Training neural pattern classifiers with a mean field theory learning algorithm

This study evaluated the mean field theory (MFT) learning algorithm with pattern classifiers and compared its performance to that of the back-propagation (BP) learning algorithm. The deterministic MFT learning algorithm replaced the time-consuming stochastic Boltzmann machine learning algorithm. In our simulation studies, the MFT was superior to the BP learning algorithm with two nonlinearly separable binary mappings, with the Iris data classification problem, and with a simulated signal classification problem. The MFF algorithm always required a substantially smaller number of training epochs than the BP algorithm, and in the case of the signal classification problem the MFT algorithm achieved much lower errors and misclassification rates.