MIWOCI Workshop - 2013

Nonlinear dimensionality reduction (DR) techniques offer the possibility to visually inspect a high-dimensional data set in two dimensions, and such methods have recently been extended to also visualize class boundaries as induced by a trained classifier on the data. In this contribution, we investigate the effect of two different ways to shape the involved dimensionality reduction technique in a discriminative way: discrimination based on the data labels of the given data set, and discrimination based on the labels as provided by the trained classifier. We demonstrate that these two different techniques lead to semantically different visualizations which allow us to further inspect the classification behavior. Both approaches can uniformly be based on the Fisher information matrix, which is estimated in two different ways.