We provide a single channel speech enhancement method leveraging the harmonic structure of voiced speech. A sinusoidal model, based on the pitch of the speaker, is used to filter noisy speech and remove any noise components that lie between the harmonics. To remove noise that lie on each harmonic frequency, we use a noise estimation procedure that exploits spectral sparsity of voiced speech. By measuring the power spectrum at frequencies that correspond to the zero crossings of the windowing function, we can estimate the noise levels even in frames that have voiced speech. We also provide a constrained linear least squares formulation to reduce “musical noise” which arises from difficulty in estimating speech and noise power spectral densities. We show that our method yields high perceptual performance over existing methods, and can easily adapt to conditions in which the noise characteristics are constantly changing.

This paper addresses three questions: How does human memory work? How could I build a memory like that? How could I use it to solve practical problems?

Regarding the workings of the human mind, memory and pattern recognition seem to be intertwined. You generally do not have one without the other. Taking inspiration from life experience, a new form of computer memory has been devised. It has been used successfully in diverse applications such as visual aircraft identification, aircraft navigation, and human facial recognition. Other uses are being explored. The basic idea has many new areas of application.

Two gradient descent adaptive algorithms are compared, the LMS algorithm and the LMS/Newton algorithm. LMS is simple and practical, and is used in many applications worldwide. LMS/Newton is based on Newton's method and the LMS algorithm. LMS/Newton is optimal in the least squares sense. It maximizes the quality of its adaptive solution while minimizing the use of training data. No other linear least squares algorithm can give better performance. LMS is easily implemented, but LMS/Newton, although of great mathematical interest, cannot be implemented in most practical applications. Because of its optimality, LMS/Newton serves as a benchmark for all least squares adaptive algorithms. The performances of LMS and LMS/Newton are compared, and it is found that under many circumstances, both algorithms provide equal performance. For example, when both algorithms are tested with statistically nonstationary input signals, their average performances are equal. When adapting with stationary input signals and with random initial conditions, their respective learning times are on average equal. However, under worst-case initial conditions, the learning time of LMS can be much greater than that of LMS/Newton, and this is the principal disadvantage of the LMS algorithm. But the strong points of LMS are ease of Implementation and optimal performance under important practical conditions. For these reasons, the LMS algorithm has enjoyed very widespread application.

The paper presents neural dynamic optimization (NDO) as a method of optimal feedback control for nonlinear multi-input-multi-output (MIMO) systems. The main feature of NDO is that it enables neural networks to approximate the optimal feedback solution whose existence dynamic programming (DP) justifies, thereby reducing the complexities of computation and storage problems of the classical methods such as DP. This paper mainly describes the background and motivations for the development of NDO, while the two other subsequent papers of this topic present the theory of NDO and demonstrate the method with several applications including control of autonomous vehicles and of a robot arm, respectively.

The paper presents neural dynamic optimization (NDO) as a method of optimal feedback control for nonlinear multi-input-multi-output (MIMO) systems. The main feature of NDO is that it enables neural networks to approximate the optimal feedback solution whose existence dynamic programming (DP) justifies, thereby reducing the complexities of computation and storage problems of the classical methods such as DP. This paper mainly describes the theory of NDO, while the two other companion papers of this topic explain the background for the development of NDO and demonstrate the method with several applications including control of autonomous vehicles and of a robot arm, respectively.

For pt.II. see ibid., p. 490-501. The paper presents neural dynamic optimization (NDO) as a method of optimal feedback control for nonlinear multi-input-multi-output (MIMO) systems. The main feature of NDO is that it enables neural networks to approximate the optimal feedback solution whose existence dynamic programming (DP) justifies, thereby reducing the complexities of computation and storage problems of the classical methods such as DP. This paper demonstrates NDO with several applications including control of autonomous vehicles and of a robot-arm, while the two other companion papers of this topic describes the background for the development of NDO and present the theory of the method, respectively.

This paper describes an application of layered neural networks to nonlinear power systems control. A single generator unit feeds a power line to various users whose power demand can vary over time. As a consequence of load variations, the frequency of the generator changes over time. A feedforward neural network is trained to control the steam admission valve of the turbine that drives the generator, thereby restoring the frequency to its nominal value. Frequency transients are minimized and zero steady-state error is obtained. The same technique is then applied to control a system composed of two single units tied together through a power line. Electric load variations can happen independently in both units. Both neural controllers are trained with the back propagation-through-time algorithm. Use of a neural network to model the dynamic system is avoided by introducing the Jacobian matrices of the system in the back propagation chain used in controller training.

Fundamental developments in feedforward artificial neural networks from the past 30 years are reviewed. the central theme of this article is a description of the history, origination, operating characteristics, and basic theory of several supervised neural network training algorithms including the Perceptron rule, the LMS algorithm, three Madaline rules, and the backpropagation technique. These methods were developed independently, but with the perspective of history they can all be related to each other. the concept which underlies these algorithms is the “minimal disturbance principle,” which suggests that during training it is advisable to inject new information into a network in a manner which disturbs stored information to the smallest extent possible. In the utilization of present-day rule-based expert systems, decision rules must always be known for the application of interest. Sometimes there are no rules, however. the rules are either not explicit or they simply do not exist. For such applications, trainable expert systems might be usable. Rather than working with decision rules, an adaptive expert system might observe the decisions made by a human expert. Looking over the expert's shoulders, an adaptive system can learn to make similar decisions to those of the human. Trainable expert systems have been used in the laboratory for real-time control of a “broom-balancing system.” © 1993 John Wiley & Sons, Inc.

The authors describe how a two-layer neural network can approximate any nonlinear function by forming a union of piecewise linear segments. A method is given for picking initial weights for the network to decrease training time. The authors have used the method to initialize adaptive weights over a large number of different training problems and have achieved major improvements in learning speed in every case. The improvement is best when a large number of hidden units is used with a complicated desired response. The authors have used the method to train the truck-backer-upper and were able to decrease the training time from about two days to four hours

Fundamental developments in feedforward artificial neural networks from the past thirty years are reviewed. The history, origination, operating characteristics, and basic theory of several supervised neural-network training algorithms (including the perceptron rule, the least-mean-square algorithm, three Madaline rules, and the backpropagation technique) are described. The concept underlying these iterative adaptation algorithms is the minimal disturbance principle, which suggests that during training it is advisable to inject new information into a network in a manner that disturbs stored information to the smallest extent possible. The two principal kinds of online rules that have developed for altering the weights of a network are examined for both single-threshold elements and multielement networks. They are error-correction rules, which alter the weights of a network to correct error in the output response to the present input pattern, and gradient rules, which alter the weights of a network during each pattern presentation by gradient descent with the objective of reducing mean-square error (averaged over all training patterns).>

An analysis is made of the sensitivity of feedforward layered networks of Adaline elements (threshold logic units) to weight errors. An approximation is derived which expresses the probability of error for an output neuron of a large network (a network with many neurons per layer) as a function of the percentage change in the weights. As would be expected, the probability of error increases with the number of layers in the network and with the percentage change in the weights. The probability of error is essentially independent of the number of weights per neuron and of the number of neurons per layer, as long as these numbers are large (on the order of 100 or more).

An adaptive pole-zero filter comprised of a bank of bandpass filters is presented. The bandpass filters permit the adaptive filter to be realized in a parallel form of first-order sections. Simple monitoring of the filter poles during adaptation is therefore possible so that stability can be ensured. This paper focuses on bandpass filters which are implemented by a frequency-sampling structure; the use of other types of bandpass filters is briefly discussed. An application in system identification is described and computer simulation results are given.

A few of the well established methods of adaptive signal processing are modified and extended for application to adaptive control. An unknown plant will track an input command signal if the plant is preceded by a controller whose transfer function approximates the inverse of the plant transfer function. An adaptive inverse modeling process can be used to obtain a stable controller, whether the plant is minimum or non-minimum phase. No direct feedback is involved. However the system output is monitored and utilized in order to adjust the parameters of the controller. The proposed method is a promising new approach to the design of adaptive control systems.

New steepest descent algorithms for adaptive filtering and have been devised which allow error minimization in the mean fourth and mean sixth, etc., sense. During adaptation, the weights undergo exponential relaxation toward their optimal solutions. Time constants have been derived, and surprisingly they turn out to be proportional to the time constants that would have been obtained if the steepest descent least mean square (LMS) algorithm of Widrow and Hoff had been used. The new gradient algorithms are insignificantly more complicated to program and to compute than the LMS algorithm. Their general form isW_{j+1} = W_{j} + 2 \mu K \epsilon_{j}^{2K-1}X_{j},whereW_{j}is the present weight vector,W_{j+1}is the next weight vector,\epsilon_{j}is the present error,X_{j}is the present input vector,\muis a constant controlling stability and rate of convergence, and2Kis the exponent of the error being minimized. Conditions have been derived for weight-vector convergence of the mean and of the variance for the new gradient algorithms. The behavior of the least mean fourth (LMF) algorithm is of special interest. In comparing this algorithm to the LMS algorithm, when both are set to have exactly the same time constants for the weight relaxation process, the LMF algorithm, under some circumstances, will have a substantially lower weight noise than the LMS algorithm. It is possible, therefore, that a minimum mean fourth error algorithm can do a better job of least squares estimation than a mean square error algorithm. This intriguing concept has implications for all forms of adaptive algorithms, whether they are based on steepest descent or otherwise.

A fundamental relationship exists between the quality of an adaptive solution and the amount of data used in obtaining it. Quality is defined here in terms of "misadjustment," the ratio of the excess mean square error (mse) in an adaptive solution to the minimum possible mse. The higher the misadjustment, the lower the quality is. The quality of the exact least squares solution is compared with the quality of the solutions obtained by the orthogonalized and the conventional least mean square (LMS) algorithms with stationary and nonstationary input data. When adapting with noisy observations, a filter trained with a finite data sample using an exact least squares algorithms will have a misadjustment given byM=\frac{n}{N}=\frac{number of weights}{number of training samples}If the same adaptive filter were trained with a steady flow of data using an ideal "orthogonalized LMS" algorithm, the misadjustment would beM=\frac{n}{4\tau_{\mse}}=\frac{number of weights}{number of training samples}Thus, for a given time constant\tau_{\mse}of the learning process, the ideal orthogonalized LMS algorithm will have about as low a misadjustment as can be achieved, since this algorithm performs essentially as an exact least squares algorithm with exponential data weighting. It is well known that when rapid convergence with stationary data is required, exact least squares algorithms can in certain cases outperform the conventional Widrow-Hoff LMS algorithm. It is shown here, however, that for an important class of nonstationary problems, the misadjustment of conventional LMS is the same as that of orthogonalized LMS, which in the stationary case is shown to perform essentially as an exact least squares algorithm.

In this paper, we present a novel technique for the design of FIR and IIR digital filters. The design approach begins with the specification of a discrete set of arbitrary magnitude and phase characteristics which describe a desired filter response. These frequency domain characteristics are used to create an ideal "pseudo-filter" whose impulse response is unknown and possibly non-causal, but whose input/output characteristics can be determined for a finite sum of sinusoids. Time-domain techniques common to adaptive system identification are then used to identify a realizable FIR or IIR digital filter which best matches the pseudo-filter. The advantages of this method include the ability to specify response at arbitrarily-spaced frequencies, to use arbitrary cost weighting, and to apply (possibly non-linear) constraints to the range of the filter coefficients.

Most adaptive receiving arrays permit beam steering in a selected "look" direction while rejecting or nulling strong interferences which arrive at angles other than the look direction. Nulling is accomplished by adjusting parameters of a signal processor connected to the array sensing elements (whether at rf, audio, or seismic frequencies) to minimize total output power. A signal arriving in the look direction would not be nulled because the adaptive process is constrained to maintain a predetermined sensitivity in the look direction. A substantial literature exists in the field. The September 1976 issue of the IEEE Transactions on Antennas and Propagation was dedicated to the subject of adaptive antennas. It is the purpose of this presentation to compare the works of Howells and Applebaum, Widrow, and Frost.

An adaptive threshold element is able to "learn" a strategy of play for the game blackjack (twenty-one) with a performance close to that of the Thorp optimal strategy although the adaptive system has no prior knowledge of the game and of the objective of play. After each winning game the decisions of the adaptive system are "rewarded." After each losing game the decisions are "punished." Reward is accomplished by adapting while accepting the actual decision as the desired response. Punishment is accomplished by adapting while taking the desired response to be the opposite of that of the actual decision. This learning scheme is unlike "learning with a teacher" and unlike "unsupervised learning." It involves "bootstrap adaptation" or "learning with a critic." The critic rewards decisions which are members of successful chains of decisions and punishes other decisions. A general analytical model for learning with a critic is formulated and analyzed. The model represents bootstrap learning per se. Although the hypotheses on which the model is based do not perfectly fit blackjack learning, it is applied heuristically to predict adaptation rates with good experimental success. New applications are being explored for bootstrap learning in adaptive controls and multilayered adaptive systems.

Abstract : This report compares the capabilities and equipment requirements of two kinds of adaptive classification networks--the 'Learning Matric,' and the Madaline. The Learning Matrix and the Madaline are structurally quite similar, forming arrays of binary output signals from linear combinations of binary or analog input signals. Both systems can learn (adjust their own parameters) from sets of patterns (with corresponding desired outputs) presented to them, and they can thereafter successfully classify additional patterns as presented. The two systems differ primarily in their training methods and in their output logic. Both the Learning Matrix and the Madaline can be realized either by the use of adaptive hardware or by digital-computer simulation. The comparison indicates that in most cases the Learning Matrix will require fewer adaptation cycles than will the Madaline when applied to a given problem; on the other hand, the Madaline will usually require less equipment than the Learning Matrix. However, in making a choice of techniques for a given situation, the overriding consideration will usually be the question of optimality. (Author)

It is possible to read information nondestructively from two- and three-dimensional magnetic-core digital computer memories in several microseconds by exciting selected cores with rf currents. If two co-ordinate lines of a core in a memory array plane are driven at slightly different frequencies, a beat-frequency signal is generated whose phase may take on one of two values which are separated by 180 electrical degrees. These two possible phases correspond to the 0 and 1 information states of the core. The beat-frequency signal, separated from the inevitable noises by tuned linear filters, may be phase detected to yield the desired information.