Computer Science Master's Theses

Applying EM to compute document relevance from crowdsourced pair preferences

Jie Wu — Fri, 14 Jun 2013 09:25:18 PDT

Traditional Information Retrieval (IR) systems evaluate over absolute relevance judgments or multi- grade judgments by hiring expert assessors. More recently there has been considerable interest in inferring document relevance, for which no true relevance labels are available, based on noisy crowdsourcing data. We introduce an algorithm that, using Expectation Maximization (EM) and the Elo Rating System on pairwise preference judgments, acquires document relevance from crowdsourced data only. We collect pairwise ternary judgments from workers with unknown accuracy, calculate the pairwise probabilities of each document pair, and transform them into document relevance probabilities. Each judgment is presented in the form, given...

Checking binding hygiene statically

Fabian Muehlboeck — Thu, 13 Jun 2013 11:16:39 PDT

Macros are a powerful facility for various applications, from simple syntax extensions to creating domain specific languages. However, they have suffered from the fact that the structure of variable bindings in a program is only implicitly encoded in the program text—because of this, simple program transformations may lead to results not intended by either the macro writer or user by violating hygiene. We present the core details of a developing system (by Stansifer and Wand) to statically ensure the preservation of α-equivalence and therefore hygiene.

Our contribution comes in two parts. The first is that we finished the implementation of...

Automated indexing of stories for conversational health intervention

Ramesh Radhakrishna Manuvinakurike — Thu, 13 Jun 2013 09:31:38 PDT

Previous studies have shown that a large percentage of people search for someone else's experience with health behavior change. Previous studies have also shown that stories encoding health information are an effective tool for promoting health behavior change. By collecting health stories from blogs, social networks, public forums and other online resources we can create a database of stories that could be tailored to users for health interventions. In this thesis I explore the delivery of health stories tailored to users based on health behavior change theory, demographics and quality metrics. The medium of delivering these stories is also explored...

BlueSANE: integrating functional blueprints with neuroevolution

Jessica H. Lowell — Tue, 28 Aug 2012 05:31:26 PDT

Neuroevolution algorithms are an important tool for optimizing neural network design in the fields of control and machine learning. We seek to improve SANE, a classic machine learning algorithm, by optimizing the size of the hidden layer in the neural networks that it generates. We use a technique called functional blueprints that guide the self-organization of systems by specifying their desired behavior, in this case avoidance of over/underfitting. We performed experiments with a simulated double cart-pole balancing benchmark problem which indicate that BlueSANE improves performance by slight to moderate amounts compared to the original SANE algorithm.
...

A 45nm CMOS, low jitter, all-digital delayed locked loop with a circuit to dynamically vary phase to achieve fast lock

Soumya Shivakumar Begur — Thu, 01 Mar 2012 09:29:27 PST

The objective of the thesis is to address the problem of clock skew between two different clock domains in modern day microprocessors due to the process, voltage and temperature (PVT) variations. In order to mitigate the misalignment of the clocks in the different clock domains, a delay line is added in all but the reference clock domain. These delay lines add or subtract the delay (as necessary) to keep the clocks continuously aligned to a common reference clock delay. This ensures error free data transfer between any two clock domains.

A novel Digital Delay Locked Loop (DLL) design is proposed...

New server - centric authentication protocols for RFID applications with simulation based analysis

Kaushik Prakash Chavan — Thu, 01 Mar 2012 09:14:01 PST

RFID is likely to emerge as a very important tool in electronically documenting the physical world to generate powerful information. In order to label millions of objects with RFID tags economically, the cost of a single tag has to be inexpensive enough. An inexpensive tag costing just a few cents would not have the necessary infrastructure to secure the data it stores and transmits. RFID has been drawing attention from privacy advocates and skeptics for the loopholes it can create in the information security infrastructure and consumer privacy. To overcome these shortcomings, security measures such as public key cryptography have...

Automated explanation of research informed consent by embodied conversational agents

Rukmal Fernando — Wed, 29 Sep 2010 05:48:06 PDT

Research has shown that the comprehension of documents written by professionals for laypersons - such as Research Informed Consent forms - is a challenging task for many due to low literacy level, lack of prerequisite domain knowledge, and the use of arcane, complex language that is above the individual's reading level.

Prior work has demonstrated that an Embodied Conversational Agent can successfully explain Research Informed Consent documents to participants, resulting in higher comprehension and satisfaction compared to self-study of the document. However, this prior work relied on scripted explanation dialogue, limiting the ability of the system to be deployed across...

Fine-grained addressability to support large-scale collaborative document development

Viral Gupta — Fri, 18 Dec 2009 08:15:08 PST

The Web has made it possible for large, distributed collaborations to develop sophisticated document bases. As these collaborations increase in size, there is a need to support reference, navigation and search through the document bases by individuals who are not computer pro-fessionals. Although collaborative tools such as wikis have been developed that address some of the requirements of these communities, the tools lack support for simple and convenient fine-grained addressability to parts of the documents. Such addressability is essential for formal documents, such as standards and legal documents. In this thesis we develop a solution to the problem of fine-grained...

Annotating decision analyses using semantic web technologies

Vishesh Duggar — Fri, 18 Dec 2009 08:15:07 PST

In this thesis we use Semantic Web techniques for the formal representation of the decision analysis and the Eclipse Process Framework (EPF) to coalesce decision making process with a software engineering life cycle. In this way, a formal artifact can be produced that expresses how and why a decision was made, which can be linked into the many other artifacts of the development process. Semantic Web search tools may then be employed to find similar decisions so that they may be reused. [Duggar07] Semantic technology can have a significant impact on retrieval, reuse,interoperability, consistency, enforcement of policy and validation of...