Non-linear dimensionality reduction techniques for unsupervised feature extraction

Dimensionality reduction techniques have been regularly used for visualization of high-dimensional data sets. In this paper, reduction to d greater than or equal to 2 is studied, with the purpose of feature extraction. Four different non-linear techniques are studied: multidimensional scaling, Sammon's mapping, self-organizing maps and auto-associative feedforward networks. All four techniques will be presented in the same framework of optimization. A comparison with respect to feature extraction is made by evaluating the reduced feature sets ability to perform classification tasks. The experiments involve an artificial data set and grey-level and color texture data sets. We demonstrate the usefulness of non-linear techniques compared to linear feature extraction. (C) 1998 Elsevier Science B.V. All rights reserved.