Unsupervised global optimization: applications on classification of handwritten digits and visual evoked potentials

The authors discuss the optical recognition of handwritten unconnected numerals and visual evoked potential (VEP) classification using two neural network learning paradigms. The first is an unsupervised approach, trained by the combinatorial optimization routine ALOPEX, while the second method uses the backpropagation algorithm. The unsupervised ALOPEX trained system classifies 1000 training digits to an accuracy of 86.3%, and 500 generalizing characters 86.0% accurately. This compares to 99.8% and 93% for a network trained with the supervised backpropagation algorithm. The system was used to cluster the VEPs of normal and multiple sclerosis (MS) subjects. The method demonstrates two distinct groups of subjects, which when histogrammed illustrate that they largely correspond to the MS and control subject groups. A suitable threshold can be chosen so that the recognizer chooses no false negatives