As mobile touch screen devices become increasingly ubiquitous, they are also becoming more accessible to the visually impaired. Someone who is blind traditionally enters text on his or her phone by using a screen reader and the default keyboard layout. However, this method is slow and it is difficult to correct mistakes as one is typing. We developed Perkinput, a nonvisual text entry method that uses the 6-bit Braille character encoding. In Perkinput, instead of relying on the device’s voice over system to announce each letter that is pressed, this input is based solely on the Braille keyboard. The user simultaneously presses down with his or her fingers that correspond to the dots for the character in Braille he or she wanted to enter. Once entered, the text can be sent in a text message, an email, or simply copied to the clipboard for later use in other applications. Results have shown that the Perkinput method is more than twice as fast as using the default keyboard layout for the blind. I have implemented this method into an iOS application that is currently deployed on the iTunes Store. I have instrumented this application in order to gain insight into how people are using the program outside of a lab setting. This will help us improve Perkinput and make it more accessible to the blind community. Perkinput aims to make text entry easier and efficient for people who cannot look at their screen.

Computer face recognition and identification is an important problem with numerous applications in intelligent user interfaces, security, and accessibility. For example, identification of faces can be used for access to restricted areas, authentication without keyboards, and personalized interfaces. Traditional algorithms that solve this problem identify faces individually or use co-occurrence information for a given set of people (For example, person X is often photographed with person Y). However, such systems have many errors, and cannot generalize well to people the system has not seen together previously. In contrast, humans are much better at this because we have "common sense." We can look at a photograph and recognize that people tend to photograph with family, friends, and co-workers. We can further guess why the photograph (birthday, conference) was taken, and even where (person's house, a certain park). This abstract knowledge is very difficult for computer algorithms to learn because there are not sets of these "rules" available explicitly. In our work, we tackle this problem using structured relational graphs like Freebase, a large online graph of billions of facts about the world. Given many specific examples of people who are in photographs together and data about those photographs, we can see what connects these people on graphs like Freebase, and generalize to the "common sense" rules that are notoriously difficult to learn. Then, given a new image and possible candidates for identification of people in it, we can compute the most likely set of correct identifications. We can also hypothesize unknown faces to our existing faces. This allows us to learn, on-the-fly, a new recognizer for that specific person. Finally, by knowing relationships behind photograph subjects, we can discover and generate useful contextual information about the photograph, such as location and possible reason for the photograph. Broadly, this means that computer systems can more accurately identify meaning behind photographs; a problem that has been historically hard for computers to accomplish.

In the classical secretary problem, an employer tries to hire the best secretary from a line of candidates. The employer considers each candidate in sequence and makes an irreversible decision to accept or reject, being able to make only one hire. A well-known result gives the optimal strategy for this problem with a probability of success of 1/e. We extend this problem by considering a setting in which there are multiple employers, which could be universities, for example, competing for the best faculty candidates. We derive the optimal strategy for this setting and demonstrate that it meets expectations because lower ranked universities tend to make earlier offers. We employ theoretical analysis using game theory and algorithms. The broader implications of this research are greater understanding of optimal stopping problems as well as better information on how employers should hire. Eric's duties include design and analysis of algorithms for solving the problem.

Physical limitations have forced computer hardware designers into providing multiple processors rather than faster ones. To take advantage of this, software engineers must structure their programs in a parallel fashion. Separately, programming languages with commands similar to those of hardware, known as "systems languages", have long been criticized for their lack of safety. This means that programs written in these languages are easily incorrect and difficult to reason about. Rust, a new systems language from Mozilla, seeks to address both parallelism and safety by using modern language concepts. To those ends, Rust’s concurrency primitive is the task, which can be thought of as a lightweight thread that does not share memory with any other task. Message passing is used between these tasks in order to coordinate their work. While this primitive is sophisticated and suited to parallelism, the development community has expressed a desire for an Actor library. At a high level, the Actor model is a concurrent object-oriented programming model, where each Actor maintains state and takes actions based on the messages they receive from other Actors. In this investigation, the suitability of the Rust language for an Actor library is evaluated by writing and refining an implementation. It is expected that the implementation will be backed by the task primitive and its associated structures. Various criteria will be evaluated throughout development, including overall throughput, library overhead, and ease of use. The success of the Actor model within the Rust language would continue its goal of bringing straightforward parallelism and safety to the systems language space.

Sleep apnea is a common syndrome where a person has an irregular breath rate only during sleep, making it difficult to be detected and diagnosed. Yet, in the U.S. about two percent of children are affected by this syndrome and thus at risk of some health issues. There are some prevalent breath detection products on the market but they remain inaccessible due to the undesired prices and limited range of detection. In the breath detection project I was participating in, we intended to detect the breath rate by wireless technology, e.g. home-based WiFi. Our design would ease the cost of production by zero assembly and installation of the hardware components and would achieve the detection for anywhere inside a house (whole-home detection) by wireless signal transmissions in line-of-sight, none-line-of-sight, and through-the-wall scenarios. The detection system broadcasts the wireless signals in a home environment, receives the reflected signals from the targeted person, and then computes the frequency of the target’s breathing over a period of detection. We expected that frequency plots would be uniform for a normal person but irregular for a patient with sleep apnea. One practical implication of performing such frequency analysis in the whole-home environment is that children’s health monitoring can still happen without parental interference.

The HOPE Study is a randomized control trial conducted by investigators from the Kenya Research Program at the University of Washington Department of Global Health, Kenyatta National Hospital, and the Kenyan Ministry of Health. The clinical trial investigates the impact of home-based HIV testing and education for partners of pregnant mothers in Kisumu, Kenya. Couples are randomized to either home-based partner education and HIV testing (HOPE) as part of routine pregnancy services or to standard antenatal care (ANC). The intervention provides education and HIV testing to men in stable relationships with pregnant HIV-infected and HIV-uninfected women in order to improve overall health of women and infants, reduce risk for vertical and horizontal transmission, and increase identification of men living with HIV and link them to care. Couples will be followed up for uptake of HIV testing, facility delivery, exclusive breastfeeding and postpartum contraceptive use, as well as linkage to HIV care. Cost-effectiveness of HOPE will be evaluated in order to inform future scale up of the intervention in Western Kenya, a region with high HIV-1 incidence during the pregnant/postpartum period and high seroprevalence among men (~10%). A custom Open Data Kit (ODK) Tables application developed at the University of Washington Computer Science and Engineering Department serves as the entry-point for data collection, allowing users to collect client health information, send data to a server, and view aggregate data on their Android device. The HOPE Study Tables application allows the nurses and community health workers to screen patients for eligibility, follow up with the study participants and collect geo point data for home visits. The HOPE Study demonstrates the capabilities of ODK Tables as a data browser and manager, highlighting the importance of developing customizable data-driven applications for disconnected environments.

The current standard eyes-free speech dictation mechanism available on mobile phones is Siri's Eyes-Free mode. There are only three functions available: reviewing the current message, redoing the entire message, and sending the message. This makes speech-input of a long text-message on a mobile device very tedious. I created an app that automatically detects potential errors by aligning alternative dictations of the message and allows users to swipe through different portions of the message. I did a study on 6 users to dictate and edit sentences using Siri's Eyes-Free mode and the app that I created, and I did another study on 10 users to dictate and edit paragraphs using Siri and my app. I compare the resultsof Siri and my app in the two studies. I found that Siri's Eyes-Free mode was generally acceptable for short messages, but that my app increased efficiency for sending paragraphs. My app decreased the amount of time users spent editing paragraphs and also decreased the average word error rate.