An overview is given of recent results coming from non- invasive brain imaging data from PET, fMRI, EEG & MEG machines. These are briefly surveyed and data analysis techniques presently being used reviewed. The results of the experiments can be divided into static and dynamical processing. The most important recent technique used in analyzing PET and fMRI, that of structural modeling, is briefly described. Results from this approach from several experiments are used to build a static model of the global flow of information in the brain. Possible modifications to this model by dynamic understanding of the brain from MEG experiments, and the further problems they present, is then considered. Approaches to these problems by neural modeling is briefly discussed to conclude the paper.

Precise temporal patterning of activity within and between neurons has been predicted on theoretical grounds, and found in the spike trains of neurons recorded from anesthetized and conscious animals, in association with sensor stimuli and particular phases of task performance. However, the functional significance of such patterning in the generation of behavior has not been confirmed. We recorded from multiple single neurons in regions of rat auditory cortex during the waiting period of a Go/NoGo task. During this time the animal waited for an auditory signal with high cognitive load. Of note is the fact that neural activity during the period analyzed was essentially stationary, with no event related variability in firing. Detected patterns therefore provide a measure of brain state that could not be addressed by standard methods relying on analysis of changes in mean discharge rate. The possibility is discussed that some patterns might reflect a preset bias to a particular response, formed in the waiting period. Others patterns might reflect a state of prior preparation of appropriate neural assemblies for analyzing a signal that is expected but of unknown behavioral valence.

The richness and complexity of data sets obtained from functional neuroimaging studies of human cognitive behavior, using techniques such as positron emission tomography and functional magnetic resonance imaging, have until recently not been exploited by computational neural modeling methods. In this article, following a brief introduction to functional neuroimaging methodology, two neural modeling approaches for use with functional brain imaging data are described. One, which uses structural equation modeling, examines the effective functional connections between various brain regions during specific cognitive tasks. The second employs large-scale neural modeling to relate functional neuroimaging signals in multiple, interconnected brain regions to the underlying neurobiological time-varying activities in each region. These two modeling procedures are illustrated using a visual processing paradigm.

We investigate different mechanisms for the control and amplification of cortical neurodynamics, using a neural network model of a three layered cortical structure. We show that different dynamical states can be obtained by changing a control parameter of the input-output relation, or by changing the noise level. Point attractor, limit cycle, and strange attractor dynamics occur at different values of the control parameter. For certain, optimal noise levels, system performance is maximized, analogous to stochastic resonance phenomena. Noise can also be used to induce different dynamical states. A few noisy network units distributed in a network layer can result in global synchronous oscillations, or waves of activity moving across the network. We further demonstrate that fast synchronization of network activity can be obtained by implementing electromagnetic interactions between network units.

We present a novel method for data clustering using temporal segmentation of spiking neurons. We use arrays of neurons whose pulse coupled interactions reflect the internal structure of the data set. The dynamical development of this system leads to temporal grouping of neurons that belong to the same cluster, while different clusters fire at different times. Grouping is achieved via two mechanisms: intra cluster synchrony and desynchronization between clusters. The former is induced by either instantaneous excitatory connections or delayed inhibitory ones, and the latter is induced by instantaneous inhibitory competition. We apply our method to a synthetic sum of gaussians and to the iris data set, demonstrating its capabilities.

For fast neural computations within the brain it is very likely that the timing of single firing events is relevant. Recently Maass has shown that under certain weak assumptions a weighted sum can be computed in temporal coding by leaky integrate-and-fire neurons. This construction can be extended to approximate arbitrary functions. In comparison to integrate-and-fire neurons there are several sources in biologically more realistic neurons for additional nonlinear effects like e.g. the spatial and temporal interaction of postsynaptic potentials or voltage-gated ion channels at the soma. Here we demonstrate with the help of computer simulations using GENESIS that despite of these nonlinearities such neurons can compute linear functions in a natural and straightforward way based on the main principles of the construction given by Maass. One only has to assume that a neuron receives all its inputs in a time interval of approximately the length of the rising segment of its excitatory postsynaptic potentials. We also show that under certain assumptions there exists within this construction some type of activation function being computed by such neurons. Finally we demonstrate that on the basis of these results it is possible to realize in a simple way pattern analysis with spiking neurons. It allows the analysis of a mixture of several learned patterns within a few milliseconds.

We present the basic architecture of a Memory Optimized Accelerator for Spiking Neural Networks. The accelerator architecture exploits two novel concepts for an efficient computation of spiking neural networks: weight caching and a compressed memory organization. These concepts allow a further parallelization in processing and reduce bandwidth requirements on accelerator's components. Therefore, they pave the way to dedicated digital hardware for real-time computation of more complex networks of pulse-coded neurons in the order of 10⁶ neurons. The programmable neuron model which the accelerator is based on is described extensively. This shall encourage a discussion and suggestions on features which would be desirable to add to the current model.

Micro and nano-satellites are important tools to explore and test new ideas and various new devices for space missions without spending extreme amounts of money. The actual launch cost per kilogram payload on a micro or nano-satellite can be as high or even higher than ordinary satellites but the turn around time and quick responses are extremely important. The HUGIN project is a nano-satellite (less than 10 kg) explicitly designed to test magnetic coils and adaptive artificial neural network (ANN) algorithms for attitude control purposes. A small PC video camera is also included and if the control function is successful then also tests of adaptive image processing using other ANN and biologically inspired methods will be performed.

ALIS (Auroral Large Imaging System) is an imaging facility in Northern Sweden. The system consists of six unmanned, remote controlled stations. Each station has a high performance CCD imager, and some stations also have other scientific instrumentation (e.g. pulsation magnetometers). ALIS is capable of producing large amounts of data in a short time. For that reason, novel (AI/VI) techniques for data analysis, are of high priority in order to be able to handle the large data sets. In this paper we will try to describe the current implementation and address the questions of how to interface AI/VI applications to an existing multi station research facility, in terms of real- time experiment control, selective imaging, real-time data analysis, etc.

In Auroral Large Imaging System (ALIS) there is need of stable methods for analysis and classification of auroral images and images with for example mother of pearl clouds. This part of ALIS is called Selective Imaging Techniques (SIT) and is intended to sort out images of scientific interest. It's also used to find out what and where in the images there is for example different auroral phenomena's. We will discuss some about the SIT units main functionality but this work is mainly concentrated on how to find auroral arcs and how they are placed in images. Special case have been taken to make the algorithm robust since it's going to be implemented in a SIT unit which will work automatic and often unsupervised and some extends control the data taking of ALIS. The method for finding auroral arcs is based on a local operator that detects intensity differens. This gives arc orientation values as a preprocessing which is fed to a neural network classifier. We will show some preliminary results and possibilities to use and improve this algorithm for use in the future SIT unit.

A method for recognition of geometrical shapes in auroral forms is presented. The method is based on the analysis of isolines of auroral luminosity shapes. The basic variables used are the angle, (phi) (s), between the tangent of the contour and the x-axis of an arbitrary coordinate system, and the differential, d(phi) (s), as a function of the distance, s, along the contour. The analysis also includes Fourier transformation of the experimental function d(phi) (s) obtained for the observed auroral forms, and the comparison of the power spectrum, F(k), with those for a series of model contours. Some dynamical characteristics of the aurora are also discussed.

The design of a silicon eye using dynamical neural networks. Our silicon eye is capable of not only able to compensate for defocus, but it can also compensate for other aberrations including astigmatism, coma, and spherical aberrations. In addition, the silicon retina acts as a reconfigurable dynamic neural network to enable real-time image processing. The silicon eye uses three key enabling technologies. First, high-speed active pixel photo-diodes are used as photo-detectors for both imaging and for wavefront sensing. The design of the active pixel photo- detectors is described along with experimental results characterizing their performance. Second, the analog signals received from the photo-detectors and processed by the active pixel circuitry is fed into a smart vision chip. The smart vision chip is a reconfigurable neural network capable of real-time reconstruction of the phase information associated with the imaging system. The micro mirrors are active optic devices that can be used to compensate for optical aberrations. Experimental results obtained from the circuit implementation of the dynamical network networks are presented. The experimental results obtained from our intelligent vision system demonstrate that dynamical neural networks offer advantages in speed, cost, size, and power consumption.

The first problem confronting the developers of algorithms for reliable automatic object recognition systems is basic intensity segmentation and noise smoothing. The benefits of using PCNNs for this are described. The next issue for the developer is the mixture of syntactical and statistical techniques. For many, only the latter is included due to the lack of abundance of fast, simple and effective syntactical algorithms. Relational maps and model-based algorithms are generally computationally intensive as compared to a straightforward statistical method such as a classifier net. It is described how the time signals of a nonadaptive PCNN incorporate some syntactical information which in turn has been shown to be compatible with a statistical classifier.

With the onset of inexpensive 2D optical detector arrays, digitized image information is readily available. Some systems, such as AOTF systems, produce up to 256 parallel channels of image information. These hyperspectral data cubes contain portions of target information in separate channels. Rarely, does a single channel contain sufficient target information. Interchannel linking of pulse image generators is capable of segmenting the many channels in concert. Spiral image fusion occurs by phase-encoding pulse images and analyzing through complex fractional power filters.

Mission planning and missile navigation control are important tasks to solve when dealing with cruise missiles. A large variation of solutions has been used all the way back to world war II and the German VI missile. Today, biologically inspired sensor analysis systems such as, e.g. pulsed coupled neural networks (PCNN), can be used in many different applications related to these two major tasks, mission planing and missile navigation. This paper discusses generally the cruise missile related problems and gives example on how they are being solved. New ideas as shown on how to use PCNN in combination with other image processing transforms, e.g. the radon transform, to solve the planning and navigation problems. This includes solving tasks such as image segmentation, target identification and maze navigation.

We make a few simple comparisons of the principles and performance for noise reduction and edge detection with conventional methods versus neural network methods. Noise reduction methods discussed include the wavelet packet transform. Edge detection is discussed from the point of view of the Sobel and Canny transforms. An approach using the IBM ZISC036 neural network chip is also discussed. In all cases, the results are compare to that of the biologically inspired PCNN. An application of the `best if both worlds' is demonstrated in a foveation/object isolation application for ATR.

Pulse coupled neural network (PCNN) algorithms for image preprocessing have minimal requirements for interconnects and essentially no memory requirements. They are an effective technique for rapid intensity segmentation and noise smoothing as an initial step in many image processing algorithms. A test array with 32 X 32 active PCNN pixels has been undergoing initial evaluation. Its performance is discussed and comparison with software versions of the PCNN algorithm is given.

This paper presents a general VHDL implementation of a Pulse Coupled Neural Network. The VHDL implementation is targeted for FPGA but can also be used with advantage for ASIC implementations. This particular case deals with images of the size 128 X 128 pixels coming at a rate of 60 images per second, each image iterated by the PCNN 70 times, i.e. a real time image processing system. Thanks to the generality, this suggested solution can easily be transformed into, e.g., a solution with images sized 32 X 32 pixels, coming at a speed of 960 images per second, assuming the same iteration length. The hardware requirement and problems are analyzed and solutions are proposed. Some problems that are dealt with are: the huge amount of data produced, the high throughput (i.e. the rate of new data produced) and the loading of coefficients during runtime.

In this paper, we will describe the basic features and capabilities of the IBM ZISC036, a massively parallel chip which implements the Restricted Coulomb Energy algorithm and the K-Nearest Neighbor algorithm. Both of the aforementioned algorithms, their learning and recognition phases, and the basic architectural structure of this hardware implementation will be discussed. The ZISC036 chip containing thirty-six neurons has the advantages of processing time reduction in comparison with classical models, adaptability, and pattern learning,; it is both easy to program and operate. A neuron is a processor capable of prototype and associated information storage as well as distance computation and communication with other neurons. At the end of this paper to show the advantage of this model and illustrate the principle of the ZISC, we will present two applications of the ZISC, one for image contour extraction, and the other for visual probe mask inspection on wafers.

In this paper, we present a neural based solution developed for noise reduction and image enhancement using the ZISC, an IBM hardware processor which implements the Restricted Coulomb Energy algorithm and the K-Nearest Neighbor algorithm. Artificial neural networks present the advantages of processing time reduction in comparison with classical models, adaptability, and the weighted property of pattern learning. The goal of the developed application is image enhancement in order to restore old movies (noise reduction, focus correction, etc.), to improve digital television images, or to treat images which require adaptive processing (medical images, spatial images, special effects, etc.). Image results show a quantitative improvement over the noisy image as well as the efficiency of this system. Further enhancements are being examined to improve the output of the system.

The visual cortex models previously proposed have some amiable hardware implementation attributes. These networks generally include only local connections (except for perhaps a global constant) and require low precision processing. There are many models of the cortex that have been proposed. Most consist of the same basic mathematical principles, but vary in the details. However, these models tend to behave similarly. This presentation will discuss basic requirements of hardware implementation and explore a few electronic and optical implementations.

Pulsed-Coupled Neural Network (PCNN) is an oscillatory model neural network where grouping of cells and grouping among the groups that form the output time series (number of cells that fires in each input presentation also called `icon'). This is based on the synchronicity of oscillations. Recent work by Johnson and others demonstrated the functional capabilities of networks containing such elements for invariant feature extraction using intensity maps. PCNN thus presents itself as a more biologically plausible model with solid functional potential. This paper will present the summary of several projects and their results where we successfully applied PCNN. In project one, the PCNN was applied for object recognition and classification through a robotic vision system. The features (icons) generated by the PCNN were then fed into a feedforward neural network for classification. In project two, we developed techniques for sensory data fusion. The PCNN algorithm was implemented and tested on a B14 mobile robot. The PCNN-based features were extracted from the images taken from the robot vision system and used in conjunction with the map generated by data fusion of the sonar and wheel encoder data for the navigation of the mobile robot. In our third project, we applied the PCNN for speaker recognition. The spectrogram image of speech signals are fed into the PCNN to produce invariant feature icons which are then fed into a feedforward neural network for speaker identification.

This paper presents an approach of large neural net emulation by using new neuroprocessor NM6403, designed by RC Module. Model of neural net, hardware supported by neuroprocessor architecture, is discussed and some key features of NM6403, such as processing of data with variable bit length and two 64-bit external buses, are highlighted. Examples of emulation of neural nets by one neuroprocessor are shown and its performance is estimated. Proposals are also examined how to build large super parallel computing systems with NM6403 as a basic block. There are two hardware supported ways to connect neuroprocessor with other ones: using of shared memory mode of any of two external buses or two communication ports, compatible with those of DSP TMS320C4x. These abilities allow to build various parallel structures as trees, rings, grids and so on.

The paper represents an architecture overview of the NeuroMatrix NM6403 neuroprocessor designed for 32-bit and 64-bit data processing. The paper includes brief description of the neuroprocessor pinout, structure and functional units. The neuroprocessor comprise original RISC core, vector coprocessor (VCP) and some peripheral units. RISC core provides general control functions, 32-bit program and data address generation, 32-bit arithmetic, logic and shift operations. The main neuroprocessor operational unit is VCP, applied for variable bit-length vector data arithmetic, logic and saturation operations. The base VCP operation is matrix by vector multiplication with accumulation. Each data vector is a 64-bit word of packed data word. It is formed by set of variable bit length operands with user defined bit length in a range from 1 to 64 bits. Neuroprocessor includes two external 64-bit buses. The programmable memory interface units allow to use static or dynamic memory having wide range of time parameters without external controller. The neuroprocessor support shared memory mode for each of the external buses. With conjunction of two byte width communication ports this one makes it easy to design multiprocessor systems. Also this paper represents addressing modes, instruction set, supported interrupts. The neuroprocessor is designed using CMOS 0.5 micrometers technology, power supply voltage is 3.3 V, clock rate is 50 MHz with one instruction per clock cycle performance.

In spite of the simple classification concept, impressing performances have been reported using the n-tuple architecture in combination with a very simple training strategy. In general, however, the performance of the n- tuple classifier is highly dependent on the choice of input connections and on the encoding of the input data. Accordingly, the simple architecture needs to be accompanied with design tools for obtaining a suitable architecture. Due to the simplicity of the architecture, it is simple to perform leave-one-out cross-validation tests and extensions of the concept. Therefore, it is also possible to operate with design methods that make extensively use of such tests. This paper describes such design algorithms and especially introduces a simple design strategy that allows the n-tuple architecture to perform satisfactorily in cases with skewed class priors. It can also help to resolve conflicts in the training material. The described methods are evaluated on classification problems from the European StatLog project. It is hereby shown that the design tools extends the competitiveness of the n-tuple classification method.

In a general purpose pulse coupled neural network (PCNN) algorithm the following parameters are used: 2 weight matrices, 3 time constants, 3 normalization factors and 2 further parameters. In a given application, one has to determine the near optimal parameter set to achieve the desired goal. Here a simplified PCNN is described which contains a parameter fitting part, in the least squares sense. Given input and a desired output image, the program is able to determine the optimal value of a selected PCNN parameter. This method can be extended to more general PCNN algorithms, because partial derivatives are not required for the fitting. Only the sum of squares of the differences is used.

This paper presents the double loop feedback model, which is used for structure and data flow modeling through reinforcement learning in an artificial neural network. We first consider physiological arguments suggesting that loops and double loops are widely spread in the exchange flows of the central nervous system. We then demonstrate that the double loop pattern, named a mental object, works as a functional memory unit and we describe the main properties of a double loop resonator built with the classical Hebb's law learning principle in a feedforward basis. In this model, we show how some mental objects aggregate themselves in building blocks, then what are the properties of such blocks. We propose the mental objects block as the representing structure of a concept in a neural network. We show how the local application of Hebb's law at the cell level leads to the concept of functional organization cost at the network level (upward effect), which explains spontaneous reorganization of mental blocks (downward effect). In this model, the simple hebbian learning paradigm appears to have emergent effects in both upward and downward directions.

Recent work on biologically motivated networks have shown that the visual system can process a natural scene more quickly by encoding the order of neural firing rather than the frequency of firing. This `order of firing' encoding scheme has led to a rank-based approach which converts activation energy into a time-dependent pulse code. This paper focuses towards the contribution of unsupervised learning to the training of integrate and fire neurons within multi-layer networks. First, we propose an unsupervised learning algorithm and we test it on a simple recognition task. Then, we propose a multilayer architecture of integrate and fire neurons to solve a more complex vision task. This architecture is efficiently trained by an algorithm combining supervised and unsupervised rank-based hebbian learning. Further improvements are proposed in the final discussion.

Pulse-Couple Neural Networks have generated quite a bit of interest as image processing tools. Past applications include image segmentation, edge extraction, texture extraction, de-noising, object isolation, foveation and fusion. These past applications do not comprise a complete list of useful applications of the PCNN. Future avenues of research will include level set analysis, binary (optical) correlators, artificial life simulations, maze running and filter jet analysis. This presentation will explore these future avenues of PCNN research.

This is a report on work in progress. Spectral recognition is central to many areas of science and technology. Classical spectral recognition analysis techniques (least squares, partial least squares, etc.) are sensitive to offset and gain drifts and errors. This sensitivity can cause excessive costs for spectrometer resources and calibrations. Neural techniques relieve some of this sensitivity but none approach human competence. It is desirable to mimic human spectral analysis not only to improve the results but to minimize detector constraints and costs. We suggest that the first step in human analysis is peak detection. We are exploring the 1D PCNN as a peak segmenter for spectral peak finding in the presence of noise and drifts in gain and offset. We present results of 1D pulse coded neural network peak detection with both simulated and actual static spectra. We also use the PCNN to form a scale and translation invariant feature vector that may be decomposed using classical techniques such as least squares. Finally, we propose using a PCNN to exploit the temporal aspects of spectral acquisition.

Here we describe an algorithm for sea ice classification in the Baltic Sea based on C-band SAR-images. The algorithm is a two-scale algorithm based on neural networks, two additional features and expert system like rules. We have tested and evaluated the algorithm based on about one hundred SAR images from the winter 1998.

A new method for extracting features from photographic images has been developed. The input image is through a pulse coupled neural network transformed to a set of signatures, well suited for classification by unsupervised neural networks. A strategy using multiple self-organizing feature maps in a hierarchical manner is developed. With this approach, using a certain degree of supervision, an acceptable classification is obtained when applied to test images. The method is applied to license plate recognition.

Sonar detection and classification of sunken wrecks and other objects is of keen interest to many. This paper describes the use of neural networks (NN) for locating, classifying and determining the alignment of objects on a lakebed in Sweden. A complex program for data preprocessing and visualization was developed. Part of this program, The Sonar Viewer, facilitates training and testing of the NN using (1) the MATLAB Neural Networks Toolbox for multilayer perceptrons with backpropagation (BP) and (2) the neural network O-Algorithm (OA) developed by Age Eide and Thomas Lindblad. Comparison of the performance of the two neural networks approaches indicates that, for this data BP generalizes better than OA, but use of OA eliminates the need for training on non-target (lake bed) images. The OA algorithm does not work well with the smaller ships. Increasing the resolution to counteract this problem would slow down processing and require interpolation to suggest data values between the actual sonar measurements. In general, good results were obtained for recognizing large wrecks and determining their alignment. The programs developed a useful tool for further study of sonar signals in many environments. Recent developments in pulse coupled neural networks techniques provide an opportunity to extend the use in real-world applications where experimental data is difficult, expensive or time consuming to obtain.

Image segmentation is one of the major application areas for Pulsed Coupled Neural Networks (PCNN). Previous research has shown that the ability of PCNN to ignore minor variations in intensity and small spatial discontinuities in images is beneficial to image segmentation as well as image smoothing. This paper describes research and development projects in progress in which PCNN is used for the segmentation of three different types of digital images. The software for the diagnosis of Pulmonary Embolism from VQ lung scans uses PCNN in single burst mode for segmenting perfusion and ventilation images. The second project is attempting to detect ischemia by comparing 3D SPECT (Single Photon Emission Computed Tomography) images of heart obtained during stress and rest conditions, respectively. The third application is a space science project which deals with the study of global auroral images obtained from Ultraviolet Imager. The paper also describes an hardware implementation of PCNN as an electro-optical chip.

A fuzzy logic-based system to classify olfactive signals is presented. The odor samples are obtained from an electronic noise that contains conducting polymer sensors with partially overlapping sensitivities to odors. The sensor responses are represented by means of the coefficients of their Fast Fourier Transform (FFT). A feature reduction method is applied to reduce the feature space dimension. Then, an unsupervised Fuzzy Divisive Hierarchical Clustering (FDHC) method is used to establish the optimal number of clusters in the data set as well as the optimal cluster structure. The output of FDHC is a binary hierarchy of fuzzy classes that are used to build a supervised fuzzy hierarchical classifier. At each level of the fuzzy hierarchy a separating hyperplane of the two corresponding fuzzy training classes is determined. The hyperplane identifies two crisp decision regions, which will be refined at the next level of the hierarchy. In this way, we obtain a hierarchy of regions, which defines a crisp decision tree. Each region is, therefore, related to a specific expected output of the system. Recognition of an unknown odor is accomplished by computing the FFT of the corresponding signal and using the decision tree to establish the region the odor belongs to. Two small-scale applications of the method yielded 100% classification accuracy on out-of-sample data.

Principal component analysis (PCA) and artificial neural networks are used to investigate electronic gas sensor responses for various alcohol chemicals. PCA is used to identify and visualize the best features to use for classification as well as for detecting outliers. A regular feed forward back propagation neural network (FBP) was used for the actual classification due to the fact that FBP determines better the non-linear borders of the various region of interest involved in the classification. Furthermore, we consider the tradeoff between classification speed and accuracy.

Sleep apnea is characterized by frequent prolonged interruptions of breathing during sleep. This syndrome causes severe sleep disorders and is often responsible for development of other diseases such as heart problems, high blood pressure and daytime fatigue, etc. After diagnosis, sleep apnea is often successfully treated by applying positive air pressure (CPAP) to the mouth and nose. Although effective, the (CPAP) equipment takes up a lot of space and the connected mask causes a lot of inconvenience for the patients. This raised interest in developing new techniques for treatment of sleep apnea syndrome. Several studies indicated that electrical stimulation of the hypoglossal nerve and muscle in the tongue may be a useful method for treating patients with severe sleep apnea. In order to be able to successfully prevent the occurrence of apnea it is necessary to have some technique for early and fast on-line detection or prediction of the apnea events. This paper suggests using measurements of respiratory airflow (mouth temperature). The signal processing for this task includes the use of a window short-FFT technique and uses an artificial back propagation neural net to model or predict the occurrence of apneas. The results show that early detection of respiratory interruption is possible and that the delay time for this is small.

The objective of this work is to experiment neural networks methodology on noisy fingerprint identification. Several models for fingerprint codification based on minutiae characteristics were designed and implemented in this work. The ENH preprocessing software was used and tested and several improvements were implemented. The Competitive Selective neural network MCSL was used in the codification procedure in all models.

The energy event reconstruction for non-leptonic decays of the Z in the DELPHI experiment is a complicated task. The energy resolution is dependent considerably from the neutral hadrons energy measurement in the electromagnetic and hadron calorimeter of the detector.

Asynchronous Transfer Mode (ATM) has been recommended by ITU-T as the transport method for broadband integrated services digital networks. In high-speed ATM networks different types of multimedia traffic streams with widely varying traffic characteristics and Quality of Service (QoS) are asynchronously multiplexed on transmission links and switched without window flow control as found in X.25. In such an environment, a traffic control scheme is required to manage the required QoS of each class individually. To meet the QoS requirements, Bandwidth Allocation and Call Admission Control (CAC) in ATM networks must be able to adapt gracefully to the dynamic behavior of traffic and the time-varying nature of the network condition. In this paper, a Neural Network approach for CAC is proposed. The call admission problem is addressed by designing controllers based on Neural Tree Networks. Simulations reveal that the proposed scheme is not only simple but it also offers faster response than conventional neural/neuro-fuzzy controllers.

The APWS allows user friendly access to several legacy systems which would normally each demand domain expertise for proper utilization. The generalized model, including objects, classes, strategies and patterns is presented. The core components of the APWS are the Microsoft Windows 95 Operating System, Oracle, Oracle Power Objects, Artificial Intelligence tools, a medical hyperlibrary and a web site. The paper includes a discussion of how could be automated by taking advantage of the expert system, object oriented programming and intelligent relational database tools within the APWS.

This paper introduces a way to locate persons in visual images of cluttered scenes using a shape-of-contour approach. The contour which we refer to is that of the upper body of frontally aligned persons.

For hadron calorimeters with a transverse structure there exists a possibility to reconstruct the particles energy with a better resolution using the neural network algorithm. For the calorimeter with a novel longitudinal structure that capability of the neural network method for the better determination of the particles energy in comparison with the traditional method was studied. The research is based on the information from the experiment at IHEP (Serpukhov) with the test (pi) ^--beam with energies 10, 20, 30 and 40 GeV. Using of the neural network improve the energy resolution of a system of electromagnetic calorimeter and hadron calorimeter with scintillators parallel to beam.

A project has been underway to add digital imaging and processing to the inspection of nuclear fuel by the International Atomic Energy Agency. The ultimate goals are to provide the inspector not only with the advantages of CCD imaging, such as high sensitivity and digital image enhancements, but also with an intelligent agent that can analyze the images and provide useful information about the fuel assemblies in real time. The project is still in the early stages and several interesting sub-projects have been inspired. Here we give first a review of the work on the fuel assembly image analysis and then give a brief status report on one of these sub-projects that concerns automatic categorization of fuel assembly images. The technique could be of benefit to the general challenge of image categorization.

We discuss possible new hardware and software techniques for handling very large databases such as image archives. In particular, we investigate how high capacity solid-state `disks' could be used to speed the database processing by algorithms that require considerably memory space. One such algorithm, for example, called the RAM neural network, or weightless neural network, needs a number of large lookup tables to perform most efficiently. The solid state disks could provide fast storage both for the algorithm and the data. We also briefly discuss development of an algorithm to cluster images of similar objects. This algorithm could also benefit from a large cache of fast memory storage.

Classical adaptive control proves total-system stability for control of linear plants, but only for plants meeting very restrictive assumptions. Approximate Dynamic Programming (ADP) has the potential, in principle, to ensure stability without such tight restrictions. It also offers nonlinear and neural extensions for optimal control, with empirically supported links to what is seen in the brain. However, the relevant ADP methods in use today--TD, HDP, DHP, GDHP--and the Galerkin-based versions of these all have serious limitations when used here as parallel distributed real-time learning systems; either they do not possess quadratic unconditional stability (to be defined) or they lead to incorrect results in the stochastic case. (ADAC or Q- learning designs do not help.) After explaining these conclusions, this paper describes new ADP designs which overcome these limitations. It also addresses the Generalized Moving Target problem, a common family of static optimization problems, and describes a way to stabilize large-scale economic equilibrium models, such as the old long-term energy mode of DOE.

Virtual Intelligence (VI) is defined as neural networks, fuzzy systems, evolutionary computation and virtual reality. This overview provides motivational material and an update on research. Some potential interactions of VI components are described. Hot topics and new approaches such as Pulse Coupled Neural Networks are reviewed. Pointers to websites and printed material are provided.