CAREER: Exploiting low-dimensional structure in data for more effective, efficient and interactive machine intelligence

点击翻译按钮获取中文摘要

The rapid increase in sensor data is revolutionizing many areas of technology, defense, and scientific discovery. Fortunately, despite data being high-dimensional, various aspects of the data can frequently be characterized as having low-dimensional geometric structure. This research project dramatically improves machine intelligence by exploiting this geometric structure for more effective, efficient and interactive data analysis systems. Complementary to these technical objectives, this project also aims to engage, recruit, and educate a diverse collection of students to STEM careers by developing novel curricular and outreach materials that illustrate how mathematics can be used in information systems. The potential benefits of this project are wide ranging in areas where data plays a fundamental role.Improving machine intelligence requires understanding how to best exploit the underlying low-dimensional structure in data for a given type of task, and this project is guided by three research objectives toward this goal. The first objective enhances machine effectiveness by exploiting the fact that multiple observations of the same phenomenon are often related by movement along a manifold, with a particular focus on the canonical computer vision problem of invariant object recognition. The second objective seeks to improve computational efficiency by developing dimensionality reduction techniques for manifold-modeled data that preserves information about nonlinear feature-space mappings. The third objective seeks to leverage interactivity to fully "close the loop" between between humans and machines while learning low-dimensional information from a human expert (an extension of the active learning paradigm). The project also pursues two educational objectives, including developing curriculum modules for pre-college outreach and integrating neural systems content into the ECE curriculum to illustrate the connections between quantitative methods and intelligent systems.