Break-out Topics and Talks

Wednesday, October 24, 2012

Session I
11:15am - 12:20pm
Wireless Power
CSE 305
HCI
CSE 403
Privacy
CSE 691
1:10 - 1:40pm Keynote Talk: GraphLab: Machine Learning for Big Data in the Cloud
Atrium  
Session II
1:45 - 2:50pm
Sensing
CSE 305
Grounded Language Learning
CSE 403
Security
CSE 691
Session III
2:55 - 4:00pm
Explorations in Education
CSE 305
Crowdsourcing
CSE 403
Computing and Health
CSE 691
Session IV
4:05 - 5:05pm
Feedback Session
CSE 305
New Faculty Presentations
CSE 403
Technology for Low Resource Settings
CSE 691
5:00 - 7:00pm RECEPTION AND LAB TOURS (WITH POSTERS AND DEMOS)
various labs  
7:15 - 7:45pm Program: Madrona Prize, People's Choice Awards
Microsoft Atrium  


Please check back for updates. Last updated 6 November 2012.

 


 

Session I

  • Wireless Power (CSE 305)

    • 11:15-11:20: Introduction and Overview: Joshua Smith
    • 11:20-11:40: High intensity wireless power transfer for hearts and phones, Benjamin Waters

      Coupled magnetic resonators can be used to wirelessly deliver large amounts of power (tens of watts) over medium distances (up to a meter). Ben will describe our FREE-D system for powering implanted heart pumps, and our new wireless power receiver that can charge any USB device.

    • 11:40-12:00: Miniaturized Magnetic Resonant Power Transfer for Brains and Bees, Vamsi Talla

      Vamsi will describe work in progress on two "small scale" applications of our wireless power techniques: powering implanted brain sensors, and powering a robotic bee.

    • 12:00-12:20: Far field wireless power, Aaron Parks (PDF slides)

      Far field wireless power can deliver smaller amounts of power (0.1 microwatts to 100s of microwatts) over larger distances (up to 20km, for a broadcast TV tower). Aaron will start by describing some recent results in wirelessly powered sensing: he'll describe our new WISP 5.0 wirelessly powered sensing platform, and then analog backscatter, a technique for collecting audio from RF-powered sensors. He'll then describe our recent results on sensor nodes that are powered by the RF signals emitted by cell phone towers.

  • Human Computer Interaction (CSE 403)

    • 11:15-11:20: Introduction and Overview: James Fogarty (PDF slides)
    • 11:20-11:40: A General-Purpose Target-Aware Pointing Enhancement Using Pixel-Level Analysis of Graphical Interfaces, Morgan Dixon, James Fogarty, Jacob Wobbrock (PDF slides)

      We present a general-purpose implementation of a target aware pointing technique, functional across an entire desktop and independent of application implementations. Specifically, we implement Grossman and Balakrishnan's Bubble Cursor, the fastest general pointing facilitation technique in the literature. Our implementation obtains the necessary knowledge of interface targets using a combination of pixel-level analysis and social annotation. We discuss the most novel aspects of our implementation, including methods for interactive creation and correction of pixel-level prototypes of interface elements and methods for interactive annotation of how the cursor should select identified elements. We also report on limitations of the Bubble Cursor unearthed by examining our implementation in the complexity of real-world interfaces. We therefore contribute important progress toward real-world deployment of an important family of techniques and shed light on the gap between understanding techniques in controlled settings versus behavior with real-world interfaces.

    • 11:40-12:00: ReGroup: Interactive Machine Learning for On-Demand Group Creation in Social Networks, Saleema Amershi, James Fogarty, Daniel S. Weld

      We present ReGroup, a novel end-user interactive machine learning system for helping people create custom, on-demand groups in online social networks. As a person adds members to a group, ReGroup iteratively learns a probabilistic model of group membership specific to that group. ReGroup then uses its currently learned model to suggest additional members and group characteristics for filtering. Our evaluation shows that ReGroup is effective for helping people create large and varied groups, whereas traditional methods (searching by name or selecting from an alphabetical list) are better suited for small groups whose members can be easily recalled by name. By facilitating on-demand group creation, ReGroup can enable in-context sharing and potentially encourage better online privacy practices. In addition, applying interactive machine learning to social network group creation introduces several challenges for designing effective end-user interaction with machine learning. We identify these challenges and discuss how we address them in ReGroup.

    • 12:00-12:20: Gesture coder: a tool for programming multi-touch gestures by demonstration, Hao Lu

      Multi-touch gestures have become popular on a wide range of touchscreen devices, but the programming of these gestures remains an art. It is time-consuming and error-prone for a developer to handle the complicated touch state transitions that result from multiple fingers and their simultaneous movements. In this paper, we present Gesture Coder, which by learning from a few examples given by the developer automatically generates code that recognizes multi-touch gestures, tracks their state changes and invokes corresponding application actions. Developers can easily test the generated code in Gesture Coder, refine it by adding more examples, and once they are satisfied with its performance integrate the code into their applications. We evaluated our learning algorithm exhaustively with various conditions over a large set of noisy data. Our results show that it is sufficient for rapid prototyping and can be improved with higher quality and more training data. We also evaluated Gesture Coder's usability through a within-subject study in which we asked participants to implement a set of multi-touch interactions with and without Gesture Coder. The results show overwhelmingly that Gesture Coder significantly lowers the threshold of programming multi-touch gestures.

  • Privacy (CSE 691)

    • 11:15-11:20: Introduction and Overview: Yoshi Kohno
    • 11:20-11:40: Blocking-Resistant Network Services using Unblock, Will Scott (PDF slides)

      The desire for uncensored access to the Internet has motivated the development of both open proxies like Tor and social graph-based overlays like FreeNet. However, neither design is sufficient, as open access networks are easily exposed and blocked, and overlays based just on social trust suffer from poor availability and performance. I present the design for a new overlay service, Unblock, constructed from an augmented social graph. In Unblock, multi-hop paths through social links protect individual participants from exposure to adversaries. Unblock achieves good performance by introducing additional links in the network graph in a manner that minimizes vulnerability. I also develop several transport level techniques for improved latency and demonstrate the practicality of the system for web traffic workloads.

    • 11:40-12:00: Privacy in an IPv6 Internet, Seungyeop Han (PDF slides)

      As personal information increases in value, the incentive for remote services to collect as much of it as possible increases as well. While there has been much research on how to improve anonymity for users on the Internet, in the current Internet, the default policy is that all behavior can be correlated using a variety of identifying information, not the least of which is a user’s IP address. Tools like Tor, Privoxy, and even NATs, provide the opposite end of the spectrum and prevent any behavior from being linked. On the contrary, we attempt to provide users with more control over linkability, not necessarily more anonymity.

      We designed a cross-layer architecture that provides users with a pseudonym abstraction. To the user, a pseudonym rep- resents a set of activities that the user is fine with linking, and to the outside world, a pseudonym gives the illusion of a single machine. Part of this abstraction is a unique IPv6 address per pseudonym; IPv6 provides such a large address space that each device does not need to limited to a single, globally-unique IP, but can have many of them. We have im- plemented and evaluated a prototype that is able to provide un-linkable pseudonyms in the Chrome web browser in order to demonstrate the feasibility, efficacy, and expressiveness of our approach.

    • 12:00-12:20: Sensor Sift: Balancing Utility and Privacy in Sensor Data, Miro Enev (PDF slides)

      We introduce SensorSift, a new theoretical scheme for balancing utility and privacy in smart sensor applications. At the heart of our contribution is an algorithm which transforms raw sensor data into a `sifted' representation which minimizes exposure of user defined private attributes while maximally exposing application-requested public attributes. We envision multiple applications using the same platform, and requesting access to public attributes explicitly not known at the time of the platform creation. Support for future-defined public attributes, while still preserving the defined privacy of the private attributes, is a central challenge that we tackle in this work.

      To evaluate our approach, we apply SensorSift to the PubFig dataset of celebrity face images, and study how well we can simultaneously hide and reveal various policy combinations of face attributes using machine classifiers.

      We find that as long as the public and private attributes are not significantly correlated, it is possible to generate a sifting transformation which reduces private attribute inferences to random guessing while maximally retaining classifier accuracy of public attributes relative to raw data (average PubLoss = .053 and PrivLoss = .075, see Figure 4). In addition, our sifting transformations led to consistent classification performance when evaluated using a set of five modern machine learning methods (linear SVM, kNearest Neighbors, Random Forests, kernel SVM, and Feed Forward Neural Networks).

Session II

  • Sensing (CSE 305)

    • 1:45-1:50: Introduction and Overview: Shwetak Patel
    • 1:50-2:10: An Ultra-Low-Power Human Body Motion Sensor Using Static Electric Field Sensing, Gabe Cohn (PDF slides)

      Wearable sensor systems have been used in the ubiquitous computing community and elsewhere for applications such as activity and gesture recognition, health and wellness monitoring, and elder care. Although the power consumption of accelerometers has already been highly optimized, this work introduces a novel sensing approach which lowers the power requirement for motion sensing by orders of magnitude. We present an ultra-low-power method for passively sensing body motion using static electric fields by measuring the voltage at any single location on the body. We present the feasibility of using this sensing approach to infer the amount and type of body motion anywhere on the body and demonstrate an ultra-low-power motion detector used to wake up more power-hungry sensors. The sensing hardware consumes only 3.3 µW, and wake-up detection is done using an additional 3.3 µW (6.6 µW total).

    • 2:10-2:30: Using the Doppler Effect on Commodity Mobile Devices for Gesture-based Interaction, Sidhant Gupta (PDF slides)

      Gesture is becoming an increasingly popular means of interacting with computers. However, it is still relatively costly to deploy robust gesture recognition sensors in existing mobile platforms. I will talk about SoundWave, a technique that leverages the speaker and microphone already embedded in most commodity devices to sense in-air gestures around the device. To do this, we generate an inaudible tone, which gets frequency-shifted when it reflects off moving objects like the hand. We measure this shift with the microphone to infer various gestures.

    • 2:30-2:50: GripSense: Using Built-In Sensors to Detect Hand Posture and Pressure on Commodity Mobile Phones, Mayank Goel (PDF slides)

      The world’s typical computer user is now holding a mobile device smaller than her hand, is perhaps in motion, and perhaps carrying more things than just a mobile device. One of the most significant contextual factors affecting mobile device use may be a user’s hand posture. And yet our devices, for the most part, have no clue how they are being held or manipulated, and therefore cannot respond appropriately with adapted user interfaces better suited to different hand postures. I will talk about GripSense, a system that uses a combination of the touchscreen and built-in gyroscope to infer hand postures (like the user of an index finger, left thumb, right thumb, which hand is holding the device). It additionally leverages the built-in vibration motors in a new way to help infer the amount of pressure being applied to the device when interacting with it.

  • Grounded Language Learning (CSE 403)

    • 1:45-1:50: Introduction and Overview: Luke Zettlemoyer
    • 1:50-2:10: Learning to Parse Natural Language Commands to a Robot Control System, Cynthia Matuszek (PDF slides)

      A critical component of understanding human language is the ability to map words and ideas in that language to aspects of the external world. This mapping, called the symbol grounding problem, has been studied since the early days of artificial intelligence. Recently, however, sensory and motor systems have advanced enough to give us access to the real-world side of the problem for the first time; in response, unconstrained natural-language interaction with robots has emerged as a significant research area. This talk discusses how to parse natural language commands to actions that can be readily implemented in a robot execution system. We describe learning a parser based on examples of English commands and corresponding robotic-language expressions; we evaluate the trained system on its ability to follow route instructions through an indoor environment, and demonstrate that our system can learn to translate English commands into correct robot behaviors. Finally, the talk will address future directions and data sources for learning language from interaction with robots.

    • 2:10-2:30: A Joint Model of Language and Perception for Grounded Attribute Learning, Nicholas FitzGerald

      As robots become more ubiquitous and capable, it becomes ever more important for untrained users to easily interact with them. Recently, this has led to study of the language grounding problem, where the goal is to extract representations of the meanings of natural language tied to the physical world. We present an approach for joint learning of language and perception models for grounded attribute induction. The perception model includes classifiers for physical characteristics and a language model based on a probabilistic categorial grammar that enables the construction of compositional meaning representations. We evaluate on the task of interpreting sentences that describe sets of objects in a physical workspace, and demonstrate accurate task performance and effective latent-variable concept induction in physical grounded scenes.

    • 2:30-2:50: Bootstrapping Semantic Parsers from Conversations, Yoav Artzi

      Human-machine natural language interaction offers an immediate and accessible interface. However, such conversations are not only a medium for interaction, they also provide rich opportunities for interactive, continuous learning. When something goes wrong, a system can ask for clarification, rewording, or otherwise redirect the interaction to achieve its goals. In this talk, we present an approach for using conversational interactions of this type to induce models of language understanding. We demonstrate learning without any explicit annotation of the meanings of user utterances. Instead, we model meaning with latent variables, and introduce a loss function to measure how well potential meanings match the conversation. This loss drives the overall learning approach, which induces a weighted CCG grammar that could be used to automatically bootstrap the semantic analysis component in a complete dialog system. Experiments on DARPA Communicator conversational logs demonstrate effective learning, despite requiring no explicit meaning annotations.

  • Security (CSE 691)

    • 1:45-1:50: Introduction and Overview: Yoshi Kohno
    • 1:50-2:10: Strengthening User Authentication through Opportunistic Cryptographic Identity Assertions, Alexei Czeskis

      User authentication systems are at an impasse. The most ubiquitous method – the password – has numerous problems, including susceptibility to unintentional exposure via phishing and cross-site password reuse. Second-factor authentication schemes have the potential to increase security but face usability and deployability challenges. For example, conventional second-factor schemes change the user authentication experience. Furthermore, while more secure than passwords, second-factor schemes still fail to provide sufficient protection against (single-use) phishing attacks.

      We present PhoneAuth, a system intended to provide security assurances comparable to or greater than that of conventional two factor authentication systems while offering the same authentication experience as traditional passwords alone. Our work leverages the following key insights. First, a user’s personal device (e.g., a phone) can communicate directly with the user’s computer (and hence the remote web server) without any interaction with the user. Second, it is possible to provide a layered approach to security, whereby a web server can enact different policies depending on whether or not the user’s personal device is present. We describe and evaluate our server-side, Chromium web browser, and Android phone implementations of PhoneAuth.

    • 2:10-2:30: User Interface Toolkit Mechanisms for Securing Interface Elements, Franzi Roesner (PDF slides)

      User interface toolkit research has traditionally assumed that developers have full control of an interface. This assumption is challenged by the mashup nature of many modern interfaces, in which different portions of a single interface are implemented by multiple, potentially mutually distrusting developers (e.g., an Android application embedding a third-party advertisement). We propose considering security as a primary goal for user interface toolkits. We motivate the need for security at this level by examining today’s mashup scenarios, in which security and interface flexibility are not simultaneously achieved. We describe a security-aware user interface toolkit architecture that secures interface elements while providing developers with the flexibility and expressivity traditionally desired in a user interface toolkit. By challenging trust assumptions inherent in existing approaches, this architecture effectively addresses important interface-level security concerns.

    • 2:30-2:50: Control-Alt-Hack(TM): White Hat Hacking for Fun and Profit (A Computer Security Card Game for Education and Outreach), Tamara Denning (PDF slides)

      We present "Control-Alt-Hack(TM): White Hat Hacking for Fun and Profit," an entertainment-focused tabletop card game that we developed to raise awareness of (and enthusiasm for) computer security issues. We licensed the game mechanics from Steve Jackson Games, then created entirely new content and art to make the game about computer security. Players take on the role of ethical hackers at a security consulting company who perform penetration tests and hacks in a variety of scenarios. Come to learn more about the game that we developed, and the ways that we highlighted emerging technologies and unexpected attack vectors.

Session III

  • Explorations in Education (CSE 305)

    • 2:55-3:00: Introduction and Overview: Dan Grossman

      UW CSE is well-known for providing unmatched curricula for our undergraduate majors, graduate students, and nonmajors across campus, but we also have a long history of broader educational impact. This session will highlight three recent initiatives to affect students and professionals at scale well beyond our UW programs: research to change math education, new outreach activities for K-12 students and teachers in Washington, and variants of several of our courses being made available for free online through a partnership with Coursera.

    • 3:00-3:20: Predicting player behavior and designing custom-tailored challenges, Erik Andersen and Yun-En Liu

      This talk will two current efforts in the Center: predicting player behavior from user data, and generating custom-tailored challenges for each player. We will show how we can use collaborative filtering combined with various approaches for finding similar players to predict with ~90% accuracy when a player is likely to quit on the level they're currently playing. We will then show how we can create new challenges for each player by mapping educational concepts to game levels and creating tasks that force the player to exercise those exact concepts. We will show results from multiple games in which we used procedural content generation to create hundreds of new levels with varying levels of difficulty, greatly increasing our ability to test specific hypotheses about player knowledge.

    • 3:20-3:40: K-12 Outreach activities, especially new initiatives, Hélène Martin (PDF slides)

      UW CSE's K-12 outreach program, DawgBytes, aims to introduce both students and teachers to the exciting world of computer science & engineering. During the last academic year, we reached over 120 teachers and 1,000 students through classroom presentations, summer camps, programming competitions, teacher workshops and more. In this session, we will discuss our motivations for being involved in K-12 outreach and highlight several of our activities.

    • 3:40-4:00: Upcoming CSE courses on Coursera: What and How, Dan Grossman (PDF slides)

      Massive Open Online Courses (MOOCs) have attracted widespread attention in the popular press, throughout academia, and, as evidenced by over one million registrations in courses, by society at large. UW CSE is not standing idly by: In the first half of 2013, we will offer five free courses through Coursera. We are currently working with a great staff of TAs to prepare for these courses, which already have tens of thousands students registered. This presentation will provide background on what makes MOOCs exciting, why we are offering MOOCs, what we are offering, how they will differ from our conventional courses and curricula, and how a course staff of a few people can offer a MOOC to tens of thousands.

  • Crowdsourcing (CSE 403)

    • 2:55-3:00: Introduction and Overview: Dan Weld
    • 3:00-3:20: Two Decision-Theoretic Solutions for Crowdsourcing, Christopher Lin

      Crowdsourcing has become immensely popular for its ability to complete a wide variety of jobs. To account for variability in worker quality, requesters often create workflows by subdividing a large task into a chain of smallsized subtasks. In this work, we address two crowdsourcing problems. First, we address the workflow choice problem. Frequently, requesters experiment with several alternative workflows to accomplish the task, and eventually deploy the one that achieves the best performance during early trials. Surprisingly, this seemingly natural design paradigm does not achieve the full potential of crowdsourcing. In particular, using a single workflow (even the best) to accomplish a task is suboptimal. We show that alternative workflows can compose synergistically to yield much higher quality output. Next, we address the problem of modelling free-response tasks. All current models assume prior knowledge of all possible outcomes of the task. While not an unreasonable assumption for tasks that can be posited as multiple-choice questions, we observe that many tasks do not naturally fit this paradigm, but instead demand a free-response formulation where the outcome space is of infinite size. For both problems, we design probabilistic generative models and design decision-theoretic agents that use these models to dynamically complete tasks. We demonstrate the effectiveness of our agents on Mechanical Turk, where they outperform state-of-the-art controllers.

    • 3:20-3:40: Measuring the Utility of a Product, Michael Toomim PDF

      This talk will show you how to measure the Utility of a user interface, or a whole website, in dollars and cents. Utility quantifies how motivated users are to complete your tasks. How much they like the interface or website. How annoying the bugs are. How much value they find in it. How pretty it is. Everything that matters to a user, and quantified in dollars and cents.

      For instance, we can measure the value of Facebook vs. Google to users—which site do they care more about? And we can show that a banner ad that pays $3 per 1,000 views actually costs users $60 in the equivalent currency of user experience.

      In the process, we derive a new foundation for Human-Computer Interaction, based on value and Economics. Instead of measuring user study survey responses, error counts, or task-completion timings, we measure what matters—how much users care, and how much they use the product.

      Crowdsourcing, and the whole internet, is fueled by human attention. This talk presents a new form of HCI that lets us measure the ability of your systems to motivate it.

    • 3:40-4:00: Cascade: Crowdsourcing Taxonomy Creation, Lydia Chilton (PDF)

      Taxonomies are a useful and ubiquitous way of organizing information. However, creating good organizational hierarchies is difficult because the process requires a global understanding of the objects to be categorized. Usually this is done by an individual or a small group of people working together. This approach does not work well for the quickly changing, large datasets found on the web. Cascade is a crowd workflow that allows crowd workers to each spend as little at 20 seconds towards collectively making a taxonomy. To achieve these results, we introduce a design pattern of generate-apply-edit and a novel technique called adaptive context filtering that allows the crowd to do robust categorization. We evaluate Cascade and show that on three datasets its quality is 80-90% of that of experts. Cascade has a competitive cost to expert information architects, but takes six times more human labor. However, that labor can be parallelized such that Cascade will run in as fast as four minutes instead of hours or days.

  • Computing and Health (CSE 691)

    • 2:55-3:00: Introduction and Overview: Shwetak Patel
    • 3:00-3:20: SpiroSmart (Mobile phone spirometer) - evaluating lung function using a cell phone, Eric Larson

      Home spirometry is gaining acceptance in the medical community because of its ability to detect pulmonary exacerbations and improve outcomes of chronic lung ailments. However, cost and usability are significant barriers to its widespread adoption. To this end, we present SpiroSmart, a low-cost mobile phone application that performs spirometry sensing using the built-in microphone. We evaluated SpiroSmart on 52 subjects, showing that the mean error when compared to a clinical spirometer is 5.1% for common measures of lung function. Finally, we showed that pulmonologists can use SpiroSmart to effectively diagnose lung ailments.

    • 3:20-3:40: Lullaby: A Capture & Access System for Understanding the Sleep Environment, Matthew Kay (PDF slides)

      The bedroom environment can have a significant impact on the quality of a person’s sleep. Experts recommend sleeping in a room that is cool, dark, quiet, and free from disruptors to ensure the best quality sleep. However, it is sometimes difficult for a person to assess which factors in the environment may be causing disrupted sleep. In this paper, we present the design, implementation, and initial evaluation of a capture and access system, called Lullaby. Lullaby combines temperature, light, and motion sensors, audio and photos, and an off-the-shelf sleep sensor to provide a comprehensive recording of a person’s sleep. Lullaby allows users to review graphs and access recordings of factors relating to their sleep quality and environmental conditions to look for trends and potential causes of sleep disruptions. In this paper, we report results of a feasibility study where par-ticipants (N=4) used Lullaby in their homes for two weeks. Based on our experiences, we discuss design insights for sleep technologies, capture and access applications, and personal informatics tools.

    • 3:40-4:00: Milk Banking - pasteurization of human breast milk in low-resource settings, Rohit Chaudhri (PDF slides)

      Developing countries are faced with a daunting challenge to lower their neonate and child mortality. Studies indicate that up to 13% of deaths of children under the age of five could be prevented by breastfeeding alone. One key barrier is the availability of breast milk for vulnerable infants (those born pre-term, to HIV-positive mothers or orphaned at birth). Establishing human milk banks that process donor milk can increase availability of breast milk. However, it has been difficult to provide safe, pasteurized donor milk to infants in developing countries due to cost and lack of infrastructure. Low-cost pasteurization methods require rigorous temperature monitoring and quality assurance processes for adoption at scale.

      We present an affordable system to monitor breast milk pasteurization. It leverages mobile and sensing technologies to enhance an existing, low-cost pasteurization method called flash heat pasteurization. A mobile application running on an Android phone monitors milk temperatures during pasteurization, and guides users performing the procedure. The pasteurization temperature curve is uploaded to a server at the end of the procedure, enabling supervisors to review remotely and perform audits. We discuss results from ongoing deployments at two locations in Durban, South Africa. Efforts are in progress to scale up the use of the system to more sites in the country.

Session IV

  • Feedback session (CSE 305)

    • 4:05-5:00: Feedback/discussion: Paul Beame

      1. What are your overall impressions of our students and the preparation they receive from our undergraduate program? What are the strengths of their preparation, and what aspects of their preparation should we try to strengthen?

      2. We recently redesigned our 300-level curriculum with the goals of streamlining and modernizing it. We are just beginning to graduate students who have gone through the new curriculum. Have you noticed any change in the preparation of our students, as interns or employees?

      3. Which schools and departments other than UW CSE do a reliably good job of preparing students? Is there something that those other departments are doing from which we could learn?

  • New faculty presentations (CSE 403)

    • 4:05-4:25: Face Modeling in the Wild, Ira Kemelmacher-Schlizerman PDF

      Billions of photos are now available on the Internet and in our personal photo collections. Naturally, a substantial portion of these photos captures people's appearance all over the world. Availability of such data opens up an opportunity to 1) model virtually the whole world's population and 2) leverage the modeling for visualization of otherwise highly unorganized pool of photos. The challenge is to come up with computer vision methods that robustly handle photos taken under completely uncontrolled imaging conditions. In this talk, I will describe some of my recent work in this space which focuses on human faces. Part of the work, we also transitioned into "Face Movie"—a feature in Google's Picasa.

    • 4:25-4:45: Toward Richer Visual Recognition, Ali Farhadi

      Object recognition is the study of understanding visual data. Recognizing objects in images or videos has several applications in various domains ranging from biomedical sciences to surveillance and security. The most common practice is to automatically attach a category label to a region in an image. But there is more to recognition than just naming. My ultimate goal is to have a machine generate a human-quality description of images. Humans can form complete sentences describing images. These sentences identify the most interesting objects, the actions that are being performed, and the scene where the action occurs. Emulating this skill demands answers to fundamental questions about recognition: how can a recognition system deal with the vast number of objects in the real world? what should a recognition system report when it sees an unfamiliar object? what are the right quanta of recognition? In this talk, I will explore novel representations that try to answer these questions. I will also show a direct attempt to generate textual descriptions for images.

    • 4:45-5:05: Embracing Interference in Wireless Systems, Shyam Gollakota

      The wireless medium is a shared resource. If nearby devices transmit at the same time, their signals interfere, resulting in a collision. In traditional networks, collisions cause the loss of the transmitted information. For this reason, wireless networks have been designed with the assumption that interference is intrinsically harmful and must be avoided. This talk takes an alternate approach: Instead of viewing interference as an inherently counterproductive phenomenon that should to be avoided, we design practical systems that transform interference into a harmless, and even a beneficial phenomenon.

  • Technology for Low Resource Settings (CSE 691)

    • 4:05-4:10: Introduction and Overview: Gaetano Borriello
    • 4:10-4:30: Mobile Tools for Point-of-Care Diagnostics in the Developing World, Nicola Dell (PDF slides)

      Remote health monitoring and disease detection in the developing world are hampered by a lack of accurate, convenient and affordable diagnostic tests. Many of the tests routinely administered in well-equipped clinical laboratories are inappropriate for the settings encountered at the point of care, where low-income patients may be best served. To address this problem, medical researchers have developed innovative rapid diagnostic tests (RDTs) that are capable of detecting diseases at the point of care within a single patient visit to a clinic. However, for these new diagnostic technologies to be effective, tools must be developed to support the health workers who will be responsible for administering the tests and interpreting their results. This talk describes the design and initial implementation of ODK Diagnostics, a smartphone application that supports health workers in three ways: (1) by facilitating the creation of digital job aids to guide users step-by-step through the process of administering each test correctly, (2) by automatically reading and processing the test results and delivering the diagnosis to health workers, and (3) by automating the process of collecting data regarding the type and outcome of the test administered.

    • 4:30-4:50: Completing the feedback loop: Sending performance information to Community Health Workers, Brian DeRenzi (PDF slides)

      Community health workers (CHWs) have been shown to be an effective and powerful intervention for improving community health. Routine visits, for example, can lower maternal and neonatal mortality rates. Despite these benefits, many challenges, including supervision and support, make CHW programs difficult to maintain. An increasing number of mHealth projects are providing CHWs with mobile phones to support their work, which opens up opportunities for information flow back to CHWs and real-time supervision of the program. In this talk, we will discuss two related projects: one where we evaluated the impact of escalating SMS reminders to improve the promptness of routine CHW visits in Tanzania, where the reminders resulted in an 86% reduction in the average number of days a CHW’s clients were overdue (9.7 to 1.4 days), with only a small number of cases ever escalating to the supervisor. We follow this by discussing a second project that is currently underway where provide simple, rich graphs containing information about individual--and peer--performance to CHWs in India.

    • 4:50-5:05: ODK Sensors: A Sensor Integration Framework for Android at the Application-Level, Waylon Brunette (PDF slides)

      Mobile devices can connect to external sensors over wired and wireless channels. However, ensuring proper device interaction can be burdensome, especially when a single application needs to integrate with a number of sensors using different communication channels and data formats. To simplify interfacing a variety of external sensors with consumer Android devices, we have developed a framework that simplifies both application and driver development by providing abstractions that separate responsibilities between the user application, sensor framework, and device driver. The sensing framework seeks to lower the barriers to creating sensing applications by providing for a high level of customization and flexibility, thereby increasing the variety of external sensors that can be connected to mobile devices. The driver architecture is implemented at the user-level to avoid issues with modifying locked consumer devices. We use four application examples to highlight the range of usage models and the ease with which the applications can be developed.