Neural Networks for Irregularly Sampled Time Series DataGregory W. Koch, Benjamin M. Marlin and Steve Cheng-Xian Li.
National Science Foundation Research Experiences for Undergraduates (NSF REU) poster session, University of Massachusetts, Amherst. (July 2012)
Abstract: Many time series classification problems involve irregularly sampled data wherein observed variables across data cases do not align in time or number due to inherent unpredictability in the sampling process. These time series are difficult to handle because few classical learning algorithms are robust to variable-length features. We motivate a direct approach to perform classification on irregularly sampled time series using neural networks without preprocessing our data set using feature extraction methods. Our results indicate promising performance at classification of noisy and irregularly sampled time series.
An Integrated Speech-to-Text Parser for a Humanoid RobotGregory Koch, Juan Fasola, and Maja Matarić.
National Science Foundation Research Experiences for Undergraduates (NSF REU) poster session, University of Southern California, Los Angeles. (August 2011)
Abstract: Current work with test beds like the Bandit II humanoid robot platform has indicated early success with the efficacy of robots in physical rehabilitation therapy and autism spectrum disorder in early childhood development. However, few researchers have taken advantage of modern natural language processing systems to develop versatile dialog systems for robots that are designed for everyday use. The principal aim of this work is to integrate a general speech-to-text system using natural language processing to facilitate better human-robot interaction in the home environment. Further work will extend the robot's capacities to promote learning within the context of active dialog with other entities in real world situations.
Minimizing networks in Snell Geometry: the Snell-Steiner criterionJack Mealy and Gregory Koch. (joint talk)
AMS/MAA Joint Mathematics Meetings, New Orleans. (January 2011)
Abstract: Further results in the category, Snell Geometry. (See various Snell Geometry abstracts from MathFests 2008-2010). Recall that a Snell Geometry is a system consisting entirely of regions of locally constant curvature, wherein Snell's Law (of optics) is in play across the boundaries between these regions of constant curvature (but which have different indices of refraction", n). After a few general remarks about this category, we report on recent work on minimizing networks in Snell systems (where all regions have zero curvature, but different values of n.) Here, both the Snell dynamic, as well as the Steiner minimizing tree conguration, are in play. Software modeling both of these phenomena has been developed and used to investigate these networks; this will be displayed. Of particular note, we are able to render a large set of non-classical Steiner point congurations. We discuss a criterion that was subsequently derived which characterizes these non-classical congurations, and further which subsumes both Snell's Law and the classical Steiner conguration. Some discussion of this dynamic within yet more complex Snell systems will be included.