Department of Computer Science and Electrical Engineering

Ph.D. Dissertation Proposal

Convexification/Deconvexification for Training Neural

Networks and Recurrent Deep Learning Machines

Yichuan Gui

9:30am Thursday, 16 May 2013, ITE 325b, UMBC

The development of artificial neural networks (ANNs) has been impeded by the local minimum problem for decades. One principle goal of this proposal focuses on devel- oping a methodology to alleviate the local minimum problem in training ANNs. A new training criterion called the normalized risk-averting error (NRAE) criterion is proposed to avoid nonglobal local minima in training multilayer perceptrons (MLPs) and deep learning machines (DLMs). Training methods based on the NRAE crite- rion are developed to achieve global or near-global minima with satisfactory learning errors and generalization capabilities.

Many advantages of DLMs have been analyzed in recent research works of ANNs, and effective architectures and training methods have been explored from those works. However, feedback structures are commonly ignored in previous research of DLMs. The next objective of this proposal is to develop recurrent deep learning machines (RDLMs) through adding feedback structures to deep architectures in DLMs. De- signing and testing works are expected to illustrate the efficiency and effectiveness of RDLMs with feedback structures comparing to feedforward DLMs.

Preliminary works presented in this proposal demonstrate the effectiveness of NRAE-based training methods in avoid nonglobal local minima for training MLPs. Methods based on the NRAE criterion will be tested in training DLMs, and the de- veloping and testing of RDLMs will be performed in subsequent works. Moreover, an approach that combining both the NRAE criterion and RDLMs will also be explored to minimize the training error and maximize the generalization capability. Contribu- tions of this proposed research are expected as (1) provide an effective way to avoid local minimum problem in training MLPs and DLMs with satisfactory performance; (2) develop a new type of RDLMs with feedback connections for training large-scale dataset efficiently; (3) apply the NRAE criterion to train RDLMs for minimizing training errors and maximizing generalization capabilities. Those contributions are expected to significantly boost research interests in ANNs' fields and stimulate new practical applications in the future.

Committee: James Lo (mentor), Yun Peng (mentor), Tim Finin, Tim Oates, Charles Nicholas

MS Defense

Social Media Analytics: Digital Footprints

Sandhya Krishnan

9:00am Monday, 13 May 2013, ITE325b

In this work we describe an approach to distinguish real and impostor/ compromised accounts on social media. Compromising a user's social media account is not only a breach of security, but can also lead to dissemination of misinformation at a fast pace on social media. There have been several such high profile attacks recently, including on Twitter feeds of AP, CBS, and Delta Airlines. A fake account for the Prime Minister's Office in India was used to spread malicious rumors last year. Our approach builds a profile or footprint of users using both the content of their tweets and the structure of their network. We analyze the real time content of users (Tweets, Facebook posts, etc.) and compare them with information about the user from reliable sources on the Web (e.g., newspapers, news channels, etc.) in order to compute a similarity metric between content from the two sources. We also compute a metric based on the social network analysis of the users: who connects to them, who they are connected with, and how central they are in their network. We have shown how such an approach can easily detect fake accounts for not just well known people such as President Obama, but also for lesser known people and organizations. We also show promising initial results on how this approach can be used to detect an account which has been hacked.

Committee: Anupam Joshi (chair), Tim Finin, Tim Oates, Ponnurangam Kumaraguru (IIIT Delhi)

On May 7, 2013, Dr. Richard Forno, Assistant Director of UMBC's Center for Cybersecurity and Director of UMBC's Graduate Cybersecurity Program, conducted an invited talk on the proposed Cybersecurity Information Sharing and Protection Act (CISPA) and moderated a discussion about general cybersecurity issues to UMBC's Delta Sigma Theta Sorority.  The evening event was the second in a series of invited guest speakers as part of the Sorority's May Week festivities.

CISPA is a proposed law that would allow and encourage the sharing of Internet traffic information between the U.S. government and technology and manufacturing companies in order to help US government agencies investigate cyber threats and ensure the security of networks against cyberattacks.

Delta Sigma Theta Sorority, Inc. is a private, non-profit organization whose purpose is to provide assistance and support through established programs in local communities throughout the world. A sisterhood of more than 200,000 predominately Black college educated women, the Sorority currently has over 900 chapters located in the United States, England, Japan (Tokyo and Okinawa), Germany, the Virgin Islands, Bermuda, the Bahamas and the Republic of Korea.

Ph.D. Dissertation Proposal

A Rapidly Deployable Image
Classification System Using Multi-Views

Adrian Rosebrock

11:00am Friday, 10 May, ITE 325, UMBC

Constructing an image classification system using strong, local invariant descriptors is time consuming and tedious, requiring many experimentations and parameter tuning to obtain an adequately performing model. Furthermore, training a system in a given domain and then migrating the model to a separate domain will likely yield poor performance. As computer vision systems become more prevalent in the academic, government, and private sectors, it is paramount that a framework to more easily construct these classification systems be created. In this work we present a rapidly deployable image classification system using multi-views, where each view consists of a set of weak global features. These weak global descriptors are computationally simple to extract, intuitive to understand, and require substantially less parameter tuning than their local invariant counterparts. We demonstrate that by combining weak features with ensemble methods we are able to outperform the current state-of-the-art methods or achieve comparable accuracy. Finally, we provide a theoretical justification for our ensemble framework that can be used to construct rapidly deployable image classification systems called "Ecosembles".

Committee: Dr. Tim Oates (chair), Dr. Jesus Caban, Dr. Tim Finin, Dr. Charles Nicholas

UMBC's cyber defense team, the Cyber Dawgs, will host an interdisciplinary talk and screening of the film Game of Pawns at 7:00pm on Tuesday, April 30, 2013 in room 102 of the ITE building (LH8). The film is a true story of an American student who was recruited by the Chinese government to infiltrate a U.S. intelligence agency.

The event is sponsored and run by InfraGuard, an organization that acts as a partnership mediator between the FBI and US businesses. The talk will be nontechnical and will present an overview of the dangers that might arise when dealing with foreign businesses or representatives. It should be of interest to students considering studying abroad, pursuing international relations or business, or anticipating working for a government agency.

Baltimore Chapter of Electron Devices and Solid-State Circuits

Quantum Engineering of Semiconductor Atomic Structures for Biosensing

Dr. Manijeh Razeghi
Center for Quantum Devices
Electrical Engineering and Computer Science
Northwestern University

5:30pm Tuesday, 30 April 2013
206 Technology Research Center, UMBC

5:30pm social hour, talk begins at 6:15pm. Free but please RSVP to by Monday, April 29

Nature offers us different kinds of atoms, but it takes human intelligence to put them together in an elegant way in order to realize functional structures not found in nature. III-V semiconductors are made of atoms from column III (B, Al, Ga, In, Tl) and column V (N, As, P, Sb, Bi) of the periodic table, and constitute a particularly rich variety of compounds with many useful optical and electronic properties. Guided by highly accurate simulations of the electronic structure, modern semiconductor optoelectronic devices are literally made atom by atom using advanced growth technology such as Molecular Beam Epitaxy (MBE) and Metal Organic Chemical Vapor Deposition (MOCVD). Recent breakthroughs have brought quantum engineering to an unprecedented level, creating light detectors and emitters over an extremely wide spectral range from 0.2 µm to 300 µm. Nitrogen serves as the best column V element for the short wavelength side of the electromagnetic spectrum, where we have demonstrated III-nitride light emitting diodes and photo detectors in the deep ultraviolet to visible wavelengths. In the infrared, III-V compounds using phosphorus, arsenic, or antimony from column V with indium, gallium, aluminum, or thallium from column III can create lasers and detectors based on quantum-dot (QD) or type-II superlattice (T2SL). These are fast becoming the choice of technology in crucial applications such as environmental monitoring and space exploration. Last but not least, on the far-infrared end of the electromagnetic spectrum known as the terahertz (THz) region, III-V semiconductors offer a unique solution of generating THz waves in a compact device at room temperature. Continued effort is being devoted to all of the above areas with the intention of developing smart technologies which meet the current challenges in environment, health, security, and energy. This talk will highlight contributions to the world of III-V semiconductor nano-scale optoelectronic devices from deep UV to THz.

Dr. Manijeh Razeghi received the Doctorat d'État es Sciences Physiques from the Université de Paris in 1980. After heading the Exploratory Materials Lab at Thomson-CSF (France), she joined Northwestern University in Evanston, IL, in the fall of 1991 as the Director of the Center for Quantum Devices, where she created undergraduate and graduate programs in solid-state engineering. Dr. Razeghi pioneered the development and implementation of major modern epitaxial techniques such as MOCVD, VPE, gas MBE, and MOMBE for the growth of entire compositional ranges of III-V compound semiconductors. She is on the editorial board of journals such as the Journal of Nanotechnology and the Journal of Nanoscience and Nanotechnology, and is an Associate Editor of the Opto-Electronics Review. Dr. Razeghi is on the International Advisory Board for the Polish Committee of Science, and is an Adjunct Professor at the College of Optical Sciences of the University of Arizona in Tucson. She ha s authored or co-authored more than 1000 papers, over 30 book chapters, and fifteen books, including the textbooks Technology of Quantum Devices and Fundamentals of Solid State Engineering, 3rd Edition. Two of her books, MOCVD Challenge Vol. 1 (1989) and MOCVD Challenge Vol. 2 (1995), discuss some of her pioneering work in InP-GaInAsP and GaAs-GaInAsP based systems. [The MOCVD Challenge, 2nd Edition (2010) represents the combined updated version of Volumes 1 and 2]. Dr. Razeghi holds 50 U.S. patents and has given more than 1000 invited and plenary talks. Her current research interest is in nanoscale optoelectronic quantum devices. Dr. Razeghi is a Fellow of MRS, IOP, IEEE, APS, SPIE, OSA, and the International Engineering Consortium (IEC), a Fellow and Life Member of the Society of Women Engineers (SWE), and a member of the Electrochemical Society, ACS, AAAS, and the French Academy of Sciences and Technology. She received the IBM Europe Science and Technology Prize in 1987, the Achievement Award from the SWE in 1995, the R.F. Bunshah! Award in 2004, and multiple best paper awards.

MS Defense

Modeling Individual Nodes In Dynamic Link Prediction

Maksym Morawski

2:00pm Thursday, 25 April 2013, ITE325b, UMBC

The question of how to predict which links will form in a graph, given the graph’s history, is an open research problem in computer science. There are many different approaches to the link prediction problem, one of which involves building a set of features for pairs of nodes and using supervised learning to build a model that predicts when these pairs of nodes will link. Typically, this model is learned over the entire graph. In this thesis, I investigate building this model over each individual node in an attempt to learn the particular ways in which that node behaves before making predictions about it. In addition, research into link prediction to date lacks intelligent ways of utilizing the graph over large timespans. To address this, I introduce a variety of ways to include temporality into the link prediction process by introducing new ways of using existing features.

Committee: Dr. Marie desJardins (Chair), Dr. Tim Oates, Dr. Tim Finin

MS Thesis Defense

A Hybrid CPU/GPU Pipeline Workflow System

Tim Blattner

11:45am Thurday, 25 April 2013, ITE 325b, UMBC

Heterogeneous architectures can be problematic to program on, particularly when trying to schedule tasks on all available compute resources, overlapping PCI express transfers, and managing the limited memory available on the architectures. In this thesis we propose a workflow system that is capable of scheduling on all available compute resources, overlaps PCI express transfers, and manages the limited memory. A procedure for creating the workflow system is described and two case studies are analyzed.

• Image Stitching, which implements the workflow system and achieves two orders of magnitude speedup over an image stitching plugin found in the popular Fiji ImageJ application. Implementing the image stitching algorithm without the workflow system yielded only one order of magnitude speedup over the image stitching plugin.
• Out of Core LU Decomposition, which does not implement the workflow system. This case study demonstrates the impact of the PCI express on a problem with a large number of dependencies. A proposed workflow system for this algorithm is provided in Future Work.

Using the workflow system, programmers have a method for scheduling any algorithm on all available compute resources and is capable of hiding the I/O impact by overlapping computation with I/O.

Committee Members: Milton Halem, Yelena Yesha, Shujia Zhou, John Dorband, Walid Keyrouz

CSEE Colloquium

Queuing and Long Lines: How to run efficient elections

Dr. William A. Edelstein
Visiting Distinguished Professor of Radiology
Johns Hopkins School of Medicine

1:00pm Friday, 3 May 2013, ITE227, UMBC

Computerized touchscreen "Direct Recording Electronic" (DRE) voting systems have been used by over 1/3 of American voters in recent elections, including in Maryland. In many places, insufficient DRE numbers, in combination with lengthy ballots and high voter traffic, have caused long lines and disenfranchised voters who left without voting. We have applied computer queuing simulation to the voting process and conclude that far more DREs, at great expense, would be needed to keep waiting times low. Alternatively, paper ballot-optical scan systems can be easily and economically scaled to prevent long lines and meet unexpected contingencies. We have developed a heuristic "Queue Stop Rule" that can be applied to prevent long lines at voting stations. We have also carried out queuing simulations of other parts of the voting process, for example, voter check-in and ballot scanning. Our results can be used to help plan cost-effective election systems that will produce expeditious elections.

William Edelstein, physicist, received BS and PhD degrees in that subject from University of Illinois and Harvard, respectively. His career has principally focused on the development of MRI, starting in Scotland in 1977 and continuing from 1980 at the GE research lab in Schenectady, NY. He has been Visiting Distinguished Professor of Radiology at Johns Hopkins School of Medicine since 2007. His MRI work has been recognized with many honors, including the 2005 Industrial Applications of Physics Prize from the American Institute of Physics. His interest in election systems began several years ago in NY State during the debate to replace lever voting machines.

MS Thesis Defense

Text and Ontology Driven
Clinical Decision Support System

Deepal Dhariwal

9:00am Tuesday 23 April 2013, ITE325b, UMBC

This thesis discusses our ongoing research in the domain of text and ontology driven clinical decision support system. The proposed framework uses text analytics to extract clinical entities from electronic health records and semantic web analytics to generate a domain specific knowledge base (KB) of patients’ clinical facts. Clinical Rules expressed in the Semantic Web Language OWL are used to reason over the KB to infer additional facts about the patient. The KB is then queried to provide clinically relevant information to the physicians. In the first phase, standard text pre processing techniques such as section tagging, dependency parsing, gazetteer lists are used filter clinical terms from the raw data.

In the second phase, a domain specific medical ontology is used to establish relation between the extracted clinical terms. The output of this phase is a Resource Description Framework KB that stores all possible medical facts about the patient. In the final phase, an OWL reasoner and clinical rules are used to infer additional facts about patient and generate a richer KB. This KB can then be queried for a variety of clinical tasks. To demonstrate a proof of concept of this framework, we have used discharge summaries from the cardiovascular domain and determined the TIMI Risk Score and San Francisco Syncope Score for a patient. The goal of this research is to combine factual knowledge about patients, procedural knowledge (clinical rules), and structured knowledge (medical ontologies) to develop a clinical decision support system.

Committee: Dr. Anupam Joshi (chair), Dr. Michael Grasso, Dr. Tim Finin, Dr. Yelena Yesha