Saint Louis University Menu Search

Upcoming talks

3:00pm Thursday 28 February 2019, Ritter 216
Michele Berno, SLU Visiting Scholar

ZipNet-GAN: Inferring Fine-grained Mobile Traffic Patterns via a Generative Adversarial Neural Network

This talk will present a summary of a recent paper by Zhang, Ouyang, and Patras

3:10pm Monday 4 March 2019, Ritter 115
Abby Stylianou, George Washington University

TraffickCam: Deep Learning and Image Search to Combat Human Trafficking

Victims of sex trafficking are often photographed in hotel rooms for online advertisements of sex services. Identifying the hotels in these photographs is a top priority for trafficking investigators and prosecutors — they show where a victim has been trafficked previously and where their trafficker may move them or others in the future. We propose recognizing the hotels in these photographs as an image search problem, where the most likely hotel is inferred from the most similar images of hotel rooms. This is, however, a challenging image search problem, due both to the properties of the victim photographs, which include unusual viewpoints and large occlusions, and the properties of hotel rooms, which may be visually dissimilar within the same hotel, but visually similar across different hotels, particularly those from the same chain. My research has focused on deep learning approaches to large scale image search that are robust to such challenging properties, as well as visualization approaches that explain why deep learning models trained on such image similarity problems find particular images to be similar. TraffickCam, our mobile application to collect images of hotel rooms from the traveling public is currently used by over 150,000 individuals who upload photos specifically to help combat trafficking. The data from this app, combined with millions of publicly available images from travel websites, support a first in the world system for image search to identify hotels in trafficking imagery that we have deployed at the National Center for Missing and Exploited Children.

3:10pm Monday 25 March 2019, Ritter 115
Erik Storrs, Washington University

Practical machine learning

Now more than ever, machine learning has been made approachable through numerous libraries and apis. I’ll show why if you can code, you can learn enough machine learning to be dangerous. And I’ll share lessons learned from implementing machine learning in a variety of fields from patent law to cancer genomics.

3:10pm Monday 8 April 2019, Ritter 115
Mark Tabor, SLU Alumnus

Past talks

3:00pm Thursday 21 February 2019, Ritter 216
Tommaso Pecorella, SLU Visiting Scholar

An Introduction to Network Simulator 3

NS3 is one of the largest open-source computer network simulators. Dr. Pecorella has been "git pushing" contributions to the project for many years. He also has a project within Google Summer of Code on it.

3:10pm Monday 18 February 2019, Ritter 115
Lyndon Coghill, LSU Center for Computation & Technology

Understanding the Tree of Life: computational approaches to unraveling the relationships between all living things

Despite a deluge of genomic sequence data pouring into data repositories in recent years, reconstructing the phylogenetic relationships that unite all lineages (the tree of life) remains a grand challenge in biology. Our ability to collect genetic sequence data has exceeded our ability to analyze that data in an informative way using traditional tools. Yet, despite collecting large amounts of data, there is still a paucity of homologous character data across disparately related lineages, rendering direct phylogenetic inference untenable in many cases. Our recent work leverages computational methods such as graph theory, statistical modeling, and high-performance computing to find solutions to these challenges. By using graph-based approaches to synthesize published phylogenies together with taxonomic classifications we have been able to generate the first "knowledge complete" draft tree of life with over 2.3 million species. Interestingly, this tree highlighted many deficiencies in our current understanding and has led to additional questions about the quality of data and methods used to generate these trees. Follow up analyses suggest that the information content in Genbank (containing most of the published sequence data in the world) is quite low, suggesting many researchers are repeating studies on the same questions with different types of data. Adding to the complexity of this challenge, many of these studies reach different, but equally strong conclusions as different types of data are added. Our explorations into the underlying cause of these challenges hint that these inconsistencies often seem to be driven by poor model adequacy and strongly influential outliers in the data sets. These findings show that one of the biggest hurdles to achieving one of the grandest goals in biology, quantifying all of life, is driven more by by a lack of appropriate models and analysis methods, than by a lack of data.

3:30pm Thursday 14 February 2019, Ritter 115
Patrick Trainor, DiscernDx Inc.

Bioinformatics for Mass Spectrometry-based Metabolomics/Lipidomics and a Novel Bayesian Approach for Systems-level Inference

The quantitation of metabolites (the intermediates of metabolic processes that enable life) from biological samples provides deep characterization of cell, tissue, and organism phenotypes. While untargeted mass spectrometry is a robust analytical platform for detecting and providing relative quantification of metabolites and lipids (a subclass of metabolites), the data generated from mass spectrometers is complex and multi-dimensional. In this talk, we will first discuss bioinformatic approaches for processing raw Liquid Chromatography–Mass Spectrometry (LC-MS) data to yield useful datasets for conducting molecular biology research. We will then discuss a novel Bayesian methodology that we have developed for making systems-level inference using processed LC-MS data. This methodology utilizes informative priors that are generated via the analysis of molecular structure to enable the estimation of metabolite “interactomes” (or probabilistic models) which are organism, sample media, and condition specific as well as comprehensive. The generated interactomes can serve as reference models for studying perturbations in metabolic processes. We will briefly discuss the software we developed for implementing the methodology and computational optimization of the underlying linear algebra routines. In addition to evaluating the performance of the developed methodology via simulation, we will discuss an application of the methodology to developing a plasma metabolite interactome for stable heart disease. The metabolite and lipid data was generated from plasma samples from human subjects who participated in a study of stable heart disease and myocardial infarction (heart attack) at two hospitals in Louisville, Kentucky.

3:10pm Monday 11 February 2019, Ritter 115
Darrel Jiang, Bayer

Writing Cleaner, More Sustainable Code with Test-Driven Development

Tests for your applications can be an afterthought or might not even be a thought at all. As projects grow and as they change hands from team to team, the absence of these fundamental principles manifests themselves in the form of tech debt and an exponential cost in maintainability. In this presentation, we’ll explore the core concepts of what test-driven development is, why you should care, and also how you implement it. Together, we’ll work through a live demonstration of how to approach a problem using basic TDD principles with Javascript, Node, and the Jest testing framework.

3:10pm Wednesday 6 February 2019, Ritter 115
Jie Hou, U. Missouri-Columbia

Improving Prediction of Protein Three-dimensional Structure using Deep Learning Techniques

Protein structure prediction is one of the most important scientific problems in bioinformatics and computational biology field. The availability of protein three-dimensional (3D) structure is crucial for studying biochemical, biological and cellular functions of proteins. Deep learning techniques have emerged as one of the most effective machine learning methods in recent years and brought revolutionary advances in computer vision, speech recognition and bioinformatics. In this talk, I will first introduce the fundamental ideas and algorithms behind deep learning and explain how the deep learning methods can be applied to protein structure prediction. Then I will present my latest research of applying deep learning techniques to tackle three major sub-problems in protein structure prediction, including protein secondary structure prediction, protein fold recognition and protein quality assessment. Finally, I will describe how to integrate all these methods to improve the prediction of protein three-dimensional structures. The methods were officially ranked among the top three out of 98 predictors in the category of structure prediction and estimating the accuracy of protein structural models in the 13th world-wide Critical Assessment of Techniques for Protein Structure Prediction (CASP13) competition, demonstrating the importance and significance of deep learning techniques in the protein structure prediction. The rigorous evaluation of these methods during the CASP13 (2018) will also be discussed in this talk.

3:10pm Monday 4 February 2019, Ritter 115
Lav Gupta, Washington University

Management and Security of Multi-cloud Applications

Single cloud platforms like Amazon’s EC2 and Microsoft Azure are common and popular today. Obtaining resources from multiple cloud systems gives clients competitive pricing, flexibility of resource provisioning, better points of presence and reduced risk of a total blackout. When these clients happen to be carriers contemplating to host their offerings over multiple clouds, there still are many research challenges that inhibit large-scale deployments. This talk revolves around some of the key issues that werethere in the 'to do' list at the beginning of the network virtualization journey and still need considerable attention on the part of the research community to seeany kind of resolution in near future. In this talk I will discuss some methods that can successfully handle optimized placement of virtual networking resources, improve availability of virtual network services and secure flow of data in the context of IoT and multi-cloud based health networks.

3:10pm Monday 28 January 2019, Ritter 115
Matthew Gottsacker, Saint Louis University

Toward Effective Visualization of Network Identifier Bindings in a Software-Defined Network

Software-defined networking allows network programmers to enforce dynamic, role-based access control policies on high-level information such as usernames at the network level. In order to enforce a policy based on a username, the system also needs knowledge of a network device’s hostname, IP address, and MAC address. We refer to the relationships between two network identifiers as bindings. This presentation will discuss employing different visualization techniques to help an analyst efficiently understand the dynamic states of those bindings in an enterprise network. No background in networks or visualization is assumed.

3:10pm Monday 26 November 2018, Ritter 115
David Letscher, Saint Louis University

The Shape of Data

We will explore the fundamentals of TDA (topological data analysis) which utilizes techniques from topology to find structural information about data. TDA lies in the intersection of computer science, mathematics, statistics and data science and has been applied to solve problems in a broad range of domains. In particular, we will examine techniques to determine the underlying shape and its properties from finite data sets. No background in topology is assumed.

3:10pm Monday 19 November 2018, Ritter 115
Ted Ahn, Saint Louis University

Computer Science for Bioinformatics: Use case of Apache Spark and Machine Learning in Bioinformatics

This will be a gentle introduction to bioinformatics research topics from the angle of computer scientist. After brief introduction of bioinformatics, I will present how Apache Spark and Machine Learning used on my large-scale data science projects to make a synergy.

3:10pm Monday 5 November 2018, Ritter 115
Erin Chambers, Saint Louis University

Computing Optimal Homotopies

The question of how to measure similarity between curves in various settings has received much attention recently, motivated by applications in GIS data analysis, medical imaging, and computer graphics. While geometric measures such as the Hausdorff and Frechet distance have efficient algorithms, measures that take into accoutn the underlying topology of the space are much newer. In this talk, we will consider using homotopy, or continuous deformation of one curve to another, in order to quantify how similar two input curves are. In particular, we will survey recent work on computing optimal homotopies in a variety of settings, discussing both algorithms and the complexity of the problem in general.

3:10pm Monday 29 October 2018, Ritter 115
Mary Hogan, Princeton University

Music-Defined Networking

For several years researchers have used the term "network orchestration" as a metaphor. In this project, we make the metaphor reality; we describe a novel approach to network orchestration that leverages sounds to augment or replace various network management operations. We test our Music-Defined Networking approach with both a real and a virtual network testbed, on several mechanisms and applications: from datacenter server fan failure detection to authentication, from load balancing to explicit congestion notification and detection of heavy hitter flows. Our approach can be used with and without a Software-Defined Network controller. Despite its limitations, we believe that sound-based network management has potential to be further explored as an effective and inexpensive out-of-band orchestration technique.

3:10pm Monday 8 October 2018, Ritter 115
Reza Tourani, Saint Louis University

Securing the Future Internet: DDoS Prevention, Anonymity, and Access Control in NDN

3:10pm Monday 1 October 2018, Ritter 115
Agnese Ventrella, Politecnico di Bari

Learning From the Past to Build a Better Internet Architecture: Is Information Centric Networking the Answer?

The current Internet architecture was born in 1971 as an academic network of fixed and trustworthy hosts, to allow the communication among the scientific community. Its usage radically changed in the last few years: it is now a global infrastructure for a massive distribution of information generated by billions of (mobile) users. To cope with the constant Internet evolution and to reduce the complexity introduced by cumbersome patches and middleware layers, the scientific community invested millions on new Internet architectures redesign. The Information Centric Networking (ICN) paradigm emerged as one of the most promising approaches. In this talk, we will first investigate some of the current Internet flaws and discuss how they originated. Then we dissect a few characteristics of some of the newly proposed Internet architectures, discussing their potential and limitations.

3:10pm Monday 10 September 2018, Ritter 115
Kevin Scannell, Saint Louis University

Recent Advances in Neural Language Models

This will be a gentle introduction to the foundational problem of probabilistic language modeling. I'll quickly review some of the traditional (non-neural) approaches and then discuss recent breakthroughs that use deep neural networks to obtain state of the art results for English. I'll close with some of the challenges these developments present for languages with more complicated morphology.