Adaptive dissimilarity measures, dimension reduction and visualization

PhD ceremony: Ms. K. Bunte, 12.45 uur, Aula Academiegebouw, Broerstraat 5, Groningen

Dissertation: Adaptive dissimilarity measures, dimension reduction and visualization

Promotor(s): prof. M. Biehl, prof. N. Petkov

Faculty: Mathematics and Natural Sciences

My thesis presents several extensions of the Learning Vector Quantization (LVQ) algorithm based on the concept of adaptive dissimilarity measures. The metric learning gives rise to a variety of applications.

This thesis includes applications of Content Based Image Retrieval (CBIR) for dermatological images, supervised dimension reduction and advanced texture learning in image analysis, which are discussed in the first part. The detailed investigation of dimensionality reduction is addressed in the second half of the thesis. We propose a general framework which facilitates the adaptation of a variety of dimension reduction methods for explicit mapping functions. This enables not only the possibility of direct out-of-sample extensions, but also the theoretical investigation of the generalization ability of dimension reduction. The concept is illustrated on several unsupervised and supervised examples. Furthermore, a novel technique for efficient unsupervised non-linear dimension reduction is proposed combining the concept of fast online learning and optimization of divergences. In contrast to most non-linear techniques, which display a computational effort growing at least quadratic with the number of points, the proposed method comprise a linear complexity. Finally, three divergence based algorithms are generalized and investigated for the use of arbitrary divergences.