BYU CS Graduate News

Michael R. Smith's MS Thesis Defense/PhD Qualifying Presentation

Computer Science Department — Wed, 04 Nov 2009 11:53:30 -0700

An Empirical Study of Instance Hardness

ABSTRACT:

Most widely accepted measures of performance for learning algorithms, such as accuracy and area under the ROC curve, provide information about behavior at the data set level. They say nothing about which instances are misclassified, whether two learning algorithms with the same classification accuracy on a data set misclassify the same instances, or whether there are instances misclassified by all learning algorithms. These questions about behavior at the instance level motivate our empirical analysis of instance hardness, a measure of expected classification accuracy for an instance. We analyze the classification of 57 data sets using 9 learning algorithms. Of the over 175000 instances investigated, 5% are misclassified by all 9 of the considered learning algorithms, and 15% are misclassified by at least half. We find that the major cause of misclassification for hard instances is class overlap, manifested as outliers and border points which can be exacerbated by class skew. We analyze these causes and show to what extent each leads to misclassifications, both in isolation and jointly. 19.8% of all misclassified instances are outliers, 71.3% are border points, and 21% belong to a minority class. We also find that 91.6% of all outliers and 38.3% of all border points are misclassified whereas only 3.5% of instances without class overlap are misclassified. We propose a set of heuristics to predict when an instance will be hard to correctly classify. Additionally, we analyze how different learning algorithms perform on tasks with varying degrees of outliers, border points and class skew.

Daniel Adams' MS Thesis Defense

Computer Science Department — Wed, 04 Nov 2009 11:47:56 -0700

Feature-Based Interactive Terrain Sketching

ABSTRACT:

Procedural generation techniques are able to quickly and cheaply produce large areas of terrain. However, these techniques produce results that are not easily directable and often require artists to edit the results by hand to achieve the desired layout. This paper proposes a sketch-based system for controlling fractal terrain that allows for a wide variety of terrain feature types. Artists sketch features rather than constrained points or elevations. The system is interactive, provides quick on-demand previews of the terrain, and allows for iterative design modifications. Interaction between features is handled in a realistic fashion. An arbitrary vertex insertion order midpoint displacement algorithm is also described which provides the necessary flexibility and constraints for the terrain generation system.

Tim van der Horst's PhD Dissertation Defense

Computer Science Department — Mon, 26 Oct 2009 11:24:27 -0600

Convenient Decentralized Authentication using Passwords

ABSTRACT:

Proof of email address ownership is typically required to create an account and to reset a password when it is forgotten. Despite its shortcomings (e.g., latency, vulnerability to passive attack), this approach is a practical solution to the difficult problem of authenticating strangers on the Internet.

This research utilizes this emergent, lightweight relationship with email providers to offload primary user authentication from service providers; thus reducing the need for service provider-specific passwords. Our goal is to provide decentralized authentication that maintains the convenience and portability of passwords, while improving its assurances (especially against phishing). Also, as existing approaches to online decentralized authentication are typically geared for web-based logins, our new protocols are designed to unify user authentication across the application and network (especially wireless) layers.

Kenneth Sundberg's PhD Dissertation Proposal

Computer Science Department — Wed, 30 Sep 2009 16:35:25 -0600

Partition Based Phylogenetic Search

ABSTRACT:

Evolutionary relationships are of critical importance in modern biology. Thus the process of inferring these relationships, phylogenetic search, is an important problem. This problem is known to be NP-Hard, and so heuristic methods are needed. This work proposes a new understanding of the solution space and exploiting this understanding looks to develop new search techniques.

Kenneth Sundberg's Research Area Exam

Computer Science Department — Wed, 30 Sep 2009 16:25:51 -0600

ABSTRACT:

Phylogenetic Search, a computationally difficult problem (NP-Hard), is of use to a wide variety of biological problems. This problem has been attacked through a number of heuristic methods,most notably hill climbing. Other heuristics such as genetic algorithms and simulated annealing have also been applied. However, work continues as there remain problems of interest that lie outside of current capabilities. A new direction in the field is the study of treespace as a whole, in an effort to devise new search techniques. This review contains a discussion of all these techniques, the major programs which implement them, and an in depth review of the most commonly used algorithm TBR.

Heather Chan's MS Thesis Proposal

Computer Science Department — Tue, 29 Sep 2009 16:31:38 -0600

Gene Network Inference and Expression Prediction Using Recurrent Networks

ABSTRACT:

We propose the use of recurrent neural networks to perform gene network in- ference and gene expression prediction. Recurrent neural networks are mathematical tools for nonlinear statistical data modeling, and they can represent important tem- poral information that is crucial to the problem of modeling gene regulation. We use a hybrid of particle swarm optimization and differential evolution to overcome the classic obstacle in training recurrent neural networks, which is the tendency to get stuck in local minima during training. We also adapt a previous statistical method for extracting network regulations from a dynamic Bayesian model and apply it to our recurrent neural network model to help discover biological meanings in inferred networks. Success in the modeling of gene regulation and prediction of gene expres- sion will lead to more rapid research and development of effective medicines, earlier diagnosis and treatment of adverse conditions, and vast advancements in the field of biology. Preliminary results on the SOS Repair dataset show striking success in accurate prediction of gene expression levels.

Derek Bunn's MS Thesis Proposal

Computer Science Department — Mon, 14 Sep 2009 14:59:50 -0600

Interactive Television News

ABSTRACT:

We propose a way to modify television news to make it interactive for viewers so they can get more of what they want and less of what they don’t want. This solution should be intuitive for users and easily fit into existing news production processes. Viewers will care because they will enjoy watching the news more. News producers will care because they can provide more transparent journalism and have viewers enjoy their news more. Our solution is to augment the existing news broadcast structure in the following ways: add a video headlines menu, provide on-demand access to additional story content, and provide interactive navigation controls between stories. For news producers we will create a video annotation program to simplify creating the interactive news. The production tools will be evaluated by using them in a news production room for a week and getting feedback from producers. We will also evaluate the home interactivity by having viewers provide feedback after watching the interactive news produced during that week.

Kristine Monteith's PhD Dissertation Proposal

Computer Science Department — Fri, 04 Sep 2009 17:30:01 -0600

Automatic Music Classification and Generation by Emotional and Physiological Responses

ABSTRACT:

The ability to automatically classify music is one of the vital elements of any music information retrieval system. Anyone interested in selling, buying, or researching music could benefit from advances in this area. This work intends to focus on classifying music by the physiological responses it produces in humans. It will employ biofeedback techniques to measure how specific acoustic features affect responses such as breathing, heart rate, and skin temperature. It will also investigate how these physiological responses, as well as a number of other descriptive features, can be used in computerized classification of music by genre and in determining musical similarity. The task of genre classification will be further explored by not only considering large numbers of features, but making refinements to the genre labels themselves using unsupervised learning strategies and dimensionality reduction techniques. This work will also investigate possibilities for computational creativity by using the information obtained through these experiments to generate music tailored to a specific mood.

Kristine Monteith's Research Area Exam

Computer Science Department — Fri, 04 Sep 2009 17:22:12 -0600

(Please also attend her Dissertation Proposal on Thursday, September 10, 2009 at 12:00 pm in the CS Conference Room (3365 TMCB). See Proposal announcement for details.)

Neha Rungta's PhD Dissertation Defense

Computer Science Department — Tue, 01 Sep 2009 16:09:32 -0600

Guided Testing for Automatic Error Discovery in Concurrent Software

ABSTRACT:

The quality and reliability of software systems, in terms of their functional correctness, critically relies on the effectiveness of the testing tools and techniques to detect errors in the system before deployment. A lack of testing tools for concurrent programs that systematically control thread scheduling choices has not allowed concurrent software development to keep abreast with hardware trends of multi-core and multi-processor technologies. This motivates a need for the development of systematic testing techniques that detect errors in concurrent programs.

The work in this dissertation presents a potentially scalable technique that can be used to detect concurrency errors in production code. The technique is a viable solution for software engineers and testers to detect errors in multi-threaded programs before deployment. We present a guided testing technique that combines static analysis techniques, systematic verification techniques, and heuristics to efficiently detect errors in concurrent programs. An abstraction-refinement technique lies at the heart of the guided test technique. The abstraction-refinement technique uses as input potential errors in the program generated by imprecise, but scalable, static analysis tools. The abstraction further leverages static analyses to generate a set of program locations relevant in verifying the reachability of the potential error. Program execution is guided along these points by ranking both thread and data non-determinism. The set of relevant locations is refined when program execution is unable to make progress. The dissertation also discusses various heuristics for effectively guiding program execution. We implemented the guided test technique to detect errors in Java programs. Guided test successfully detects errors caused by thread schedules and data input values in Java benchmarks and the JDK concurrent libraries for which other state of the art analysis and testing tools for concurrent programs are unable to find an error.