Overview of CS Department research - part 1

Presenters: Clem Izurieta, Hunter Lloyd, Mike Wittie, Qing Yang

Description: We will present past and ongoing work taking place in our research groups.

Date: Monday, Nov 26 2012 - 4:10 pm
Location: EPS 108


Technical Debt Reduction Using a Game Theoretic Competitive Source Control Approach

Presenter: Sarah Morrison-Smith

Description: The management of technical debt and the use of productivity games are important aspects of developing software projects. A productivity game was created in the form of a competitive source control plug-in that rewards technical debt-reducing actions. The plug-in was tested by simulating source control usage with in a small sample project. Analysis showed that the plug-in appropriately assigned scores to developers for debt-reducing and debt-increasing actions. The plug-in has potential practical applications in the management of technical debt in workplace environments. The approach described in this paper is promising, and in future work we plan to test the Build Game plug-in with a wider variety of existing and simulated projects. Additional research is also planned to investigate the impact of the Build Game plug-in on workplace productivity.

Date: Monday, Nov 5 2012 - 4:10 pm
Location: EPS 108


Medical Robotics and Computer-Integrated Interventional Medicine

Presenter: Dr. Russell Taylor, Johns Hopkins University

Abstract: This talk will discuss ongoing research at the JHU Engineering Research Center for Computer-Integrated Surgical Systems and Technology  (CISST ERC) to develop systems that combine innovative algorithms, robotic devices, imaging systems, sensors, and human-machine interfaces to work cooperatively with surgeons in the planning and execution of surgery and other interventional procedures.  This talk will describe past and emerging research themes and illustrate them with examples drawn from our current research activities in medical robotics and computer-integrated interventional systems.

Bio: Russell H. Taylor received his Ph.D. in Computer Science from Stanford in 1976.  He joined IBM Research in 1976, where he developed the AML robot language and managed the Automation Technology Department and (later) the Computer-Assisted Surgery Group before moving in 1995 to Johns Hopkins, where he is a the John C. Malone Professor of Computer Science with joint appointments in Mechanical Engineering, Radiology, and Surgery and is also Director of the Engineering Research Center for Computer-Integrated Surgical Systems and Technology (CISST ERC).  He is the author of over 275 peer-reviewed publications, a Fellow of the IEEE, of the AIMBE, of the MICCAI Society, and of the Engineering School of the University of Tokyo.  He is also a recipient of numerous awards, including the IEEE Robotics Pioneer Award, the MICCAI Society Enduring Impact Award, and the Maurice Müller Award for Excellence in Computer-Assisted Orthopaedic Surgery.

Date: Friday, Oct 26 2012 - 3:10 pm
Location: Roberts 101


DOSI: Training Artificial Neural Networks using Overlapping Swarm Intelligence with Local Credit Assignment

Presenter: Nathan Fortier

Description: A novel swarm-based algorithm is proposed for the training of artificial neural networks. Training of such networks is a difficult problem that requires an effective search algorithm to find optimal weight values. While gradient-based methods, such as backpropagation, are frequently used to train multi-layer feedforward neural networks, such methods may not yield a globally optimal solution. To overcome the limitations of gradient-based methods, evolutionary algorithms have been used to train these networks with some success. This paper proposes an overlapping swarm intelligence algorithm for training neural networks in which a particle swarm is assigned to each neuron to search for that neuron's weights. Unlike similar architectures, our approach does not require a shared global network for fitness evaluation. Thus the approach discussed in this paper localizes the credit assignment process by first focusing on updating weights within local swarms and then evaluating the fitness of the particles using a localized network. This has the advantage of enabling our algorithm's learning process to be fully distributed.

Date: Monday, Oct 22 2012 - 4:10 pm
Location: EPS 108


Taxonomic Dimensionality Reduction in Bayesian Text Classification

Abstract: Lexical abstraction hierarchies can be leveraged to provide semantic information that characterizes features of text corpora as a
whole.  This information may be used to determine the classification utility of the dimensions that describe a dataset.  This paper presents a
new method for preparing a dataset for probabilistic classification by determining, a priori, the utility of a very small subset of taxonomically-related dimensions via a Discriminative Multinomial Naïve Bayes (DMNB) process.  We show that this method yields significant improvements over both DMNB and Bayesian network classifiers alone.

Date: Monday, Oct 15 2012 - 4:10 pm
Location: EPS 108


The Future of Computing

Presenter: Professor Peter A. Freeman, Emeritus Dean and Professor, Georgia Institute of Technology; (Former) Assistant Director of NSF for CISE.

Abstract:  Without knowing the past, it is difficult to know the future, most especially in computing.  We will take a look at some of the developments of the past fifty years, including hardware such as supercomputers and the PC, the Internet, Google, PDA's such as the iPhone, and massive software systems and ask the question: What is common among these developments?  In the process, we will look at some of the individuals that were key to these developments and ask the question: What did these people do about the future, and how did they do it?  

Date: Friday, Oct 12 2012 - 4:10 pm
Location: Roberts 113


Entrepreneurship

Presenter: Greg Gianforte, creator and CEO of RightNow Technologies.  The company was founded in 1997 in Bozeman and acquired by Oracle in 2011 for more than $1.8 billion.  RightNow Technologies created more than 1000 new jobs.

Abstract: After making a few remarks on the topic of entrepreneurship, the rest of the period will be devoted to questions and answers.

Date: Monday, Oct 8 2012 - 4:10 pm
Location: EPS 108


An Experimentalist’s Perspective on Biological Complexity and Computation

Presenter: Eric D. Smidansky, D.D.S., Department of Biochemistry and Molecular Biology, Penn State University, University Park, PA

Abstract: It is increasingly apparent that biological systems are much more functionally complex than previously imagined. The highly integrated nature of biological systems presents severe challenges to scientists seeking to understand them and, even more, to manipulate their outputs. These challenges are clearly revealed by the large overestimation of the expected ability to interpret and exploit human genomic information. The integrated nature of biological systems is indicated, for example, by the difficulties in understanding how hepatitis C virus achieves and maintains persistent infection in humans. Perhaps the best way to appreciate the sources of biological complexity and to get a sense of its scope is to begin at the lowest hierarchical level, that of an individual biological macromolecule. This talk will consider biological complexity by examining, as a starting point, a fundamental type of biological system, a nucleic acid polymerase, which is a member of an extremely interesting family of protein enzymes. Experiments will be described that provide insight into biological meaning, illustrate why computational approaches are finding explosively growing application to the study of biological systems and also help to define some of the substantial obstacles computational biology faces.

Date: Monday, Oct 1 2012 - 4:10 pm
Location: EPS 108


The Western Transportation Institute – An Overview, Including Connections to Computer Science

Description: Founded in 1994 by the Montana and California Departments of Transportation in cooperation with MSU, the Western Transportation Institute (WTI) is charged with advancing the field of transportation and developing the next generation of professionals by conducting cutting-edge, multidisciplinary research. Our concentration has been rural transportation research, yet many of our projects have been applicable in urban areas. A surprising amount of work at WTI involves software, hardware and communication systems development and integration, and we have employed quite a few computer science students and graduates over the past ten years. These staff members have applied their expertise to the development of technology for intelligent transportation systems to address the challenges for rural areas. In this presentation, we will provide an overview of WTI with an emphasis on connections to computer science.

Date: Monday, Sep 10 2012 - 4:00 pm
Location: EPS 108


Welcome Seminar

Description: Important general information about the department will be shared with graduate students.  Graduate students will have the opportunity to meet other students, the faculty and the staff.  Organizer: John Paxton

Date: Monday, Aug 27 2012 - 4:10 pm


Bayesian Approaches to Musical Instrument Classification Using Timbre Segmentation

Description: PHD Comprehensive Examination

Presenter: Patrick J. Donnelly

Abstract: The task of identifying musical instruments in an audio recording is a difficult  problem.  While there exists a body of literature on single instrument identification, little research has been performed on the more complex, but real-world, situation of more than one instrument present in the signal.  This work proposes a Bayesian method for multi-label classification of musical instrument timbre.  

Preliminary results demonstrate the efficacy of Bayesian networks on the single instrument classification problem. Peak spectral amplitude in ten frequency windows were extracted for each of twenty time windows to be used as features. Over a dataset of 24,000 audio examples covering the full musical range of 24 different common orchestral instruments, four different Bayesian network structures, including  na"{i}ve Bayes, were examined and compared to two support vector machines and a  $k$-nearest neighbor classifier.  Classification accuracy was examined by instrument, instrument family, and dataset size.  Bayesian networks with conditional dependencies in the time and frequency dimensions achieved 98% accuracy in the instrument classification task and 97% accuracy in the instrument family identification task. These results demonstrated a significant improvement over the previous approaches in the literature on this dataset.  

The remainder of this proposal outlines my approach for the identification of musical instrument timbre when more than one instrument is present in the signal. First signature matching Bayesian networks will be trained on single instruments to recognize the timbral signature of individual instruments.  Secondly, those signatures will be used to extract the features relevant to a single instrument from the spectral analysis of a multi-instrument signal.  Finally, a binary-relevance Bayesian classifier will determine if each specific instrument is present in the signal.  

This system proposes a novel approach to template matching allowing for probabilistic segmentation of musical spectra.  Furthermore the proposed approaches outline a novel approach to multi-label classification of music instrument timbre which supports both harmonic and inharmonic instruments, scales to a large number of musical instruments, and allows for efficient classification of new examples given the trained models.

Date: Friday, May 25 2012 - 9:00 am
Location: CS Conference Room


AY 2011-2012 Awards Ceremony

Presenter: John Paxton, Ph.D.

Abstract: Awards will be awarded, and refreshments/snacks will be provided!

Date: Monday, Apr 23 2012 - 4:10 pm
Location: EPS 108


Wireless Vehicular Ad Hoc Networks: Standards, Protocol Design, and Research Challenges

Presenter: Qing Yang, Ph.D.

Abstract: This talk covers the fundamental technologies and most recent progress in wireless vehicular ad hoc networks (VANETs). In a first part, we investigate the requirements on vehicle-to-vehicle (V2V) communications, ranging from traffic information systems to safety applications with real-time communication constraints. Typical V2V approaches are introduced including fully distributed as well as infrastructure-based, and centralized 3G/4G solutions. Emphasis is laid on the most recent standardization activities in the DSRC/WAVE context. We continue to discuss relevant protocols and communication principles to provide detailed information on which communication methods can be applied and how V2V protocols are developed. We study ad hoc routing approaches and their limitations to cover wide areas as well as recent geo-routing and broadcast-based data dissemination techniques. At the end of this talk, potential research problems and open issues in vehicular networks will be addressed.

Date: Monday, Apr 9 2012 - 4:10 pm
Location: EPS 108


The Life and Intelligence of Alan Turing

Presenter: Denbigh Starkey, Ph.D.

Date: Monday, Apr 2 2012 - 4:10 pm
Location: EPS 108


Image Recognition and Feature Detection in Solar Physics

Presenter: Piet Martens, Ph.D., Physics and Computer Science Departments, Montana State University, Bozeman, MT, Smithsonian Astrophysical Observatory, Cambridge, MA

Abstract: The NASA funded Solar Dynamics Observatory (SDO) Feature Finding Team (FFT) consists of collaborating institutions from Europe and the US that work on developing feature finding modules.  SDO produces about 1.5 terabytes of data per day, including over 80.000 full disk images from the AIA telescopes.  The SDO data repository will dwarf the archives of all previous solar physics missions put together.  NASA recognized early on that the traditional methods of analyzing the data -- solar scientists and grad students in particular analyzing the images by hand -- would simply not work and tasked our team with developing automated feature recognition modules for solar events and phenomena likely to be observed by SDO.  Having these metadata available on-line will enable solar scientist to conduct statistical studies involving large sets of events that would be impossible now with traditional means.

Mindful of the considerations mentioned above we have followed a two-track approach in our project:  we have been developing some existing task-specific solar feature finding modules to be "pipe-line" ready for the stream of SDO data, plus we are designing a few new modules.  Secondly, we took it upon us to develop an entirely new CBIR "trainable" module that would be capable of identifying different types of solar phenomena starting from a limited number of user-provided examples.  Both approaches are now reaching fruition, and in my lecture I will show solar movies with results from several of our feature finding modules overlaid.

In the remainder of my presentation I will focus on our CBIR module, which is the most innovative in character.  I will describe its design, early results, and highlight some unexpected but very illuminating aspects.  First there is the strong similarity between solar and medical X-ray images with regard to their texture, which has allowed us to apply some advances made in medical image recognition.  Second, we have found that there is a strong similarity between the way our trainable module works and the way our brain recognizes images.  As is well known, the human brain does not store entire images like an image archive would. Instead the brain extracts from the images key characteristics (say, color, outline, texture) and stores these data in different centers for each characteristic.  The brain can quickly recognize similar images from similar key characteristics, just as our code does.  We conclude from that that our approach represents the beginning of a more human-like procedure for computer image recognition.

Date: Monday, Mar 26 2012 - 4:10 pm
Location: EPS 108


Digital Isolation in the Developed World: Extending Internet Innovation to Rural and Mountain Communities

Presenter: Mike Wittie, Ph.D.

Abstract: Increasingly, full participation in modern society requires access to communication services like Twitter, Skype, and even online games, such as World of Warcraft. Future versions of such services will rely heavily on interactive communication streams, including voice, video, and control traffic, as part of their functionality. Unfortunately, the existing designs of application infrastructure struggle to deliver low-latency services users enjoy, gradually isolating rural communities from the modern society.

While the reach of the Internet has grown dramatically in recent decades, the delivered quality of online services has significant geographic variation. To benefit from the economies of scale, services are instantiated on large datacenters often quite distant from rural users. Physical separation, made worse by indirect Internet routes, introduces significant delays to rural users’ packets that impairs interactive service usability.

Without rethinking how interactive services are delivered, the indiscriminate infrastructure expansion along existing designs cannot reduce such distance-induced delays. Unless the problem of latency disparity is addressed, the future will see the expansion of interactive communications to new types of services that will be available only to a dwindling percentage of well-connected urban users.

In this talk I will present my research vision and recent results towards improving access to interactive services in rural and mountain communities.

Date: Monday, Mar 19 2012 - 4:10 pm
Location: EPS 108


Sensitivity Analysis of Continuous Time Bayesian Networks Using Perturbation Realization

Description: PhD Qualifying Exam Presentation

Presenter: Liessman Sturlaugson

Abstract: The continuous time Bayesian network (CTBN) model can be thought of as a factored Markov process. Sensitivity analysis of a Markov process is done by calculating partial derivatives of a user-defined performance measure with respect to changes in the transition intensities of the Markov process. Sensitivity analysis has yet to be applied to the CTBN model. This talk presents work on one method for Markov process sensitivity analysis called perturbation realization which works on a sample path of the process. In a CTBN, the number of states is exponential in the number of nodes, making it difficult to create a sample path of the entire model that visits every state. We show how to exploit the conditional independence structure of the model to build sample paths and compute performance measure derivatives independently for different subnetworks.

Date: Monday, Feb 27 2012 - 4:10 pm


A Comparative Evaluation of Automated Solar Filament Detection

Description: PhD Qualifying Exam Presentation

Presenter: Mike Schuh

Abstract: We present a comparative evaluation for automated filament detection in H-alpha solar images. By using metadata produced by the Advanced Automated Filament Detection and Characterization Code (AAFDCC) module, we adapted our Trainable Feature Recognition (TFR) component to accurately detect regions in solar images containing filaments. We first analyze the module's metadata and then transform it into labeled datasets for machine learning classification. Visualizations of data transformations and classification results are presented and accompanied by statistical findings. Our results confirm the reliable event reporting of the AAFDCC module as well as our ability to effectively detect solar filaments with our TFR component.

Date: Monday, Jan 23 2012 - 4:10 pm
Location: EPS 108


Other Years:   2014, 2013, 2011