9th Undergraduate Research Symposium (2015)
Held 10 April 2015 in 011 Sitterson Hall
3:40-4:40 Session 1
“Testing the Effectiveness of AQMs at Improving Network Performance” by Ryan Doyle, supervised by Kevin Jeffay
“Evaluating Net-score as a Measurement Platform for Broadband Access Performance” by James Martin, supervised by Jasleen Kaur
“Scene Generation and Visualization for an HDR Camera Sensor” by Colin Arnott, supervised by Montek Singh and Leandra Vicci
4:50-5:30 Session 2
“Exploring Design Space of Schemes for Detecting Programming Difficulty from Interaction Logs” by Kevin Wang, supervised by Prasun Dewan
“A Testbed for Evaluating and Visualizing Facebook Friend-List Recommendation” by Ziyou Wu, supervised by Prasun Dewan
5:35-6:15 Session 3
“Design Optimization Algorithms for Concentric Tube Medical Robots” by Cenk Baykal, supervised by Ron Alterovitz
“A Framework for Incorporating DTI Atlas Builder Registration into Tract-Based Statial Statistics and a Simulated Comparison to Standard TBSS” by Matt Leming, supervised by Martin Styner
6:20 Reception — hors d’oeuvres will be served
Presentation Abstracts and Presenter Bios
Ryan Doyle, supervised by Kevin Jeffay
As use of the Internet and the number of real-time applications that depend on it increases, network performance statistics such as link utilization and average response time become more important. Today, most routers use a basic drop tail style of queueing in which the queue is unmanaged. When the queue is full, every packet arriving at the router is dropped. TCP, the Internet’s dominant transport protocol, has a congestion control mechanism that responds to dropped packets by limiting the rate at which the affected connection sends packets. If many packets from different TCP connections are dropped by a router at the same time, all of those connections will limit their throughput, creating the potential for significant loss in link utilization. In addition to this link utilization problem, drop tail queueing allows for the potential of a buffer to fill and remain full, adding delay to every packet that arrives at the router.
Active queue management algorithms (AQMs) run on routers, replace the standard drop tail method of queueing, and manage the queue before it becomes full. By selectively dropping or marking packets before the router becomes full, AQMs theoretically limit buffer-related delay and increase link utilization by preventing the dropping of packets from many different TCP connections at the same time. In our research, we have constructed and calibrated a network designed to replay actual Internet traffic in a laboratory setting. By implementing different AQMs on the routers in our network, we can test the effectiveness of these algorithms compared to drop tail queuing and will report on the results of these experiments.
Ryan Doyle grew up in New Bern, NC. He spent much of his childhood sailing, but, as an Eagle Scout, also enjoys hiking and camping. Just prior to high school, Ryan taught himself how to program. He worked for OptoSonics, Inc. on photoacoustic angiography of the breast and was published in Medical Physics in 2010. He then attended the North Carolina School of Science and Mathematics before deciding to attend UNC Chapel Hill. Ryan is currently a BS/MS student researching the effects of active queue management algorithms on network performance.
James Martin, supervised by Jasleen Kaur
Net-score is a new platform for measuring broadband access connections for Internet users. In this paper, we outline the architecture of the net-score measurement platform and discuss its advantages over current state-of-the-art. We also describe how we have evaluated net-score through a seven month long case study, describe the data we collected, and show how net-score reaches its goal of being a low-cost and low barrier to entry method of measuring general broadband performance.
James Martin is a 5-year B.S./M.S. student in his 8th semester from Washington, D.C. He will be working in IBM’s research lab in Austin, Texas as an Extreme Blue intern this summer before returning in the Fall to finish his Master’s Degree. Outside of the classroom, James gives tours for the university, is an avid cyclist, and loves vinyl records played over quality speakers. Hoping to work in the field of data science or system engineering, James is looking forward to sharpening his skills as a researcher and computer scientist before entering the workforce full time.
Colin Arnott, supervised by Montek Singh and Leandra Vicci
We are developing a novel camera sensor that provides frameless capture, and has significantly higher dynamic range (20 bits), finer color sensitivity, and lower noise as compared to the current state-of-the-art sensors. The strength of the approach lies not in developing new types of photodetectors or analog circuitry, but in the manner in which information is extracted from the pixel sensor, transported to the processing logic, and processed to yield intensity values. In this talk, I will focus on my contribution to this project: creating synthetic scenes to evaluate the sensor design. Since current commodity cameras offer only 8-12 bits of dynamic range, real-world photographs are inadequate for use as source scenes to demonstrate the capabilities of our sensor. I have used the open-source Blender framework for generating ray-traced scenes with 20-30 bits of dynamic range, and the open-source EXR framework for developing a visualization tool.
Colin Arnott is a senior undergraduate student at UNC majoring in Computer Science. This semester he participated in mentored research with Montek Singh and Leandra Vicci, helping them on their on their high dynamic range image processor design. His interests include in security, computer hardware and cryptography and he has worked in internships at IBM and Moody’s Investor Service.
Kevin Wang, supervised by Prasun Dewan
This paper will utilize classifiers and surrounding toolkits to reliably mine programming activity in Eclipse to promote difficulty detection. Such program can be used effectively in various settings to promote learning and pinpoint the programming issues. We hope to achieve this by examining input vectors of user actions taken directly from Eclipse IDE. Out of the five major chosen categories of input attributes, the number of website accesses and the frequency of debugging are deemed to be the most critical to difficulty detection. The classification algorithm explored in the process are J48, Decision Stump, Adaboost.M1, and Bagging. The classification algorithms’ shortcomings and strengths are all explored with regards to the true positive, true negative, and overall prediction accuracy. J48 and tree-base algorithms excel predicting the majority class, the true negative, while Decision Stumps exceeds the capability of J48 at predicting the minority class. Finally, statistical analysis of the attribute distribution is conducted to discover any statistical anomaly.
Kevin (Kun) Wang is currently a senior undergraduate student majoring in computer science and minoring in chemistry. With a biology, chemistry, and computer science background, he is interested in statistical learning and the manner in which it can be used to solve common problems across all the aforementioned disciplines.
Ziyou Wu, supervised by Prasun Dewan
This project involves implementing a named-group prediction algorithm in Facebook and creating analytics tools for analyzing and visualizing recommendation results. It not only focuses on recommending friend-lists but also studies how friend-lists change over time and how to predict such evolutions. It has three components: a recommendation system, a friend-list editor, and a tool for visualizing recommendations. The recommendation system implemented in Facebook mines end-users’ social graphs to make predictions on friend-lists. The friend-list editor allows editing recommendations and measures the efforts a user takes to create friend-lists with and without recommendations. The visualization tool provides a set of visualization methods of social graphs and friend-lists. We have planned user studies to evaluate both the friend-list editor and the visualization tool.
Ziyou “Will” Wu is a senior from Beijing, China, majoring in Computer Science and Mathematical Decision Sciences. He has a variety of research experience, two years under the mentorship of Prof. Prasun Dewan on Facebook Friend-lists Recommendation and one year and a half under the mentorship of Prof. Denis Tsygankov on Cell Image Processing. His career interest lies in data analytics and machine learning in the financial market. After graduation, he will be joining in a bank as a technology analyst.
Cenk Baykal, supervised by Ron Alterovitz
Concentric tube robots are tentacle-like surgical robots that can bend around anatomical obstacles to access hard-to-reach surgical targets. These robots have potential to enable safer, minimally invasive interventions to many sites in the human body. Concentric tube robots can be built to facilitate easy swapping out of their components to deeply customize their behavior. Customizing the design of these robots can potentially enable different surgical procedures on a variety of patients. In this presentation, I will present design optimization algorithms for concentric tube robots capable of generating a single or sets of concentric tube robot designs that can collectively maximize the reachable volume of a given goal region in the human body. I will summarize mathematical analysis proving the asymptotic optimality of our algorithm used to optimize a single design and results supporting the favorable algorithmic properties of our method. I will also show results demonstrating the effectiveness of our algorithms in a clinical scenario involving minimally invasive lung biopsy for early-stage lung cancer diagnosis.
Cenk Baykal is a senior undergraduate student majoring in Computer Science and Mathematics. In addition to his coursework, he enjoys participating in undergraduate research in Robotics and Artificial Intelligence (AI). He is particularly interested in developing algorithms based on sound mathematical foundations that have potential to enable or facilitate the use of intelligent agents to complete a wide variety of tasks in a successful way. After he graduates from UNC in May, he will be a Ph.D. student at MIT in the Computer Science and Artificial Intelligence Laboratory.
A Framework for Incorporating DTI Atlas Builder Registration into Tract-Based Statial Statistics and a Simulated Comparison to Standard TBSS
Matt Leming, supervised by Martin Styner
Tract-based spatial statistics (TBSS) is a software pipeline widely used in comparative analysis of the white matter integrity of groups of diffusion tensor imaging datasets. However, several components of TBSS have been criticized over the years. Many of these criticisms stem, in part, from its white matter skeletonisation and projection and the lack of directional data in its calculations. Nonetheless, criticisms of TBSS have not provided a consistent method for comparing TBSS to its counterparts, and these criticisms similarly ignore the need for convenient software tools in running TBSS, which contributes to its widespread use. In this study, I have altered the TBSS into a new framework, DAB-TBSS (DTI Atlas Builder Tract-Based Spatial Statistics) by using directionally-based registration offered by DTI Atlas Builder and incorporating blurring into the skeletal projection step, and I propose a framework for simulating differences in groups of diffusion tensor imaging data, providing a more substantive means of empirically comparing DTI group analysis programs such as TBSS.
Matthew Leming is a senior computer science major at UNC Chapel Hill. He works under Dr. Martin Styner in the Neuro Image Research and Analysis Laboratories to research applications of diffusion tensor imaging group comparison software.