STOCHASTIC SENSITIVITY ANALYSIS METHOD FOR NEURAL-NETWORK LEARNING

A new, efficient algorithm is developed for the sensitivity analysis of a class of continuous-time recurrent neural networks with additive noise signals. The algorithm is based on the stochastic sensitivity analysis method using the variational approach, and formal expressions are obtained for the functional derivative sensitivity coefficients. The present algorithm uses only the internal states and noise signals to compute the gradient information needed in the gradient descent method, where the evaluation of derivatives is not necessary. In particular, it does not require the solution of adjoint equations of the back-propagation type. Thus, the algorithm has the potential for efficiently learning the network weights with significantly fewer computations. The effectiveness of the algorithm in a statistical sense is shown, and the method is applied to the familiar layered network.