exchange rate as close as possible to a given target rate. We solve the problem by using techniques of stochastic impulse control with regime switching. |
| Title: Deep Models for Gene Regulation |
|---|
| Defense: Dissertation |
| Speaker: Olgert Denas of Emory University |
| Contact: Olgert Denas, odenas@emory.edu |
| Date: 2014-03-27 at 4:00PM |
| Venue: W302 |
| Download Flyer |
| Abstract: The recent increase in the production pace of functional genomics data has created new opportunities in understanding regulation. Advances range from the identification of new regulatory elements, to the prediction of gene expression from genomic and epigenomic features. At the same time, this data-rich environment has raised challenges in retrieving and interpreting information from these data.\\ \\ Based on recent algorithmic developments, deep artificial neural networks (ANN) have been used to build representations of the input that preserve only the information needed to the task at hand. Prediction models based on these representations have achieved excellent results in machine learning competitions. The deep learning paradigm describes methods for building these representations and training the prediction models in a single learning exercise.\\ \\ In this work, we propose ANN as tools for modeling gene regulation and a novel technique for interpreting what the model has learned.\\ \\ We implement software for the design of ANNs and for training practices over functional genomics data. As a proof of concept, use our software to model differential gene expression during cell differentiation. To show the versatility of ANNs, we train a regression model on measurements of protein-DNA interaction to predict gene expression levels.\\ \\ Typically, input feature extraction from a trained ANN is formulated as an optimization problem whose solution is slow to obtain and not unique. We propose a new efficient technique for classification problems that provides guarantees on the class probability of the features and their norm. We use this technique to identify input features used by the trained model in classification and show how these features agree with previous empirical studies.\\ \\ Finally, we propose building representations of functional features from protein-DNA interaction measurements using a deep stack of nonlinear transformations. We train the model on a small portion of the input and compute small dimensional representations for the rest of the genome. We show that these reduced representations are informative and can be used to label parts of the gene, regulatory elements, and quiescent regions.\\ \\ While widely successful, deep ANNs are considered to be hard to use and interpret. We hope that this work will help increase the adoption of such models in the genomics community. |
| Title: A computational model of drug delivery through microcirculation to compare different tumor treatment options |
|---|
| Seminar: Numerical Analysis and Scientific Computing |
| Speaker: Paolo Zunino of University of Pittsburgh |
| Contact: TBA |
| Date: 2014-03-27 at 4:00PM |
| Venue: MSC N304 |
| Download Flyer |
| Abstract: Starting from the fundamental laws of filtration and transport in biological tissues, we develop a mathematical model able to capture the interplay between blood perfusion, fluid exchange with the interstitial volume, mass transport in the capillary bed, through the capillary walls and into the surrounding tissue. These phenomena are accounted at the microscale level, where the capillary bed and the interstitial volume are viewed as two separate regions. The capillary bed is described as a network of vessels carrying blood flow. We complement the model with a state of art numerical solver, based on the finite element method. The numerical scheme is based on the idea to represent the capillary bed as a network of one-dimensional channels that acts as a concentrated source of flow immersed into the interstitial volume, because of the natural leakage of capillaries. As a result, it can be classified as an embedded multiscale method. We apply the model to study drug delivery to tumors. Owing to its general foundations, the model can be adapted to describe and compare various treatment options. In particular, we consider drug delivery from bolus injection and from nanoparticles, which are in turn injected into the blood stream. The computational approach is prone to perform a systematic quantification of the treatment performance, enabling the analysis of interstitial drug concentration levels, drug metabolization rates, cell surviving fractions and the corresponding timecourses. Our study suggests that for the treatment based on bolus injection, the drug dose is not optimally delivered to the tumor interstitial volume. Using nanoparticles as intermediate drug carriers overrides the shortcomings of the previous delivery approach. Being directly derived from the fundamental laws of flow and transport, the model relies on general foundations and it is prone to be extended in different directions. On one hand, we are planning to combine it with a poroelastic description of the interstitial tissue, in order to capture the interplay of mechanical deformations and transport phenomena. On the other hand, the model may be adapted in future to study different types of cancer, provided that suitable metrics are available to quantify the transport properties of a specific tumor mass. |
| Title: Solving geometric variational problems by gluing |
|---|
| Colloquium: Analysis and Differential Geometry |
| Speaker: Professor Matthew J. Gursky of University of Notre Dame |
| Contact: Vladimir Oliker, oliker@mathcs.emory.edu |
| Date: 2014-03-27 at 4:00PM |
| Venue: MSC W301 |
| Download Flyer |
| Abstract: In this talk I will describe give a schematic overview of a technique for constructing new solutions of geometric variational problems from known ones, called "gluing." I will focus on two examples which share a key feature, namely, conformal invariance. I will begin with an outline of the work of D. Joyce, who used gluing techniques to construct metrics of constant scalar curvature. Then, I will then describe some recent work with J. Viaclovsky, in which we create new examples of four-manifolds which are critical points of L^2-curvature functionals. |
| Title: Forms of Toric Varieties |
|---|
| Seminar: Algebra |
| Speaker: Alex Duncan of University of Michigan |
| Contact: David Zureick-Brown, dzb@mathcs.emory.edu |
| Date: 2014-03-25 at 4:00PM |
| Venue: W302 |
| Download Flyer |
| Abstract: A toric variety is a special kind of compactification of a torus. A basic example of a toric variety is a projective space $\mathbb{P}^n$. Given an n+1-dimensional vector space $V$ there is a canonical projection map from $V$ (minus the origin) to $\mathbb{P}^n$. A construction of Cox generalizes this situation to toric varieties. I will introduce toric varieties, then show how one may use Cox’s construction to classify their forms over non-algebraically closed fields using Galois cohomology. |
| Title: Enabling Highly Accurate Large-Scale Phylogenetic Estimation |
|---|
| Seminar: General Colloquium |
| Speaker: Shel Swenson of University of Southern California |
| Contact: Steve Batterson, sb@mathcs.emory.edu |
| Date: 2014-03-25 at 4:00PM |
| Venue: MSC W303 |
| Download Flyer |
| Abstract: Evolutionary histories of sets of molecular sequences are a fundamental tool in many biological and biomedical questions of societal importance, including biodiversity conservation, drug development, and even forensic investigations. The best methods for estimating these evolutionary histories, or phylogenetic trees, are based on NP-hard optimization problems, and thus phylogenetic analyses of large-scale datasets is extremely computationally intensive. The continually diminishing costs and increasing throughput of DNA sequencing technologies will lead to an ever greater demand for methods capable of producing accurate phylogenetic trees on complex, large-scale molecular datasets. In this talk, I will describe algorithms my collaborators and I have developed to address this demand. I will present SuperFine and ASTRAL, two divide-and-conquer approaches with desirable theoretical properties and excellent empirical performance. Both methods are supertree approaches in that they divide a larger taxon set into subsets, estimate trees on those subsets, and apply a supertree method which assembles a tree on the entire set of taxa from the smaller "source" trees. SuperFine is designed to handle datasets with source tree conflict only due to estimation error, while ASTRAL is designed to handle source tree conflict due to estimation error and incomplete lineage sorting which can cause gene trees to differ from the underlying species tree. I will present supertree methods in a mathematical context, focusing on some theoretical properties of MRP (Matrix Representation with Parsimony), the most popular supertree method, and SuperFine which outperforms MRP. I will also describe a desirable statistical property of ASTRAL and this method's potential to enable highly accurate genome-scale phylogenetic analysis. |
| Title: Algebraic cycles and degeneration |
|---|
| Seminar: Algebra |
| Speaker: Jaya Iyer of IMSC Chennai |
| Contact: David Zureick-Brown, dzb@mathcs.emory.edu |
| Date: 2014-03-25 at 5:00PM |
| Venue: W302 |
| Download Flyer |
| Abstract: We will introduce some questions on the theory of algebraic cycles, and later discuss degenerations of certain one-cycles on jacobian of a curve and on triple product of a curve. These correspond to elements in higher Chow groups. |
| Title: Multi-Structured Inference in Text-to-Text Generation |
|---|
| Seminar: Computer Science |
| Speaker: Kapil Thadani of Columbia University |
| Contact: Vaidy Sunderam, vss@emory.edu |
| Date: 2014-03-21 at 3:00PM |
| Venue: MSC W201 |
| Download Flyer |
| Abstract: Automated personal assistants and summarization tools are increasingly prevalent in the modern age of mobile computing but their limitations highlight the longstanding challenges of natural language generation. Focused text-to-text generation problems present an opportunity to work toward general-purpose statistical models for text generation without strong assumptions on a domain or semantic representation. In this talk, I will present recent work on a supervised sentence compression task in which a compact integer linear programming formulation is used to simultaneously recover heterogenous structures which specify an output sentence. This inference strategy avoids cyclic and disconnected structures through commodity flow networks, generalizing over several recent techniques and yielding significant performance gains on standard evaluation corpora. I will then discuss a number of extensions to this multi-structured generation approach. One line of research explores approximation strategies using Lagrangian relaxation, dynamic programming and linear programming in order to speed up inference while preserving performance. Other extensions exploit the flexibility of the formulation and extend it with minimal additions to new problems such as the more challenging task of merging sentences, as well as to new structures including directed acyclic graphs that represent frame semantics. Finally, I will briefly discuss our use of multi-structured inference in other natural language applications such as summarization and alignment. |