Jacques H. de Villiers

Senior Scientific Programmer
jacques@cslu.ogi.edu
http://www.deVilliers.com

Education

M.S. Electrical Engineering: Oregon Graduate Institute, 1997
B.Eng., Electronics Engineering: University of Pretoria, 1990

Research Interests

Pattern recognition, generalization. The main property that distinguishes a classifier is its ability to generalize: whereas any of a range of memory types can be used to recall input-output training pairs, classifiers are also expected to give sensible answers when inputs not occurring in the training set are presented to it. The generalization of a classifier is measured by the extent to which it exploits the a priori relationship between data items at different places in the feature space.

My proposed research will seek to understand the underlying principles and to create an optimal generalizer, by discovering what ``optimal" means in different circumstances. To augment and test this theoretical work, I would like to apply my ideas to interesting applications which can benefit from improved generalization. The prime examples of such applications are to be found in fields such as image recognition, signal processing and speech recognition.

Publications

Cole, R. A., D.G. Novick, P.J.E. Vermeulen, S. Sutton, J.H. de Villiers and L.F.A. Wessels, ``Experiments With a Spoken Dialogue System for Taking the U.S. Census, International Journal on Human Computer Interaction, in press, 1996.
de Villiers, J., P. Vermeulen and M. Fanty, "Digital Data Collection at CSLU," Technical Report No. CSLU-004, Center for Spoken Language Understanding, Oregon Graduate Institute, May 10, 1994.
J.H. de Villiers, E. Barnard, ``Basis functions and Generalization,'' Proceedings of the Second South African Workshop on Pattern Recognition, Stellenbosch, South Africa, November 1991.
J.H. de Villiers, E. Barnard, ``Backpropagation neural networks with one and two hidden layers'', IEEE transactions on Neural Networks, January 1993, pp. 136-141.