October 9, 1998
Fachbereich Physik, University of Oldenburg, GERMANY
Computing the Bayes Perceptron
Perceptrons are simple thresholded linear machines used widely in pattern recognition. In this talk we are going to discuss several geometric algorithms for computing the single Perceptron with the best average generalization ability given a finite set of labeled training examples. Although Perceptrons are useful mainly on the context of linearly separable functions, these ideas can be easily extended to Vapnik's support vector machines and provide powerful tools for building optimal classifiers. The solution of several practical examples will be presented on-line.