Using saturation detection to shorten the training duration for Gaussian ANNs
Minesh A. Shah, Peter H. Meckl · 2005
Gaussian networks utilizing gradient descent are often characterized as being slow to learn. This precludes their application in real time adaptation and control of dynamic systems undergoing parameter variation. For Gaussian networks, the slow convergence may in part be due to the continued adaptation of network parameters that have reached their final value before the completion of the training process. Hence, these parameters are said to be saturated and incapable of further learning. A training algorithm that enhances the performance of gradient descent by detecting saturation and terminating parameter adaptation is developed. Initial results indicate that the algorithm reduces training times without degrading the quality of the input-output mapping.