A Maximum Likelihood Training Scheme
An error function E for a mixture model is derived from a maximum likelihood approach. The derivation of a gradient descent scheme is performed for both the DSM and the GM networks, and leads to a modified form of the backpropagation algorithm. However, a straightforward application of this method is shown to suffer from considerable inherent convergence problems due to large curvature variations of the error surface. A simple rectification scheme based on a curvature-based shape modification of E is presented.
KeywordsLearning Rate Gradient Descent Training Scheme Output Weight Error Surface
Unable to display preview. Download preview PDF.