talk: Real-time Causal Anomaly Detection for Hyperspectral Imagery

UMBC CSEE Colloquium

Real-time Causal Anomaly Detection for Hyperspectral Imagery

Yu-Lei Wang
Information and Communication Engineering College
Harbin Engineering University, China

1:00pm Friday, 12 October 2012, ITE 227, UMBC

Due to availability of very high spectral resolution, a hyperspectral imaging sensor is capable of uncovering many subtle signal sources which cannot be visually inspected or known by prior knowledge. Such signal sources generally appear as anomalies in the data. As a result, anomaly detection has received considerable interest in hyperspectral imaging. In anomaly detection real time causal processing is particularly important and crucial. This is because many anomalies, such as moving targets, may not stay long enough and the duration of their presence is very short. Most importantly, they may show up suddenly and instantly, then disappear quickly afterwards. Therefore, for an algorithm to be able to detect these targets in a timely fashion, the process must be real time. In addition, the data that can be used should be only those which have been visited and processed. So, the data processing must be also causal as well. Such causality is a very important pre-requisite to real time processing. Our work is believed to be the first work devoted to exploring this concept into anomaly detection. Specifically, it further derives a causal innovations information update equation for implementing real time causal anomaly detection. This concept which makes use of only innovations information provided by the pixel currently being processed without re-processing previous pixels is similar to those derived in Kalman filtering.

Yu-Lei Wang received her BS degree in Electrical Engineering from Harbin Engineering University, China in 2009 and is currently a Ph.D. student in the same university. Since December 2011 Ms. Wang has been working in the Remote Sensing Signal and Image Processing Laboratory at UMBC on hyperspectral anomaly detection under a China State Scholarship awarded by China Scholarship Council for a two-year visit to UMBC. Ms. Wang's research interest includes remote sensing image processing and vital sign signal processing.

more information and directions

CSEE professor Tinoosh Mohsenin to speak at Grace Hopper conference, 10/3

Tomorrow, Wednesday, October 3rd, Dr. Tinoosh Mohsenin will speak at the Grace Hopper Celebration of Women in Computing Conference at the Baltimore Convention Center. She will talk about a "A Many-core Platform for Intelligent Biomedical Systems". Her presentation will be on "Data Intensive Computing"  in the New Investigators session starting at 10:45 am. For more details, visit: http://gracehopper.org/2012/schedule-at-a-glance/10-3/

 
Abstract
This talk presents a low power programmable many-core platform well suited for portable biomedical and DSP applications and contains 64 cores routed in a hierarchical network. For demonstration, Electroencephalogram (EEG) seizure detection and analysis and ultrasound spectral doppler are mapped onto the cores. The seizure detection and analysis algorithm takes 900 ns and consumes 240 nJ of energy. Spectral doppler takes 715 ns and consumes 182 nJ of energy. The prototype is implemented in 65 nm CMOS which contains 64 cores, occupies 19.51 mm2 and runs at 1.18 GHz at 1 V.
 
Speaker’s bio
Dr. Tinoosh  Mohsenin is an assistant professor in the  Department of Computer Science and Electrical Engineering at the University of Maryland Baltimore County since 2011. Prior to joining UMBC, she was finishing her PhD at the University of  California, Davis. Dr. Mohsenin’s research interests lie in the areas of high performance and energy-efficiency in programmable and special purpose processors. She is the director of Energy Efficient High Performance Computing (EEPC) Lab where she leads projects in architecture, hardware, software tools, and applications for VLSI computation with an emphasis on digital signal processing workloads. She has been consultant to early stage technology companies and currently serves inTechnical Program Committees of the IEEE Biomedical Circuits & Systems Conference (BioCAS), Life Science Systems and Applications Workshop (LiSSA), International Symposium on Quality Electronic Design (ISQED) and IEEE Women in Circuits and Systems (WiCAS). 

MS defense: Using Mobile Data Collectors to Federate Clusters of Disjoint Sensor Network Segments

MS Thesis Defense

Using Mobile Data Collectors to Federate Clusters
of Disjoint Sensor Network Segments

Bhuvana Kalyanasundaram

11:00am Tuesday 2 October 2012, ITE 346

Wireless Sensor Networks (WSN) operating unattended in harsh environments have the higher probability of suffering from large scale damage, where many nodes fail simultaneously and the network gets partitioned into several disjoint segments. Restoring connectivity of structurally damaged WSN’s segments may be very urgent considering that they are employed to assist in risky missions. A similar scenario is when multiple standalone networks are to be federated to serve an emerging event such as an earthquake and conduct search-and-rescue. To deal with these scenarios, Mobile Data Mules (MDMs) are employed to establish intermittent links by moving around and carrying data from one segment to another. To limit data delivery latency and minimize the motion overhead, the travel path of the MDM should be shortened. We present a novel algorithm that groups the segments into k overlapping clusters based on the inter-segment proximity. Each cluster is assigned a distinct MDM to tour its segments. A segment that belongs to two clusters serves as a gateway that enables data transfer across clusters. Our algorithm minimizes the tour length for each MDM and sets the speed of the individual MDMs to rendezvous at the gateway nodes so that buffering space and time for inter-cluster traffic are minimized.

Committee: Drs. Mohamed Younis (chair), Charles Nicholas and Chintan Patel

talk: introduction to the OpenACC parallel programming standard

Introduction to OpenACC

Mark EbersoleNVIDIA

1:00pm Thursday, 27 September 2012, ENG 005a

Modern GPUs have grown past their graphics heritage and evolved into the world's most successful parallel computing architecture. The introduction of this talk will briefly cover where the GPU came from and how it turned into this processing powerhouse. We will then look into how to access this power by using the relatively new standard called OpenACC. This method is a balance between the maximum flexibility you get by writing your own kernels and the ease of use you get using existing libraries. We will then end the lecture looking at the existing GPU Computing ecosystem that works well with OpenACC.

As CUDA Educator at NVIDIA, Mark Ebersole teaches developers and programmers about the NVIDIA CUDA parallel computing platform and programming model, and the benefits of GPU computing. With more than ten years of experience as a low-level systems programmer, Mark has spent much of his time at NVIDIA as a GPU systems diagnostics programmer in which he developed a tool to test, debug, validate, and verify GPUs from pre-emulation through bringup and into production. Before joining NVIDIA, he worked for IBM developing Linux drivers for the IBM iSeries server. Mark holds a BS degree in math and computer science from St. Cloud State University.

Host: Marc Olano,

more information and directions

talk: Volume Calculation of Magnetic Resonance Tissues via Image Classification

CSEE Colloquium

Volume Calculation of Magnetic Resonance
Tissues via Image Classification

Shih-Yu Chen
Remote Sensing Signal & Image Processing Laboratory
UMBC Computer Science and Electrical Engineering

1:00pm Friday 5 October 2012, ITE 227

Magnetic resonance (MR) tissue volume calculation is very important in medical diagnosis. A general approach is to first perform image classification of desired tissue substances slice by slice and then calculate tissue volumes via classified data samples in each slice. Two issues are generally involved; (1) selection of training samples which are slice-dependent, i.e., each slice requires its own specific training samples and (2) classification which must be carried out slice by slice individually because training samples obtained from one slice are not necessarily applicable to another. We develop a volume sphering analysis (VSA) approach which can process all MR image slices as one single image cube to calculate tissue volumes via image classification using only one set of training samples that is obtained from a single image slice. The proposed VSA using one set of training samples not only performs comparably to that using training samples specifically selected for individual image slices, but also saves significant amounts of selecting training samples and computing time.

Shih-Yu Chen received the BS degree in Electrical Engineering from Da-Yeh University in 2005, and the MS EE degree from National Chung Hsing University in 2010. He is currently a PhD (EE) student at UMBC. Mr. Chen's research interest includes medical image, remote sensing image and vital sign signal processing.

more information and directions

talk: Geometric Modeling and Visualization for Science, 3pm Wed 10/3

CSEE Colloquium

Geometric Modeling and Visualization for Science

Dr. Liz Marai
Associate Professor of Computer Science
University of Pittsburgh

3:00pm Wednesday 3 October 2012, ITE325b

The incredible array of measurement technologies available to the scientific community is changing fundamentally our understanding of physical and biological processes. However, scientific data acquisition marks only the first step. To turn numbers into insight, computer graphics and visualization help us model complex systems, make predictions about their behavior, and finally harness the immense power of the human visual perception system to make insights into complex processes possible. In this talk I will present several novel geometric representations, computational modeling, and visual analysis tools to facilitate the simulation and analysis of such complex scientific phenomena. These representations and tools were developed at the Pitt Interdisciplinary Visualization Research lab I direct, and have applications in domains as diverse as neuroimaging, astronomy, biology, turbulent combustion, or machine translation.

Liz Marai is an Assistant Professor of Computer Science at the University of Pittsburgh, with joint and adjunct appointments in the Pitt Department of Computational Biology and at the CMU Robotics Institute. She is the Director of the Interdisciplinary Visualization Research lab at Pitt, featuring interdisciplinary research in computational modeling, data visualization, and computer graphics. She is a recipient of an NSF CAREER award, of a recent Best Paper Award at BioVis 2011, and of multiple teaching awards for courses that blend research and teaching.

Host: Dr. Jian Chen,

More information and directions

talk: Wolff on Local Thresholding for Structured and Unstructured Graphs, 2pm 9/28

UMBC CSEE Colloquium

Local Thresholding for Structured and Unstructured Graphs

Dr. Ran Wolff, Haifa University, Israel

2:00pm Friday, 28 September 2012, ITE 325B
note unusual time and room

Local thresholding algorithms were first offered a decade ago as a communication thrifty alternative for computation in large distributed environments. Their disadvantage, however, has always been in their brittleness. A single cycle in the communication graph could mean the algorithm converges to the wrong value. This talk describes two advances in local thresholding algorithms which overcome the demand for cycle freedom. The first is a local tree induction protocol for structured peer-to-peer networks which seamlessly integrates with the local thresholding algorithm. The second are new local stopping and update rules which permit execution of the local thresholding algorithm on general graphs. The first solution vastly outperforms a gossip based algorithm on simple computation tasks in a Chord-like peer-to-peer network. The second may transform the way data is processed in wireless sensor networks, where gossip is mostly considered impermissibly costly.

UMBC Host: Hillol Kargupta,

More information and directions

PhD proposal: Birrane on Virtual Circuit Provisioning in Challenged Sensor Internetworks

Ph.D. Dissertation Proposal

Virtual Circuit Provisioning in Challenged Sensor Internetworks,
with Application to the Solar System Internet

Ed Birrane

9:00am Friday, 21 September 2012, ITE 325b, UMBC

As sensing devices are applied to increasingly diverse tasks the network architectures that connect them must handle increasingly complex sets of operational constraints. One dimension in which these networks must scale is in their spatial footprint: there is a desire to distribute sensing devices over areas from miles to hundreds of miles to millions of miles. A second dimension in which these networks scale is in their media access heterogeneity: to gradually cover larger distances, existing networks (that may not otherwise communicate amongst themselves) must be stitched together. Examples of such networks include the Solar System Internet (SSI), Autonomous Underwater Surveillance (ASU), National Border Protection (NBP) and Intelligent Highway Initiatives (IHI).

I propose that the non-random sensing performed in these networks supports the establishment of virtual circuits that communicate information more efficiently than in broadcast mesh networks. Specifically, virtual circuits may be pre-negotiated using data-link-agnostic overlay techniques based on directed, weighted, time-variant graphs. The construction and maintenance of these circuits is feasible in non-random networks and may be accomplished through proposed protocols and stochastic processes. My first contribution will define an emerging, useful special case of networks. I label this architecture the "Challenged Sensor Internetwork" (CSI) and provide models relating to data motion and path selection. My second contribution will provide algorithms and associated analysis for path selection and synchronization. The network topology created by a CSI is graphically modeled as a multi-hypergraph. Since transmission in a CSI is wireless, a single transmission may be received by multiple nodes in the network, hence a hypergraph. However, as a challenged network, link opportunities amongst nodes will change as a function of time, hence a multigraph. I will show that the multi-cast problem, as formulated for CSIs, is NP-Complete, propose an approximation algorithm for the generation of paths in such a multi-hypergraph, and provide an analysis of the performance of this algorithm. My third contribution will provide heuristic algorithms and performance measurements. Each node in the CSI must store its own copy of the network graph so as to make local routing decisions. Synchronization of these network graphs across the network is often impossible. I propose two heuristic mechanisms, based on my proposed principle of path locality, to synchronize preferred path information in the network: exchanging relevant sub-graphs along paths as part of nominal messaging and altering local graphs based on predicted congestion based on observed traffic. Finally, I propose a method for inferring overlay-level contact opportunities from routing information available to local nodes via existing physical and data link layer mechanisms. My final contribution will demonstrate this work in the context of a real-world CSI deployment. I will provide a case study demonstrating how the SSI networking concept exemplifies the definition and characteristics of a CSI and showing how my proposed algorithms are mission enabling to existing, published SSI scenarios.

Several portions of the proposed dissertation work have been completed and validated through simulation and peer-reviewed publication. To complete the dissertation, I plan to finalize the problem statements, proofs, and algorithm analysis supporting achieved heuristic results. I will also apply these algorithms to scaled simulations and emulations of increasingly complex CSIs.

Committee: Drs. Dr. Mohammed Younis (Chair), Alan Sherman (Co-Advisor) Dhananjay Phatak, Vinton Cerf, Keith Scott, Hans Kruse

More information and directions

talk: Simon on A Novel Dynamic Task Scheduling Environment for High Performance Distributed Systems

CSEE Colloquium

A Novel Dynamic Task Scheduling Environment
for High Performance Distributed Systems

Tyler Simon

Faculty Research Assistant, UMBC

1:00pm Friday, 21 September 2012, ITE 227, UMBC

The number of concurrently executing tasks required for a single application to perform at the petascale is on the order of hundreds of thousands. Given current manycore hardware trends, future peta- and exa-scale class systems will require applications to run tasks on the order of hundreds of millions to billions. To address the problem of creating, running and managing jobs of this scale, both from a system user and administration perspective we have developed, ARRIA, an Autonomic Runtime for Resource Intensive Applications. ARRIA uses a decentralized bag of tasks and workload scheduler that increases individual job priorities based on weighed factors that are of interest to the application programmer or the system administrator. ARRIA is designed to run millions of independent tasks reliably and efficiently without explicit message passing from the user. In previous work, using the ARRIA scheduler for scientific MapReduce workloads, we have shown a 2.1x speedup over the Hadoop Fair Share scheduler. We investigate novel scheduling parameters and strategies that guarantee efficient job execution for a wide range of realistic and simulated workloads with both user and administrator objectives, such as increased throughput and maximized utilization with minimal wait times for specific job classes. Finally our experiments investigate the long tail phenomenon for mixed workloads and the overheads incurred for increased system size.

Mr. Simon has undergraduate degrees in Computer Science and Philosophy with a Master of Science in Computer Science from the University of Mississippi, he is currently pursuing a PhD in Computer Science at the University of Maryland Baltimore County. Mr. Simon has worked professionally in the high performance computing (HPC) field for over a decade. In 2005 he earned a Department of Energy graduate research fellowship at Oak Ridge National Laboratory, where he worked for in the Computer Science and Mathematics Division developing and implementing the Freeloader distributed storage system. Mr. Simon has worked as a computational scientist for the Department of Defense High Performance Computing Modernization Office based at the U.S. Army Engineer Research and Development Center in Vicksburg, MS, evaluating both current and future HPC system requirements for applications of interest to the Department of Defense. Since 2009 Mr. Simon has been a computational scientist and manager of HPC user services at the NASA Center for Climate Simulation at Goddard Space Flight Center and is currently a Faculty Research Assistant at the University of Maryland Baltimore County working at the NSF Center for Hybrid Multicore Productivity Research. Mr. Simon’s research involves the study of dynamic distributed runtime environments, parallelization strategies and scheduling of large scale scientific applications for current petascale and future HPC architectures.

For more information and directions see http://bit.ly/UMBCtalks.

talk: Ҫağatay Demiralp on Computational Brain Connectivity Using Diffusion MRI

 

CSEE Colloquium

Computational Brain Connectivity Using Diffusion MRI

Ҫağatay Demiralp
Brown University

1:30pm Tuesday, 18 September 2012, ITE 325B

In my talk, I’ll present examples from modeling, visualization, and analysis of diffusion-derived structural brain connectivity. I’ll first introduce two interactive visual analysis tools that use novel planar representations of the brain. I’ll show that two-dimensional map representations that are viewed, interacted with, and enriched like online geographical maps result in faster and more accurate exploration of brain connectivity.

Second, I’ll introduce neural tract-based probability density functions, including joint densities of tract arc length and scalar diffusivity measures, as biomarkers. I’ll demonstrate their simple and effective use in detecting individual and group differences. I’ll also describe a new coherence measure for neural tract clusters based on geometric slicing. I’ll show that a refinement of neural tract clustering based on this measure leads to a significant improvement in clustering results that is not possible directly using standard methods.

Third, I’ll describe a new coloring method for three-dimensional line fields based on Boy's real projective plane immersion. This coloring method is smooth and one-to-one, except on a set of measure zero. I’ll demonstrate its use in visualization of neural tracts and cross-sectional diffusion MRI brain images.

Çağatay Demiralp is a PhD candidate in computer science at Brown University. His research interests are in characterizing patterned structures in data both qualitatively and quantitatively using topological, geometric as well as statistical approaches. While computational brain connectivity using diffusion MRI has been the focus of his thesis research, he has published on a diverse set of topics ranging from surface deformation to semantic segmentation. He received Brown University’s Brain Sciences Research award, IEEE Vis Best Poster award, and ASSH Best Layout and Best Scientific Presentation awards.

1 41 42 43 44 45 58