Comparing Support Vector Machines and Feed-forward Neural Networks with Similar Parameters
From a computational point of view, the main differences between SVMs and FNNs are (1) how the number of elements of their respective solutions (SVM-support vectors/FNN-hidden units) is selected and (2) how the (both hidden-layer and output-layer) weights are found. Sequential FNNs, however, do not show all of these differences with respect to SVMs, since the number of hidden units is obtained as a consequence of the learning process (as for SVMs) rather than fixed a priori. In addition, there exist sequential FNNs where the hidden-layer weights are always a subset of the data, as usual for SVMs. An experimental study on several benchmark data sets, comparing several aspects of SVMs and the aforementioned sequential FNNs, is presented. The experiments were performed in the (as much as possible) same conditions for both models. Accuracies were found to be very similar. Regarding the number of support vectors, sequential FNNs constructed models with less hidden units than SVMs. In addition, all the hidden-layer weights in the FNN models were also considered as support vectors by SVMs. The computational times were lower for SVMs, with absence of numerical problems.
KeywordsSupport Vector Activation Function Radial Basis Function Network Hide Unit Relevance Vector Machine
Unable to display preview. Download preview PDF.
- Bishop, C.M.: Neural Networks for Pattern Recognition. Oxford University Press Inc., New York (1995)Google Scholar
- Vincent, P., Bengio, Y.: Kernel Matching Pursuit. Machine Learning 48(1-3), 165–187 (2002); Special Issue on New Methods for Model Combination and Model SelectionGoogle Scholar
- Chang, C.C., Lin, C.J.: LIBSVM: A Library for Support Vector Machines (2002), http://www.csie.ntu.edu.tw/~cjlin/libsvm