Applications of neural networks are prevailing in speech recognition research. In this paper, first, suitable role of neural network (mainly back-propagation based multi-layer type) in speech recognition, is discussed. Considering that speech is a long, variable length, structured pattern, a direction, in which neural network is used in cooperation with existing structural analysis frameworks, is recommended. Activities are surveyed, including those intended to cooperatively merge neural networks into dynamic programming based structural analysis framework. It is observed that considerable efforts have been paid to suppress the high nonlinearity of network output. As far as surveyed, no experiment in real field has been reported.

This paper introduces a modeling of the human rotation invariant recognition mechanism at the neural level. In the model, mechanisms of memory search and mental rotation are realized in the process of minimizing the energy of a bi-directional connection network. The thrust of the paper is to explain temporal mental activities such as successive memory retrievals and continuous mental rotation in terms of state transitions of collective neurons based on nonequilibrium dynamics. We conclude that regularities emerging in the dynamics of intermittent chaos lead the recognition process in a structural and meaningful way.

Threre are two types of nonlinear associative silicon retinas. One is a sparse Hopfield type neural network which is called a H-type retina and the other is its dual network which is called a DH-type retina. The input information sequences of H-type and HD-type retinas are given by nodes and links as voltages and currents respectively. The error correcting capacity (minimum basin of attraction) of H-type and DH-type retinas is decided by the minimum numbers of links of cutset and loop respectively. The operation principle of the regeneration is based on the voltage or current distribution of the neural field. The most important nonlinear operation in the retinas is a dynamic quantization to decide the binary value of each neuron output from the neighbor value. Also, the edge is emphasized by a line-process. The rates of compression of H-type and DH-type retinas used in the simulation are 1/8 and (2/3) (1/8) respectively, where 2/3 and 1/8 mean rates of the structural and binarizational compression respectively. We could have interesting and significant simulation results enough to make a chip.

This paper reviews the capability of the three layer neural network (TLNN) with one output neuron. The input set is restricted to a finite subset S of En, and the TLNN implements a function F such as F : S I={1, -1}, i,e., F is a dichotomy of S. How many functions (dichotomies) can it compute by appropriately adjusting parameters in the TLNN? Brief historical review, some theorems on the subject obtained so far, and related topics are presented. Several open problems are also included.

A theoretical conjecture on fractal dimensions of a dendrite distribution in neural networks is presented on the basis of the dendrite tree model. It is shown that the fractal dimensions obtained by the model are consistent with the recent experimental data.

Quantization of the impulse response coefficients due to finite word length causes the moments to deviate from their ideal values. This deviation is found to have a linear variation with the output roundoff noise of the filter realized in direct form. Since the zeros and poles of a given filter also move away from their designed locations due to quantization, we show a relation between the zeros and poles and the moments of the impulse response.

This paper proposes a Mean-Separated and Normalized Vector Quantizer with edge-Adaptive Feedback estimation and variable bit rates (AFMSN-VQ). The basic idea of the AFMSN-VQ is to estimate the statistical parameters of each coding block from its previous coded blocks and then use the estimated parameters to normalize the coding block prior to vector quantization. The edge-adaptive feedback estimator utilizes the interblock correlations of edge connectivity and gray level continuity to accurately estimate the mean and standard deviation of the coding block. The rate-variable VQ is to diminish distortion nonuniformity among image blocks of different activities and to improve the reconstruction quality of edges and contours to which the human vision is sensitive. Simulation results show that up to 2.7dB SNR gain of the AFMSN-VQ over the non-adaptive FMSN-VQ and up to 2.2dB over the 1616 ADCT can be achieved at 0.2-1.0 bit/pixel. Furthermore, the AFMSN-VQ shows a comparable coding performance to ADCT-VQ and A-PE-VQ.

The encoding of high quality and super high definition images requires new approaches to the coding problem. The nature of such images and the applications in which they are used prohibits the introduction of perceptible degradation by the coding process. In this paper, we discuss techniques for the perceptually transparent coding of images. Although technically lossy methods, images encoded and reconstructed using these techniques appear identical to the original images. The reconstructed images can be postprocessed (e.g., enhanced via anisotropic filtering), due to the absence of structured errors, commonly introduced by conventional lossy methods. The compression, ratios obtained are substantially higher than those achieved using lossless means.

A new set of self-consistent linear equations is presented for the analysis of the startup characteristics of gyrotron oscillators with an open cavity consisting of weakly irregular waveguides. Numerical results on frequency detuning and oscillation starting current for a whispering-gallery-mode gyrotron are described in which these equations were utilized. Experiments for making a check on the effectiveness of the derived equations showed that they well express the operation of gyrotrons in comparison with the linear theory using an empty cavity field as the wave field.

This paper addresses fault tolerance of a processor array that is reconfigurable by replacing faulty processors with spare processors. The fault tolerance of such a reconfigurable array depends on not only an algorithm for spare processor assignment but also the folloving factor of an organization of spare processors in the reconfigurable array: the number of spare processors; the number of processors that can be replaced by each spare processor; and how spare processors are connected with processors. We discuss a relationship between fault tolerance of reconfigurable arrays and their organizations of spare processors in terms of the smallest size of fatal sets and the reliability function. The smallest size of fatal sets is the smallest number of faulty processors for which the reconfigurable array cannot be failure-free as a processor array system no matter what reconfiguration is used. The reliability function is a function of time t whose value is the probability that the reconfigurable array is failure-free as a processor array system by time t when the best possible reconfiguration is used. First, we show that the larger smallest size of fatal sets a reconfigurable array has, the larger reliability function it has by some time. It suggests that it is important to maximize the smallest size of fatal sets in orer to improve the reliability function as well. Second, we present the best possible smallest size of fatal sets for nn reconfigurable arrays using 2n spare processor each of which is connected with n processors. Third, we show that the nn reconfigurable array previously presented in a literature achieves the best smallest size of fatal sets. That is, it is optimum with respect to the smallest size of fatal sets. Fourth, we present an uppr bound of the reliability function of the optimum nn reconfigurable array using 2n spare processors.

This paper deals with an adaptive credit assignment algorithm to select strategies having higher capabilities in the learning classifier system (LCS) based upon the genetic algorithm (GA). We emulate a kind of prizes and incentives employed in the economies with imperfect information. The compensation scheme provides an automatic adjustment in response to the changes in the environment, and a comfortable guideline to incorporate the constraints. The learning process in the LCS based on the GA is realized by combining a pair of most capable strategies (called classifiers) represented as the production rules to replace another less capable strategy in the similar manner to the genetic operation on chromosomes in organisms. In the conventional scheme of the learning classifier system, the capability s(k, t) (called strength) of a strategy k at time t is measured by only the suitableness to sense and recognize the environment. But, we also define and utilize the prizes and incentives obtained by employing the strategy, so as to increase s(k, t) if the classifier provide good rules, and some amount is subtracted if the classifier k violate the constraints. The new algorithm is applied to the portfolio management. As the simulation result shows, the net return of the portfolio management system surpasses the average return obtained in the American securities market. The result of the illustrative example is compared to the same system composed of the neural networks, and related problems are discussed.

Artificial neurons and neural networks have been shown to perform Principal Component Analysis (PCA) when gradient ascent learning rules are used, which are related to the constrained maximization of statistical objective functions. Due to their parallelism and adaptivity to input data, such algorithms and their implementations in neural networks are potentially useful in feature extraction and data compression. In the companion paper(9), two such learning rules were derived from two criteria, the Subspace Criterion and the Weighted Subspace Criterion. It was shown that the only solutions to the latter problem are dominant eigenvectors of the data covariance matrix, which are the basis vectors of PCA. It was suggested by a simulation that the corresponding learning algorithm converges to these eigenvectors. A homogeneous neural network implementation was proposed for the algorithm. The learning algorithm is analyzed here in detail and it is shown that it can be approximated by a continuous-time differential equation that is obtained by averaging. It is shown that the asymptotically stable limits of this differntial equation are the eigenvectors. The neural network learning algorithm is further extended to a case in which each neuron has a sigmoidal nonlinear feedback activity function. Then no parameters specific to each neuron are needed, and the learning rule is fully homogeneous.

A tag-partitioned join algorithm is described. The algorithm partitions only one relation, while other partition-based algorithms partition both relations. It is performed as the joinable tuples of one relation are rearranged and some of them are duplicated according to the original sequence of the join attribute values of the other relation. To do this, the algorithm first finds the positions of all the tuples of the other relation which are joinable with each tuple of one relation, and then partitions joinable tuples of one relation into buckets by using the positions found. Final joining is performed on the partitioned relation and the other relation. We analyze and compare the performance of the algorithm with that of other partition-based join algorithms. The comparison shows that our method is better than other partition-based methods under the practical values of the analysis parameters.

Principal Component Analysis (PCA) is a useful technique in feature extraction and data compression. It can be formulated as a statistical constrained maximization problem, whose solution is given by unit eigenvectors of the data covariance matrix. In a practical application like image compression, the problem can be solved numerically by a corresponding gradient ascent maximization algorithm. Such on-line algoritms can be good alternatives due to their parallelism and adaptivity to input data. The algorithms can be implemented in a local and homogeneous way in learning neural networks. One example is the Subspace Network. It is a regular layer of parallel artificial neurons with a learning rule that is completely homogeneous with respect to the neurons. However, due to the complete homogeneity, the learning rule does not converge to the unique basis given by the dominant eigenvectors, but any basis of this eigenvector subspace is possible. In many applications like data compression, the subspace is not sufficient but the actual eigenvectors or PCA coefficient vectors are needed. A new criterion, called the Weighted Subspace Criterion, is proposed, which makes a small symmetry-breaking change to the Subspace Criterion. Only the true eigenvectors are solutions. Making the corresponding change to the learning rule of the Subspace Network gives a modified learning rule, which can be still implemented on a homogeneous network architecture. In learning, the weight vectors will tend to the true eigenvectors.

The problems of cache coherency in multiprocessor systems are directly related to their architectural structures. Small scale multiprocessor systems have focused on the use of bus based memory interconnection networks using centrally shared memory and a sequential consistency model for coherency. This has limited scalability to but a few tens of processors due to the limited bus bandwidth used for both coherency updates and memory traffic. Recently, large scale multiprocessor systems have been proposed that use general interconnection networks and distributed shared memory. These architectures have been proposed using weak consistency models and various directory map schemes to hide the overhead for coherency maintenance within the memory hieratchy, interconnection network or process context switch latencies. The coherency and memory traffic are still maintained over the same interconnection network. In this paper, we present the architecture of a new general purpose medium scale multiprocessor system. This Cache Coherent Multiprocessor System (C2MP), supports distributed shared memory using a general memory interconnection network for memory traffic and a separate bus based coherency interconnection network for coherency maintenance. Through the use of a special directory based coherency protocol and cache oriented distributed coherency controllers, direct cache-to-cache coherency maintenance is performed over the dedicated coherency bus. This minimizes coherency updates to only those processor nodes needing coherency maintenance. An aggressive sequential coherncy model is used, which reduces the hardware penalty to support an ideal sequential consistency programmers model. The system can scale up to 256-512 processors depending on the degree of shared data and is expected to have higher per processor utilization in this range than currently proposed medium and large scale multiprocessor systems. The C2MP system is analyzed utilizing a Generalized Timed Petri-Net model of a processor node. A stochastic model for internode interactions over the general memory interconnection network and coherency bus are used . The model of the proposed architecture is analyzed under steady-state conditions for varying system work load parameters.

This paper discusses the design, implementation, and performance of a bus-connected multiprocessor, called Presto, for a Rete-based production system. To perform a match, which is a major phase of a production system, a Presto match scheme exploits the subnetworks that are separated by the top two-input nodes and the token flow control at these nodes. Since parallelism of a production system can only increase speed 10-fold, the aim is to do so efficiently on a low-cost, compact bus-connected multi-processor system without shared memory or cache memory. The Presto hardware consists of up to 10 processisng elements (PEs), each comprising a commercial microprocessor, 4 Mbytes of local memory, and two kinds of newly developed ASIC chips for memory control and bus control. Hierarchical system software is provided for developing interpreter programs. Measurement with 10 PEs shows that sample programs run 5-7 times faster.

The Batcher banyan network is well known as a non-blocking switching fabric. However, it is conflict free only when there is no packets for the same destination. To cope with the arbitrary combination of packets, an additional network or special control sequence which causes the increase of the hardware or performance degradation is required. A Batcher Double Omega network with Combining (BDOC) is an elegant solution of this problem. It consists of a Batcher sorter and two double sized Omega networks. Like in the Batcher banyan network, packets are sorted by the destination label in the Batcher sorter. In the first Omega network called the distributer, a packet is routed by a tag corresponding to the sum of the label at the output of the Batcher sorter and the destination label. In the second (Inverse) Omega network called the concentrator, the original destination label is used as the routing tag, and packets are routed without any conflict. The BDOC is useful for an interconnection network to connect processors and memory modules in multiprocessor. Unlike conventional multistage interconnection networks for multiprocessors, packets are transferred in a serial and synchronized manner. The simple structure of the switching element enables a high speed operation which reduces the latency caused by the serial communication. Using the pipelined circuit switching, the address and data packets share the same control signal, and the structure of the switching element is much simplified. Moreover, packets combining which avoids the hot spot contention is realized easily in the concentrator.

This paper describes a new evaluation technique for Si surfaces. A laser/microwave method using two lasers of different wavelengths for carrier injection is proposed to evaluate Si surfaces. With this evaluation system, the effect of impedance mismatching between the microwave probe and the Si wafer can be eliminated. These lasers used in this experiment are He-Ne (wavelength633 nm, penetration depth3 µm) and YAG lasers (wavelength1060 nm, penetration depth500 µm). Using a microwave probe, the amount of injected excess carriers can be detected. These carrier concentrations are mainly dependent on the condition of the surface, when carriers are excited by the He-Ne laser, and the condition of the bulk region, when carriers are excited by the YAG laser. We refer to microwave intensities detected by the He-Ne and YAG lasers as the surface-recombination-velocity-related microwave intensity (SRMI) and bulk-related microwave intensity (BRMI), respectively. We refer to the difference between SRMI and BRMI as relative SRMI (R-SRMI), which is closely related to the surface condition. A theoretical analysis is performed and several experiments are conducted to evaluate Si surfaces. It is found that the R-SRMI method is better suited to surface evaluation then conventional lifetime measurements, and that the rdliability and reproducibility of measurements are improved.

High-speed digital land mobile communications suffer from frequency-selective fading due to a long delay difference. Several techniques have been proposed to overcome the multipath propagation problem. Among them, an adaptive array antenna is suitable for very high-speed transmission because it can suppress the multipath signal of a long delay difference significantly. This paper describes the LMS adaptive array antenna for frequency-selective fading reduction and a new diversity technique. First, we propose a method to generate a reference signal in the LMS adaptive array. At the beginning of communication, we use training codes for the reference signal, which are known at a receiver. After the training period, we use detected codes for the reference signal. We can generate the reference signal modulating a carrier at the receiver by those codes. The carrier is oscillated independently of the incident signal. Then, the carrier frequency of the reference signal is in general different from that of the incident signal. However, the LMS adaptive array works in such a way that the carrier frequency of the array output coincides with that of the reference signal. Namely, the frequency difference does not affect the performance of the LMS adaptive array. Computer simulations show the proper behavior of the LMS adaptive array with the above reference signal generator. Moreover, we present a new multipath diversity technique using the LMS adaptive array. The LMS adaptive array reduces the frequency-selective fading by suppressing the multipath components. This means that the transmitted power is not used sufficiently. We propose a multiple beam antenna with the LMS adaptive array. Each antenna pattern receives one of the multipath components, and we combine them adjusting the timing. Then, we realize the multipath diversity. In addition to the multipath fading reduction, we can improve a signal-to-noise ratio by the diversity technique.

The energy distribution and emission efficiency of electrons emitted from a superconductor-insulator-superconductor (SIS) junction have been investigated by numerical calculation adopting the free electron model. The emission efficiency of an SIS junction cold cathode was found to be about 0.3% of tunneling current flowing to the SIS junction when the energy gap voltage of superconductor was 20 meV, the work function of counter electrode 1 eV, the bias voltage 0.96 V, the thickness of the counter electrode 100 , the electric field strength between the plate and the counter electrode 106 V/m, and the relaxation time 0.01 ps. It is clear that the SIS junction cold cathode can emit electrons with sharper energy distributions at much the same efficiency as compared with a metal-insulator-metal (MIM) junction cold cathode.