"CAREER: Thinking that is 'just right': Query-Specific Probabilistic Reasoning and its Application to Large-Scale Sensor Networks" PI: Carlos Guestrin Carnegie-Mellon University
This project develops a novel approach for probabilistic reasoning in complex systems. Whereas most current approaches work by first learning a probabilistic model from data, committing to this model, and then applying probabilistic inference techniques to answer user queries, this project is pursuing a significantly different approach: learn a model specific for the query at hand. This project addresses the problem that complex real-world systems require complex models, and inference in these models can be intractable, thus forcing most practitioners to apply approximate inference techniques that are unstable and inaccurate
This projects aims to demonstrate that many queries can be answered by simple models that enable exact, stable inference. This project will develop algorithms for building such query-specific models, addressing both static and dynamic inference problems, distributed reasoning, and modular or relational query-specific models.
This project's general approach, query-specific probabilistic reasoning, enables the efficient solution of many real-world reasoning problems. Specifically, the project addresses practical problems in sensor networks, including: emergency response, surveillance with camera networks and monitoring of large-scale computer systems. Results from this work will be used to develop a publicly available Machine Learning class, including class projects (data), exercises, notes, slides and lecture videos.
Most of the current approaches for understanding data work by firstlearning a model from data, committing to this model, and thenapplying computational techniques to answer users' questions about themodel. Unfortunately, complex real-world systems require complexmodels, and answering questions in these models is intractable. Thus,most practitioners are forced to apply approximate computationaltechniques that are unstable and inaccurate. The goal of this project was a significant shift: rather than learninga complex model a priori, we learn a model specific for the questionat hand. We seek to demonstrate that many queries can be answered bysimple models that enable exact, stable computations. We can thenprovide algorithms for building such query-specific models, addressingboth static and dynamic reasoning problems, distributed reasoning, andmodular or relational query-specific models. We also had a major focus on scalability. As part of this effort, westarted the GraphLab project for large-scale machine learning. Thiseffort has lead to a major open-source software effort with tens ofthousands of downloads. We have held two GraphLab workshops in thelast couple of years. The first one in 2012 had 318 people inattendance. The second one in 2013 had 570 people. We also had a very successful education plan. For example, the PI'sMachine Learning Class was the most popular graduate class at CMU,with about 120 registered students. The class has greatly benefitedfrom class projects derived from data collected by the PI and otherresearchers, and by data given to the PI through industrycollaborators. In addition, the new PAC-learning result from thisaward has been incorporated into the class. Finally, the PIs classslides have been requested by a number of other instructors worldwide,and have been incorporated into their classes. One important aspect of this project is to support under-representedminority students in science. We had undergraduates working on thisand related projects, supported through CMU's IFYRE program, whichhelps expose undergraduates from under-represented minorities toresearch. We have also hosted a undergraduate from Caltech through aprogram for introducing female students to undergraduate research.
