Case studies in neural networks and an exploration into data reduction

We describe the application of learining with neural networks and support vector machines on different data sets. We report high results for hepatitis, sonar, and ionosphere, fail completely with a eye tracking data set, and for a natural language learning data set, after desastrous classification rates with more conventional kernel methods, MLPs, and others, we obtain good results with custom kernels that take into account the discrete nature of features. We conclude that the use of learning methods differed very much over data sets. At last, we present a method of unsupervised feature selection based on the unmixing matrix of ICA, which we found promising.