Data Intelligence and Visualization

 
Photo of a weather station
Ensuring Data Quality for the Internet of Things (Video) (Presentation slides)
Tom Dietterich

Sensors drift and break. We need automatic methods for detecting this and either compensating in the downstream data analysis (e.g., ignoring bad sensor values) or dispatching a technician to replace/repair the sensor. I have been developing methods to do this for a large existing sensor networks — namely, weather station networks.

 
Image of sample data as charts and graphs
Personalizing Data (Video) (Overview slide)
Ron Metoyer

Sensors produce a tremendous amount of data that must be interpreted by some end user. Presenting data to end users in an appropriate manner is critical. We are studying the effects of matching visual feedback to people in a personalized manner, taking into account their tasks, context, and individuality.

 
A graphic representing database analysis
Seamless Exploration of IoT Data (Video) (Overview slide)
Arash Termehchy

Database analytics algorithms leverage quantifiable structural properties of the data to extract interesting insights. The same information, however, are represented using many different forms and the structural properties observed over particular representations do not necessarily hold for alternative forms. Enterprises spend a great deal of time and resources to convert databases to the representations over which their analytics algorithms are effective. Thus, current data analytics algorithms cannot handle the volume, variety, and the velocity of the data generated in the Internet of things. In order to make database analytics scalable and usable, we develop database analytics algorithms that are effective over a wide range of choices of structural organizations.

 
Graphic of the State of Oregon
From Sound to Knowledge: Bird Bioacoustics (Video) (Overview slide)
Xiaoli Fern (in collaboration with Raviv Raich, Mathew Betts)

Birds are widely used as biological indicators of the health of our ecosystem. Our project aims to develop a rich set of capabilities that allows us to extract ecologically meaningful information and knowledge from bio-acoustic data continuously collected with in-situ field recorders. In addition to the standard species recognition problem, our efforts also address the problem of detecting rare species and estimating the abundancy of the detected species based on their song activities. Collaborating with ecologists, we also hope to gain a better understanding of the singing behavior of birds by studying their song activities in relation to different aspects of their habitats.

 
Photo of a Alan Fern at a football practice
Computer Vision for Sports Analytics (Video) (Overview slide)
Alan Fern

Advanced sports analytics hinges on the availability of detailed, high-quality data. While the technological and human cost of collecting such data is manageable for teams at elite levels, it is beyond the means of most teams at non-elite levels, e.g. high school athletics. Our research aims at lowering the cost of producing the type of data necessary for advanced sports analytics at non-elite levels. Due to the fact that the vast majority of teams regularly record video of their sporting events, the focus of our work is to push the limits of reliably extracting data directly from video. The presentation will overview our past and recent work on applying computer vision to video of American football and basketball.

 
Screen shots of videos
Inferring Dark Matter and Dark Energy from Videos
Sinisa Todorovic

Recognizing functional objects in images is a long-standing open problem in computer vision. This problem is challenging, because functional objects are not characterized by their appearance and shape (which can be directly observed in the image), but by their function (e.g., sittable surfaces, looking glass) or affordance (e.g., slots are for inserting things into). In our work, we focus on objects whose function (and affordance) can be defined in relation to people and their activities in the scene. Given a video, our goal is to detect functional objects in the scene that can be viewed as attracting people to approach them — called dark matter — as well as functional objects that repel people to move away from them. To this end, we analyze noisy behavior of people in the scene using agent-based, probabilistic Lagrangian mechanics.

 
Picture of a yellow and black bird
Machine Learning for Citizen Science Data (Overview slide)
Weng-Keen Wong

The Internet of Things has enabled the general public to participate in scientific research in a paradigm known as citizen science. Many citizen science projects allows participants to submit observations of interest, such as sightings of a particular bird species. A major concern with this approach is the quality of the observations submitted by the general public. My work will highlight the use of machine learning to extract the signal from the noise in the context of the eBird project, which is one of the largest citizen science projects in existence.

 
Photo of a fire station
Optimal Sensing, Planning, and Control (Overview slide)
Prasad Tadepalli

Internet of Things offers the unique possibility of tightly integrating sensing, planning, and control to optimize the performance of a whole system. We have been working on different applications of IoT including fire and medical emergency response, wild fire and invasive species management, and endangered species protection. These problems are characterized by many interacting distributed entities, highly control-sensitive costs, random exogenous events, noisy sensors, and the need to control multiple parallel activities. Our research develops near-optimal planning and control algorithms for such domains using techniques in artificial intelligence and machine learning.

 
Picture of a yellow and black bird
Leveraging Big Data to Fight Cardiovascular Diseases (Overview slide)
Stephen Ramsey

Patient information from personal genotyping, "omics" profiling, and mobile diagnostic devices promises to usher in the era of precision medicine, in which intelligent systems analyze patient data to provide tailored treatment recommendations; but such systems can only be built on a foundation of predictive models that span from organ systems to cells to molecular pathways. Our lab uses a data-driven approach to map the molecular networks that underlie atherosclerosis, a condition that causes most of the 1.7 million heart attacks and strokes that occur each year. We develop and use statistical machine learning and network analysis to pinpoint the molecules that control gene expression changes in cells in the plaque, with the eventual goal of identifying new molecular targets for inducing atherosclerosis regression or beneficial plaque remodeling in humans.

 
Screen shot of visual analysis tool
Interactive Visual Analytics of Sensor Data (Overview slide)
Eugene Zhang

Data is being collected every moment, everywhere, on everything. Without analysis and interpretation, data remains just that, data. While automatic data analysis is highly desirable, it is often difficult to know which analysis tools to use, or even what questions to ask of the analysis, given a new data application domain. Visualization helps the domain experts and data stakeholders to explore their data sets visually for patterns that can lead to proper choice of analysis tools as well as additional questions. We have been working with civil engineers on visualizing the rating of bridges based on the load data collected from sensors located at various parts of the bridges as well as data from simulation.

 
Graphic of using analysis tools for home appliances
Home Energy Management: Recognition of Home Appliances from Voltage Measurements (Overview slide)
Raviv Raich

Electric supply and demand is becoming a source of concern due to the increase in home power consumption. Home energy management can provide further flexibility in energy demand management. A system which can learn, monitor, and control home energy usage patterns is of interest. We illustrate how voltage signatures can be learned and used to provide online detection of home appliance activations.