From sherman@umbc.edu Thu Apr 30 11:42:17 2009 Date: Thu, 30 Apr 2009 11:41:36 -0400 From: Dr. Alan T. Sherman To: CSEE ALL Subject: [Csee-faculty-lecturer] Abstracts for CSEE Research Review Talks Reminder: The Annual CSEE Research Review is this Friday 9am-3:45pm at South Campus. There will be a free lunch. All graduate students and research faculty are expected to come, to celebrate some of the best research from our department this year. One highlight: Marc Olano will present some of his recent research on computer gaming and progress in our game-development track. CSEE Research Review - Talk Abstracts Friday, May 1, 2009 Department of Computer Science and Electrical Engineering University of Maryland, Baltimore County (UMBC) Session I Charles Nicholas "Who wrote this document?" Questions of authorship have fascinated historians, theologians, and other scholars for centuries. In recent years statisticians and now at last computer scientists are also addressing these issues. We will present an overview of the study of authorship attribution, including famous examples such as "The Federalist Papers," the various "Wizard of Oz" books, as well as the Hebrew Bible and the Christian New Testament. We will present results from our own work in applying Latent Semantic Analysis (LSA), a well-known technique in information retrieval, to the authorship attribution problem. Adam Anthony (Advisor: Marie desJardins), MAPLE Lab - Honorable mention research by PhD student Fast relational clustering using the block modularity clustering objective I discuss a new algorithm for clustering in relational data that emphasizes scalability to large relational data sets. The algorithm uses a unique clustering objective, referred to as Block Modularity, which measures the quality of any general relational pattern, including community, hierarchical and k-partite relational patterns. First, I motivate the area of relational clustering in general, and then compare and contrast two popular approaches to relational clustering. Then, I describe the block modularity objective and the resulting optimization algorithm, and present empirical evidence that the new algorithm finds relatively high quality clusterings for a very low cost in runtime. Finally, I briefly discuss areas of current and future work, including the remaining contributions in my dissertation: the probabilistic relational clustering framework, multi-relational clustering, and relation selection. Peter A. Hamilton (Advisor: Marie desJardins), MAPLE Lab - Best Research by BS student Applying swarm rule abstraction to a wireless sensor network domain Rule abstraction is a powerful tool for modeling abstract behaviors in swarm systems. The research presented in this paper examines the application of rule abstraction to the wireless sensor network domain. I analyze the potential of rule abstraction to accurately model and control the connectivity, coverage, and density of a simple sensor network. I also present a simulation tool developed to facilitate the discovery and creation of new abstract rules and discuss preliminary experimental results that will lead to the development of new abstract rules. Session II Tim Finin, ebiquity Group Creating and exploiting a web of (semantic) data Twenty years ago Tim Berners-Lee proposed a distributed hypertext system based on standard Internet protocols. The Web that resulted fundamentally changed the ways we share information and services, both on the public Internet and within organizations. That original proposal contained the seeds of another effort that has not yet blossomed: a Semantic Web designed to enable computer programs to share and understand structured and semi-structured information easily. We will review the evolution of the idea and technologies to realize a Web of Data and describe how we are exploiting them to enhance information retrieval and information extraction. Aaron Curtis (Advisor: Marc Olano), VANGOGH Lab - Honorable mention research by MS student Real-time soft shadows on the GPU via Monte Carlo Sampling Realistic shadows present a difficult problem in real-time rendering. While techniques for rendering hard edged shadows from point light sources are well established, attempts to incorporate soft shadows typically suffer from inaccuracies or poor performance. Our algorithm makes use of recent advances in GPU randomization to perform Monte Carlo sampling of points on an area light source. Rays are then traced to the sampled points, using the shadow map as a discretized representation of occluders in the scene. The accuracy of this method can be improved through the use of multiple shadow maps, which together are able to better approximate the scene geometry. As with conventional shadow mapping, our method is performed entirely on the GPU, does not require any precomputation, and can handle fully dynamic scenes with arbitrary geometric complexity. The quality of the generated shadows is comparable to that of off-line rendering algorithms such as ray tracing, while performance remains real-time, on par with existing techniques. Deepak Chinavle (Advisor: Tim Oates), Coral Lab - Best research by MS student Adversarial classification: An ensemble-based approach Spam has been studied and dealt with extensively in the email, web and, recently, the blog domain. Recent work has addressed the problem of non-stationarity of data using ensemble based approaches. Adversarial classification has been handled by retraining base classifiers using labeled samples obtained from the ensemble. However, frequent retraining is expensive. The need is to determine dynamically when the classifiers should be retrained and to retrain only those classifiers that are performing poorly. We show how mutual agreement between classifiers can be use to reduce retraining time, measure runtime performance, and keep track of the weakest performing classifier. We back our research with experimental results using real life data from blogs as a special case of spam Session III Marc Olano, VANGOGH Lab Simulation in real-time computer games Modern computer games perform physical simulation in a variety of contexts. It is perhaps most visible in kinematic simulation of articulated bodies, the so-called "rag doll physics" for character animation. However, simulation can also be found for cloth, fluid, optical effects with participating media, and models of surface reflectance. These methods must run effectively on a range of consumer hardware, and must be fast, finishing within a fraction of the 10-30 ms time available per frame. This talk will present some of the mainstream and research methods, demonstrate their results, and discuss the state of consumer-level hardware to accelerate physical simulation. Fusan Yaman, Research Assistant Professor, MAPLE & Coral Labs A context driven approach for workflow mining Workflows play an important role in automation. The key challenge in building workflows is the need for a domain expert. However, in many realistic domains, experts are either not readily accessible or cannot express their knowledge in a declarative way. A solution to this bottleneck is workflow mining, which aims to discover a workflow given sample executions. In this talk, I will present a novel workflow mining algorithm, WIT, which applies grammar inferencing techniques to discover the target workflow with the help of the dataflow. WIT is designed to cope with the challenge of learning when training examples are scarce. It has been deployed as a major component in the POIROT system, which is built to take DARPA's integrated learning challenge, where several learners interact to solve different parts of a complex learning problem given only one example solution. Muhammad A. Talukder (Advisor: Curtis Menyuk), Photonics Lab - Best research by PhD student Analytical and computational study of self-induced transparency mode locking in quantum cascade lasers The possibility of using the self-induced transparency effect to modelock lasers has been discussed since the late 1960s, but has never been observed. It is proposed that quantum cascade lasers are the ideal tool to realize self-induced transparency modelocking due to their rapid gain recovery times and relative long coherence times, and because it is possible to interleave gain and absorbing layers. Designs of quantum cascade lasers are presented here that satisfy the requirements for self-induced transparency modelocking. Analytical modelocked solutions of the coupled Maxwell-Bloch equations that define the dynamics in quantum cascade lasers that have both gain and absorbing layers have been found under the conditions that there is no frequency detuning, the absorbing layers have a dipole moment twice that of the gain layers, the input pulse is a ð pulse in the gain medium, and the gain recovery times in the gain and the absorbing layers are much longer than the coherence time T2 and are short compared to the round-trip time. It is shown that the modelocked pulse durations are on the order of T2, which is typically about 100 fs. The Maxwell-Bloch equations have been solved computationally to determine the robustness of the modelocked solutions when frequency detuning is present, the dipole moment of the absorbing layers differs from twice that of the gain layers, the gain relaxation time is on the order of 1-10 ps, as typically obtained in quantum cascade lasers, and the initial pulse is not a ð pulse in the gain medium. We find that modelocked solutions exist over a broad parameter range. _______________________________________________ Csee-faculty-lecturer mailing list Csee-faculty-lecturer@cs.umbc.edu http://lists.cs.umbc.edu/mailman/listinfo/csee-faculty-lecturer