Optimal Brain Surgeon AlgorithmSummary of AlgorithmComputational Hurdles of the OBS algorithmSimplifying the OBSResultsOptimal Brain Surgeon AlgorithmRandy SkalaECE 539Section 1Summary of Algorithm1) Train the multilayer perceptron to a minimum in mean square error2) Compute the inverse Hessian (H-1)3) Eliminate weights will small saliencies4) Repeat until no more weights can be deleted without a large increase in the mean-square error.iiiiiiiiHHwwHwS 1][,][21,12,12Computational Hurdles of the OBS algorithmComputing the inverse Hessian matrixRecursive formula has order greater than n2Many Recursions are needed for convergenceUpdating the Weights, Hessian matrix and Saliencies every time a weight is eliminatedMany weights are typically eliminatedSimplifying the OBSEstimate dF/dw by finding the change in the cost function when each weight is eliminatedUse the initial Hessian matrix for all updatesDo “batch elimination” of weightsResultsIncompleteResults so far seem to show that classification problems are handled well with the simplification.Estimation problems do not seem to work as
View Full Document