Neural Networks For Signal Processing Research Articles

A parallel CORDIC architecture dedicated to compute the Gaussian potential function in neural networks

In this paper, a new parallel hardware architecture dedicated to compute the Gaussian potential function is proposed. This function is commonly utilized in neural radial basis classifiers for pattern recognition as described by Lee [(Neural Networks for Signal Processing. Prentice-Hall, Englewood Cliffs, NJ, 1992)], Girosi and Poggio [(Neural Comput. 1 (1989) 465)], and Musavi et al. (Neural Networks 5 (1992) 595). Attention to a simplified Gaussian potential function which processes uncorrelated features is confined. Operations of most interest included by the Gaussian potential function are the exponential and the square functions. Our hardware computes the exponential function and its exponent at the same time. The contributions of all features to the exponent are computed in parallel. This parallelism reduces computational delay in the output function. The duration does not depend on the number of features processed. Software and hardware case studies are presented to evaluate the new CORDIC.

Millimeter-wave active imaging using neural networks for signal processing

A neural network has been successfully implemented in an active-mode millimeter-wave (60 GHz) imaging system with a Yagi-Uda antenna array in order to recognize objects and reconstruct images that appear distorted under coherent millimeter-wave illumination. With 10 /spl times/ 10 sampling points and five teaching trials, a recognition rate of 98% has been obtained for ten dissimilar alphabetical letters used as objects. The success rate of reconstruction of distorted millimeter-wave images was 80% when five dissimilar letters were used for the reconstruction. The recognition rate after changing the spatial resolution of the optical system and sampling interval of the image is also discussed.

Handbook of Neural Network Signal Processing [Book Review

Handbook of Neural Network Signal Processing [Book Review

Handbook of Neural Network Signal Processing

From the Publisher: The use of neural networks is permeating every area of signal processing. They can provide powerful means for solving many problems, especially in nonlinear, real-time, adaptive, and blind signal processing. The Handbook of Neural Network Signal Processing brings together applications that were previously scattered among various publications to provide an up-to-date, detailed treatment of the subject from an engineering point of view.The authors cover basic principles, modeling, algorithms, architectures, implementation procedures, and well-designed simulation examples of audio, video, speech, communication, geophysical, sonar, radar, medical, and many other signals. The subject of neural networks and their application to signal processing is constantly improving. You need a handy reference that will inform you of current applications in this new area. The Handbook of Neural Network Signal Processing provides this much needed service for all engineers and scientists in the field.

A Clustering Method Using Hierarchical Self-Organizing Maps

We describe a method of clustering that uses self-organizing maps (SOMs) in a method of image classification. To ensure that this clustering method is fast, we defined a hierarchical SOM and used it to construct the clustering method (M. Endo, M. Ueno, T. Tanabe, and M. Yamamoto, in Proc. of the IEEE Int. Workshop on Neural Networks for Signal Processing X, 2000, pp. 261---270). We define the clustering method in detail and outline its behavior as determined on the basis of both theory and experiment. We also propose a cooperative learning algorithm for the hierarchical SOM. Experiments on artificial image data confirmed the basic performance and adaptability of the SOM in clustering images. We also confirmed, both experimentally and theoretically, that our method is faster SOM, for the objects used in these experiments, than a method based on a non-hierarchical SOM.

BOOK REVIEW: Applied Neural Networks for Signal Processing, by F. L. Luo, & R. Unbehauen, 1997

BOOK REVIEW: Applied Neural Networks for Signal Processing, by F. L. Luo, & R. Unbehauen, 1997

Exploitation and usage of signal evaluating concepts of nature—Ultrasonic data acquisition and neural network signal processing

The objective of the work is the mapping of the main features of data acquisition and neural signal processing concepts, which can be found in nature like the biosonar of bats and dolphins, to technical applications like medical ultrasound or material testing. A first attempt in this direction was the exploitation of higher harmonics of an emitted ultrasound wave for the measurement of very weak signals by an acousto-optical Mach-Zehnder- and Sagnac-Interferometer. A second ongoing work is the analysis of a frequency-modulated signal by neural network signal processing methods for the purpose of medical ultrasound application. The first promising results give rise to the hope of further technical applications. a)Former address. Currently with BRAIN.

Neural Networks for Signal Processing

<i>Neural Networks for Signal Processing</i>

Inversion scheme based on optimization for 3-D eddy current flaw reconstruction problems

An inversion scheme based on first-order optimization is developed for eddy current flaw reconstruction problems with arbitrary specimen, probe and defect shapes. As an essential component of this scheme, a new 3-D forward solver is introduced for the purpose of rapid flaw signal prediction in the inversion loop. This forward solver, whose numerical formulation is basically a discrete reaction variational technique, relies on a reaction data set in the form of an equation system, constructed before entering the inversion loop by a finite element electromagnetic field simulator. The anomalous region is subdivided into small subregions, called flaw cells, and a flaw is represented by a complete set of current dipole density pulses defined in these flaw cells. The coefficient matrix of the equation system consists of reactions between the dipole current density pulses while the elements of the right-hand-side vector are reactions between the pulses and the probe coils. The gradient of the error function, which represents the sensitivity with respect to the flaw parameters, can also be computed quickly from the same pre-calculated reaction dataset, thereby ensuring the efficient implementation of a first-order optimization algorithm. In order to avoid being trapped in a local minimum of the error function, good initial flaw estimates are generated by a neural network signal processing system developed recently by the authors. Various reconstruction examples demonstrate the efficiency of the reconstruction system.

Guest Editors' Introduction: Neural Networks For Signal Processing

Guest Editors' Introduction: Neural Networks For Signal Processing

A study on a position sensitive time of flight detector based on a neural network signal processing

The capabilities of a position sensitive time of flight scintillator (TOF-PS), viewed by an array of subminiature PMTs and based on a neural network signal processing, are studied. For each PMT, time and charge informations are fed to an artificial neural network to define on-line the event position. The first prototype of this detector consists of two 1.5 cm × 1.5 cm × 100 cm scintillator strips viewed by four directly coupled subminiature PMTs. A TOF time resolution of 166 ps and a spatial resolution of 2 cm (single counter) have been achieved.

The past, present, and future of neural networks for signal processing

The article provides a review of the fundamental of neural networks and reports recent progress. Topics covered include dynamic modeling, model-based neural networks, statistical learning, eigenstructure-based processing, active learning, and generalization capability. Current and potential applications of neural networks are also described in detail. Those applications include optical character recognition, speech recognition and synthesis, automobile and aircraft control, image analysis and neural vision, and several medical applications. Essentially, neural networks have become a very effective tool in signal processing, particularly in various recognition tasks.

Two kinds of neural network algorithms suitable for fiber optic sensing array signal processing

Neural networks are employed to process signals from fiber optic sensing arrays embedded in physical structures to monitor structure parameters and to estimate structural damage in real time. The backpropagation neural network, the Kohonen neural network, and its variants the LVQ1 to LVQ5 algorithms (learning-vector quantization) are discussed in detail. Simulation experiments have been done to test the suitability of neural networks for signal processing of fiber optic sensing arrays. The results of these experiments are also given.

A hybrid digital signal processing-neural network CDMA multiuser detection scheme

We present a new hybrid digital signal processing-neural network two-step multiuser detection scheme whose small computational complexity makes it attractive for real-time CDMA multiuser detection. An investigation on the nature of the local minima of the Optimal Multiuser Detector's (OMD) objective function leads to the development of an efficient algorithmic stage that can reduce significantly the size of the OMD optimization problem. This stage may then be followed by a Hopfield neural network employed to solve a smaller size residual problem of the same form. The performance of the hybrid detector is evaluated via simulations and it is shown to exceed that of other suboptimal receivers at a much lower computational cost in both synchronous and asynchronous CDMA transmission cases.

A spectral network model of pitch perception

A model of pitch perception, called the spatial pitch network or SPINET model, is developed and analyzed. The model neurally instantiates ideas from the spectral pitch modeling literature and joins them to basic neural network signal processing designs to stimulate a broader range of perceptual pitch data than previous spectral models. The components of the model are interpreted as peripheral mechanical and neural processing stages, which are capable of being incorporated into a larger network architecture for separating multiple sound sources in the environment. The core of the new model transforms a spectral representation of an acoustic source into a spatial distribution of pitch strengths. The SPINET model uses a weighted "harmonic sieve" whereby the strength of activation of a given pitch depends upon a weighted sum of narrow regions around the harmonics of the nominal pitch value, and higher harmonics contribute less to a pitch than lower ones. Suitably chosen harmonic weighting functions enable computer simulations of pitch perception data involving mistuned components, shifted harmonics, and various types of continuous spectra including rippled noise. It is shown how the weighting functions produce the dominance region, how they lead to octave shifts of pitch in response to ambiguous stimuli, and how they lead to a pitch region in response to the octave-spaced Shepard tone complexes and Deutsch tritones without the use of attentional mechanisms to limit pitch choices. An on-center off-surround network in the model helps to produce noise suppression, partial masking, and edge pitch. Finally, it is shown how peripheral filtering and short-term energy measurements produce a model pitch estimate that is sensitive to certain component phase relationships.

A systematic investigation on the use of time-dependent sensor signals in signal-processing techniques

Due to the lack of selectivity and separability of most common chemical sensors, the use of sensor arrays in combination with artificial neural network (ANN) signal processing is widely used. The time-dependent behaviour of gas sensors is one of the principal problems in signal-processing methods for sensor arrays. These time dependencies of semiconductor sensors can be classified into three main time domains: first, the rise time of the sensor after a sudden concentration change; secondly, the short-term drift after switching on; and finally, the long-term deterioration of the semiconductor material. An important improvement in the treatment of these time-dependent processes is the compensation of sensor rise time and short-term drift effects. The short-term drift of semiconductor sensors can be modelled and equalized by a preprocessing step. The rise-time effects can be reduced by measuring the sensor signals with a high measurement rate and training ANN's with these measurements. Preprocessing the sensor signal and applying the rise-time behaviour leads to an improvement of ANN generalization ability, and attains a minimum of prediction error after a minimum of time after gas exchange.

A novel approach to the convergence of neural networks for signal processing

A novel deterministic approach to the convergence analysis of (stochastic) learning algorithms is presented. The link between the two is the new concept of time-average invariance, which is a property of deterministic signals but resembles that of stochastic signals which are ergodic and stationary.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Full analog CMOS integration of very large time constants for synaptic transfer in neural networks

A method for the full on-chip analog implementation of large time constants in a CMOS technology is presented. These time constants are used for delayed synaptic transfer in neural networks for signal processing. For real-time speech recognition time constants from 1 to 500 ms are necessary, that vary logarithmically with a 4-bit digital code. An RC-type circuit is used to make a continuous time implementation. A modified operational transconductance amplifier (OTA) in a negative feedback configuration is used as a resistor and a gate-oxide capacitor is used as load. The output current of the OTA is divided by current dividers to obtain an electronic multiplication of the time constant. An adapted current mirror design is applied. Much attention is paid to offset-voltage reduction. This new multiplication scheme has several important advantages compared to other schemes. A global area minimization procedure is explained. The total active are of a 3-μm CMOS chip for a 500-ms time constant is 0.5 mm2. The realized circuit has a typical measured offset voltage of 130 mV. The area efficiency and the offset sensitivity are shown to be one order of magnitude better than in other designs.

Field‐Effect Transistors as Sensors of Neural Systems

In the analysis of the signal processing in small neural networks, the application of field-effect transistors as potential probes has led to significant progress. Recently Fromherz et al. were successful in growing neurons directly on a thin insulating layer of the gate-oxide on n-silicon (shown right). The changes in potential in the cell during an action potential via capacitative coupling have the same effect on the source–drain current as changes of tension on a metal gate. The goal of measuring electrical potentials in neuron cultures on one chip with resolutions of 1 μm or less has become feasible.

Applications of neural-network (NN) signal processing in brain research

The application is reviewed of neural-network (NN) signal processing methods to neurological waveform detection and pattern analysis. NN methods are shown to be an excellent way of incorporating expert knowledge about the brain into a mathematical framework with minimal assumptions about the statistics of signal and noise. Constrained by expert knowledge, NN algorithms can search for optimal and near-optimal connections between, and weightings of, application specific features in data spaces for which human knowledge is incomplete. Applying NN algorithms to electrical signals noninvasively recorded from the human brain, the neurological effects of different types of sleeping pills have been differentiated, and insights have been gained as to how our brains produce higher cognitive functions.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>