Information Preserving Embeddings for Discrimination

Dimensionality reduction is required for "human in the loop" analysis of high dimensional data. We present a method for dimensionality reduction that is tailored to tasks of data set discrimination. As contrasted with Euclidean dimensionality reduction, which preserves Euclidean distance or Euler angles in the lower dimensional space, our method seeks to preserve information as measured by the Fisher information distance, or approximations thereof, on the data-associated probability density functions. We will illustrate the approach for multi-class object discrimination problems.