Optimization in companion search spaces: the case of cross-entropy and the Levenberg-Marquardt algorithm

We present a new learning algorithm for the supervised training of multilayer perceptrons for classification that is significantly faster than any previously known method. Like existing methods, the algorithm assumes a multilayer perceptron with a normalized exponential (softmax) output trained under a cross-entropy criterion. However, this output-criteria pairing turns out to have poor properties for existing optimization methods (backpropagation and its second order extensions) because second-order expansion of the network weights about the optimal solution is not a good approximation. The proposed algorithm overcomes this limitation by defining a new search space for which a second-order expansion is valid and such that the optimal solution in the new space coincides with the original criterion. This allows the application of the Levenberg-Marquardt search procedure to the crossentropy criterion, which was previously thought applicable only to a mean square error criteria.