Seminar on Computational Learning and Adaptation

  Constructing Features from High-Dimensional Data

Geoffrey J. Gordon
Center for Automated Learning and Discovery
Carnegie Mellon University
Pittsburgh, PA 15213

An important machine learning problem is feature construction: learning functions which reduce raw, high-dimensional data to a lower-dimensional set of features, while still retaining as much information from the raw data as possible. This sort of dimensionality reduction can help filter out noise, and it can be an important pre-processing step before running learning algorithms whose computational complexity or sample complexity grow quickly with input dimensionality. In this talk, I will describe our experiments with feature construction from data recorded using our team of mobile robots. These algorithms include nonlinear principal components analysis, which we have used to compress belief states in planning problems, and predictive state representations, which we have used to compress time series of laser rangefinder readings. While our experiments are based on robot sensor data, the algorithms and results should be relevant to other applications of dimensionality reduction such as text, link structure, or images.

Return to the seminar schedule