The process of generating the F0 contour of speech has been modeled quite accurately in mathematical terms by Fujisaki and his coworkers, but the extraction of parameters of the underlying commands from an observed F0 contour is an inverse problem that can be solved only by successive approximation. In order to guarantee an efficient and accurate search for the solution, one needs to start with a set of initial values that are close enough to the optimum. This paper presents a method for pre-processing a measured F0 contour to obtain its approximation consisting of thirdorder polynomial segments that are continuous and differentiable everywhere. It is shown that the proposed method allows one to obtain first-order approximations to the parameters of accent commands for about 90% of all the accent commands, and of phrase commands for about 84% of all the phrase commands.

Investigation of adaptive control systems using neuronlike networks for the optimization of multitasking control of an unknown object has revealed that the identification of the unknown object should precede the main adaptation process. The Adaptive Neuronlike Network (ANN) is used for the simulation of an "inverted object model". In the result of the identification procedure a joined block composed of the unknown object and the ANN may be described by a matrix close to the identity matrix. This procedure considerably simplifies the optimization of multitasking control.A new model of neuronlike element with nonlinear presynaptic inhibition was introduced.Applying this model and a modified learning process makes it possible to simulate a broad class of nonlinear multidimensional objects. AbstractA new learning algorithm for multi-layered neural networks is presented. This algorithm, called minimal disturbance backpropagation, approximates a least mean squared error minimization of the error function while minimally disturbing the connection weights in the network. This means that the information previously trained into the network is disturbed to the smallest amount possible while achieving the desired error correction. Simulation results indicate that this algorithm is more robust and yields much faster convergence rates than the standard back-propagation algorithm. AbstractThis report presents a back-propagation algorithm that varies the number of hidden units. This algorithm i s expected t o escape l o c a l minima and.makes i t no longer necessary t o decide the number of hidden units. We e x p l a i n e x c l u s i v e OR t r a i n i n g and 8 x 8 d o t alphanumeric font training using t h i s algorithm. In exclusive OR t r a i n i n g , the p r o b a b i l i t y of being trapped i n local minima i s reduced. In alphanumeric font t r a i n i n g , the network converged two t o three times faster than the conventional back propagation algorithm. Abstract THE EFFECTS OF PRECISION CONSTRAINTS INA BACK-PROPAGATION LEARNING NETWORK Primacy and recency effects are analyzed mathematically for back propagation algorithms (generalized delta rule), which use momentum. Our results show that when the conventional momentum parameter is used, a primacy effect occurs: The current values of the weights are biased towards the first presentations in a sequence of training patterns. To produce a recency effect, we introduce a different momentum parameter. The current values of the weights depend more on recent presentations of training patterns under this recency effect. A method is provided for selecting a momentum parameter based on the effect desired : primacy or recency.ABSTRACT This paper presents a study of precision constraints imposed by a hybrid chip architecture with analog neurons and digital back-propagation calculations. Conversions between the analog and digital domains and weight storage restrictions impose precision limits on both analog and digital calculations. It is shown through simulations that a learning s...

No abstract

No abstract

No abstract