The MLP Iterative Construction Algorithm (MICA) designs a Multi-Layer Perceptron (MLP) neural network as it trains. MICA adds Hidden Layer Nodes one at a time, separating classes on a pair-wise basis, until the data is projected into a linear separable space by class. Then MICA trains the Output Layer Nodes, which results in an MLP that achieves 100% accuracy on the training data. MICA, like Backprop, produces an MLP that is a minimum mean squared error approximation of the Bayes optimal discriminant function. Moreover, MICA's training technique yields novel feature selection technique and hidden node pruning technique

A new universal one-chart smooth manifold model for vector information sources is introduced. Natural coordinates (a particular type of chart) for such data manifolds are then defined. Uniformly quantized natural coordinates form an optimal vector quantization code for a general vector source. Replicator neural networks (a specialized type of multilayer perceptron with three hidden layers) are the introduced. As properly configured examples of replicator networks approach minimum mean squared error (e.g., via training and architecture adjustment using randomly chosen vectors from the source), these networks automatically develop a mapping which, in the limit, produces natural coordinates for arbitrary source vectors. The new concept of removable noise (a noise model applicable to a wide variety of real-world noise processes) is then discussed. Replicator neural networks, when configured to approach minimum mean squared reconstruction error (e.g., via training and architecture adjustment on randomly chosen examples from a vector source, each with randomly chosen additive removable noise contamination), in the limit eliminate removable noise and produce natural coordinates for the data vector portions of the noise-corrupted source vectors. Consideration regarding selection of the dimension of a data manifold source model and the training/configuration of replicator neural networks are discussed.

This paper considers the relevance of nonlinear feature extraction for the analysis of real-world single channel wake EEG signals. It is demonstrated that it is feasible to extract structured patterns which possibly reflect the state of mind of the subject. This is exhibited by a clustering and a dynamics in a feature space derived by a dynamical systems approach of projecting the information into the space spanned by the lowest order singular vectors determined from a matrix of delay vectors. An embedding of the signal was obtained in a 11-dimensional Euclidean space indicating a relatively small number of intrinsic degrees of freedom in the data. Feature extraction and clusterings in the signal have been obtained using linear methods (principal component projections) and nonlinear approaches (the neural network technique known as `NEUROSCALE'). Although most of the analysis was performed in an unsupervised manner (without using any task-specific information), a final clustering was demonstrated which used some of the task-related knowledge to obtain more distinct clusters. The interesting aspect was that in both linear and nonlinear methods the characteristic clusters did not align themselves in an order which reflected the time of day of the tasks, or even the type of tasks. Our supposition is that the self-organized clusters are driven by a higher level cognitive state such as the `attentiveness' of the subject through no data is available to test the hypothesis.

A new feature space trajectory (FST) description of 3D distorted views of an object is advanced for active vision applications. In an FST, different distorted object views are vertices in feature space. A new eigen-feature space and Fourier transform features are used. Vertices for different adjacent distorted views are connected by straight lines so that an FST is created as the viewpoint changes. Each different object is represented by a distinct FST. An object to be recognized is represented as a point in feature space; the closest FST denotes the class of the object, and the closest line segment on the FST indicates its pose. A new neural network is used to efficiently calculated distances. We discuss its uses in active vision. Apart from an initial estimate of object class and pose, the FST processor can specify where to move the sensor to: confirm close and pose, to grasp the object, or to focus on a specific object part for assembly or inspection. We advanced initial remarks on the number of aspect views needed and which aspect views are needed to represent an object.

A study of time series predictions has been conducted using artificial neural networks and several concepts of non- linear dynamics. The ability of the proposed approach for modelling the underlying dynamics of complex systems in the presence of noisy signals is demonstrated. In our approach the embedding theorem of non-linear dynamics provides a basis for determining the time delay ((tau) ) and the embedding dimension of the complex system. The embedded vectors are then extracted from the time series data by the method of delays. Experiments were conducted by introducing gaussian noise of varying degrees in the systems with known dynamics and the robustness of multistep predictions was studied. Various ANN architectures including backpropagation, backpropagation through time, recurrent networks, etc. were studied and their performances are compared. Several applications including sine series prediction and chaotic systems such as the Lorenz series are presented. For a deterministic system like the sine series prediction, it was observed that the neural network was able to extract salient features such as frequency, amplitude and phase in case of relatively low Signal to Noise Ratio (SNR). In case of the chaotic system, the ANN model extracts the dynamics for high SNR values. Realistic application to prediction of sunspot numbers is also discussed.

Despite dramatic recent advances in speech recognition technology, speech recognition still perform much worse than humans. The difference in performance between humans and machines is most dramatic when variable amounts and types of filtering and noise are present during testing. For example, humans readily understand speech that is low-pass filtered below 3 kHz or high-pass filtered above 1 kHz. Machines trained with wide-band speech, however, degrade dramatically under these conditions. An approach to compensate for variable unknown sharp filtering and noise is presented which uses mel-filter-bank magnitudes as input features, estimates the signal-to-noise ratio (SNR) for each filter, and uses missing feature theory to dynamically modify the probability computations performed using Gaussian Mixture or Radial Basis Function neural network classifiers embedded within Hidden Markov Model recognizers. The approach was successfully demonstrated using a talker-independent digit recognition task. It was found that recognition accuracy across many conditions rises from below 50% to above 95% with this approach. These promising results suggest future work to dynamically estimate SNR's and to explore the dynamics of human adaptation to channel and noise variability.

The Decompositional Hierarchical Self-Organizing Network (DHSON) is derived from earlier research which developed the Hierarchical Self-Organizing Network. DHSON decomposes input vectors and creates a separate multi-layer 1D self-organized mapping for each component. This approach eliminates the scaling problems typical of Kohonen-like architectures. The end objective is for DHSON to prepare input data for presentation to recurrent networks developed through evolutionary strategies by reducing dimensionality, deriving an effective data encoding for parallel processing, and/or reducing complexity within a data set.

Accurate total system electrical load forecasting is a necessary part of resource management for power generation companies. The better the hourly load forecast, the more closely the power generation assets of the company can be configured to minimize the cost. Automating this process is a profitable goal and neural networks should provide an excellent means of doing the automation. However, prior to developing such a system, the optimal set of input parameters must be determined. The approach of this research was to determine what those inputs should be through a parametric study of potentially good inputs. Input parameters tested were ambient temperature, total electrical load, the day of the week, humidity, dew point temperature, daylight savings time, length of daylight, season, forecast light index and forecast wind velocity. For testing, a limited number of temperatures and total electrical loads were used as a basic reference input parameter set. Most parameters showed some forecasting improvement when added individually to the basic parameter set. Significantly, major improvements were exhibited with the day of the week, dew point temperatures, additional temperatures and loads, forecast light index and forecast wind velocity.

Neural networks represent a promising approach to time series forecasting; however, the problem of obtaining good network generalization continues to present a challenge. As a means of improving network generalization ability for time series forecasting applications, this paper investigates the utility of a biologically inspired scheme that employs receptive fields for encoding network inputs. Both single- and multi-step forecasting performances are studied in the context of the sunspot series. Additionally, a heuristic for selecting the placement and dilations of the receptive field functions is presented. The performance of multi-layered perceptron networks trained using the data arising from the encoding scheme is assessed. The heuristic for placing and dilating the receptive fields yielded networks that learn rapidly and have consistently good multi-step prediction capability as compared to other published results.

Radial basis function neural network models of a time series may be developed or trained using samples from the series. Each model is a continuous curve that can be used to represent the series or predict future vales. Model development requires a tradeoff between a measure of roughness of the curve and a measure of its error relative to the samples. For roughness defined as the root integrated squared second derivative and for error defined as the root sum squared deviation (which are among the most common definitions), an optimal tradeoff conjecture is proposed and illustrated. The conjecture states that the curve that minimizes roughness subject to given error is a weighted mean of the least squares line and the natural cubic spline through the samples.

In this work we describe a method of estimating and characterizing appropriate data and model complexity in the context of long term iterated time series forecasting using embeddings and multiple time-scale decomposition techniques. An embedding of a signal is obtained which decouples multiple time scale effects such as seasonality and trend. The complexity and stability of networks are estimated and the performance of long term iteration is examined. The performance of the technique is tested using the real world time series problems of electricity load forecasting, and financial futures contracts.

Predicting time series in the economic field calls for special techniques. Instead of having the feed-forward network predict the future value, the problem is cast as a classification task. Codings are proposed that give a probability density of tomorrows change in value. Methods to decide for the most likely or profitable direction of the change of value are introduced. The methods have been used to predict the daily change for the Bundfuture.

A hybrid approach, incorporating concepts of nonlinear dynamics in artificial neural networks, is proposed to study complex dynamic system for more accurate predictions of time series. In this approach important features of nonlinear dynamics are used in the framework of neural networks to construct a model for the underlying dynamics of time series. During pre-analysis phase the series has been characterized in terms of auto-correlation, power spectrum, average mutual information, number of false nearest neighbors, Lyapunov exponents, DVS (deterministic Vs stochastic) plot, and surrogate data sets to find if the series is nonlinear, deterministic or chaotic. The series is then projected in the embedding space by constructing embedding vectors using the method of delays. We examine the dynamics of the system in the embedding space, and note that the time development now follows vector to vector mapping. The artificial neural network has been used to obtain this mapping function. It was observed that time series prediction is better if vector at time T is mapped with more than one vectors immediately preceding it. We illustrate our method by considering a time series generated from the Lorenz equations. The model thus developed gave excellent quality of fit both for the training and test sets. Using the model, the multistep predictions for future 50 values have been made and are found to be very accurate.

This paper discusses the developed Adaptive Neural Control architecture for on-line system identification and real-time adaptive control. After reviewing existing literature involving controls of structural vibration, we report new developments carried out under the adaptive neural control program for the USAF Phillips Laboratory. The new results include a neural control architecture suitable for MIMO systems subjected to tonal disturbances that is capable of optimizing vibration suppression in the presence of sensor or actuator failures. This architecture was demonstrated in a series of tests on the ASTREX facility.

In 1995 we have presented a new approach to intelligent control based on Restricted Coulomb Energy concept which derives from Radial Basis Function (RBF) like neural network. A parallel implementation of adaptive process control using this technique has been discussed. In this paper we present a Zero Instruction Set Computer based parallel implementation of the real time intelligent adaptive control (using RBF like neural network). We expose the architecture of our parallel neural controller discussing the neurons consumption (use up) in learning phase and it's capabilities in the operation phase (relaxation). Analysis of feature space (parameter space) mapping obtained from experimental validation and experimental results are presented validating our concepts.

The use of neurocomputing to solve real world problems is some times penalized by the long computing time, especially in the training phase of the neural network. A lot of approaches were suggested in the literature that allows to reduce the computing time and to enhance the generalization factor. The approach we propose in this paper is original by the fact that it is based on the following paradigm: Divide To Simplify or DTS. The task of learning is decomposed in 2 steps. In the first step, we analyze the set of the input patterns and decompose the input space in several regions of interest. This is done by using an unsupervised prototype based neural network: a Kohonen Self Organized feature Map (SOM). The second step consists of training several feature based networks to learn the behavior of each region of interest. In this way, we obtain a set of specialized neural networks. We will present in this paper various data driven methods that use the DTS paradigm to build efficient Multi- Neural Networks systems. The efficiency of this context means fast learning time, fast execution or relaxation time, enhanced generalization factor and intrinsic implementation on a parallel machine. Our approach has been used, with success, in an industrial real world process control to analyze such a complex problem with multi-variable data.

This paper presents that experimental results obtained on indoor tissue culture using the adaptive image segmentation system. The performance of the adaptive technique is contrasted with different non-adaptive techniques commonly used in the computer vision field to demonstrate the improvement provided by the adaptive image segmentation system.

This paper presents a new approach for structural adaptation in multi-layer neural networks in general and the application of the proposed method to land mine target detection and classification problem. The new algorithm uses time and order update formulations of the orthogonal projection theorem to derive a recursive weight updating procedure and architectural variation of the network during the training process. The proposed approach provides optimal network structure in the sense that the mean-squared error is minimized for the newly created topology. This algorithm is used in conjunction with a data representation scheme to perform land mine target detection and classification. The simulation results on targets with different composition indicated superior detection and classification performance when compared to the conventional methods.

This paper describes a method for classifying image pixels as belongings to one of several known classes or as being unrecognized. A Haar wavelet transform is utilized to generate a feature vector per image pixel. This provides information about the local brightness and color, as well as about the texture of the surrounding area. Hand-labeled images are used to generate a codebook using the Optimal Learning Rate Learning Vector Quantization algorithm. The system is applied to outdoor images recorded with a video camera from a moving automobile. The classification is used to produce annotated images, in which recognized areas are replaced by uniform colors denoting the corresponding class, and unrecognized areas are left unchanged. For small numbers of classes, the pixel classification accuracy is as high as 99%.

VIGILANTE is an automated recognition and tracking system that closely integrates a sensing platform with a very large processing capability (over 2 TeraOPS). The architecture currently consists of an optical bench with multiple sensors, a large parallel analog pre-processor, and a digital 512 processor, parallel machine. Preliminary results on target detection and orientation are presented for an algorithm that is suitable for the VIGILANTE architecture. The technique makes use of eigenvectors calculated from image blocks (size 32 X 32) drawn from video sequences containing rocket targets. The eigenvectors are used to reduce the dimensionality of frame-lets (size 32 X 32) from the larger sensor images. These frame-lets are projected on to the eigenvectors and the resultant values are then used as an input pattern to a feed forward neural network classifier. A description and evaluation of this algorithm (including precision limitation) with respect to VIGILANTE is provided. Experiments using this technique have generated near 99target and non-target images and close to 97% identification of the rocket type.

The recognition of human faces is a phenomenon that has been mastered by the human visual system and that has been researched extensively in the domain of computer neural networks and image processing. This research is involved in the study of neural networks and wavelet image processing techniques in the application of human face recognition. The objective of the system is to acquire a digitized still image of a human face, carry out pre-processing on the image as required, an then, given a prior database of images of possible individuals, be able to recognize the individual in the image. The pre-processing segment of the system includes several procedures, namely image compression, denoising, and feature extraction. The image processing is carried out using Daubechies wavelets. Once the images have been passed through the wavelet-based image processor they can be efficiently analyzed by means of a neural network. A back- propagation neural network is used for the recognition segment of the system. The main constraints of the system is with regard to the characteristics of the images being processed. The system should be able to carry out effective recognition of the human faces irrespective of the individual's facial-expression, presence of extraneous objects such as head-gear or spectacles, and face/head orientation. A potential application of this face recognition system would be as a secondary verification method in an automated teller machine.

In this paper, we will present a Neural Net Benchmark. In many neural algorithms, learning is defined as a change in connection weight values that result in the capture of training information that can later be recalled. A typical learning algorithm is: if neuron A was active and A's activity caused a connected neuron B to fire, then the synaptic connection between A and B should be increased by certain amount. The benchmark systematically tests a neural network for this ability. The benchmark can be written as: (L3, L4, L5, L6, ... ...). L3 tests the net's ability for 3- neuron correlation: if A and B were active and A and B's activity caused a connected neuron C to fire, can the network recall it later? Similarly, L4, L5, ... tests a net for 4-neuron, 5-neuron correlation. We will also present a neural net classification based on the benchmark.

This paper investigates a traditional clustering algorithm (K-means) and two neural networks (SOM and ART-F). The characteristics of each algorithm are illustrated by simulating geometric space data clustering. Then each algorithm is applied to image data sets to compress the size by reducing the number of colors from 256 to 16.

This paper focuses on use of a new image filtering technique, Pulsed Coupled Neural Network factoring to enhance both the analysis and visual interpretation of noisy sinusoidal time signals, such as those produced by LLNL's Microwave Impulse Radar motion sensor. Separation of a slower, carrier wave from faster, finer detailed signals and from scattered noise is illustrated. The resulting images clearly illustrate the changes over time of simulated heart motion patterns. Such images can potentially assist a field medic in interpretation of the extent of combat injuries. These images can also be transmitted or stored and retrieved for later analysis.

Sequences of thermographic images of burning residue produced by M198 155 (unicharge) test rounds fired at Yuma Proving Ground (YPG) have been collected for analysis with the objective being to elucidate the evolution of conditions in the breech after firing and to provide guidance in determining safe loading protocols for future autoloaders. In order to better understand the thermal environment in the breech, we are developing advanced analytical tools that can be used to quantitatively characterize sequences of thermographic images. However, the calibration data required to extract the temperature profiles of the YPG thermographic images for these analyses is unavailable. No analytic solution could be determined to perform the highly nonlinear reverse transformation from RGB space to intensities, therefore, a neural network has been employed. Furthermore, the experimental data provided by YPG is only measurable over a restricted range of temperatures extending from approximately 80 degree(s)C up to 110 degree(s)C. Since the highest temperatures measured in the thermographic data do not correspond to a hazardous condition, more complex measures than simple statistical averages of the temperature must be employed. A new numerical technique is introduced for measuring the scaling properties of single valued surfaces in 3-space represented by sparse data sets.

Neural networks are well known for their ability to perform pattern recognition tasks. This paper discusses the use of parallel neural network hardware for performing pattern recognition tasks. We address the need for neural network hardware and how it can dramatically improve system performance both in training and in actual applications. The use of specialized parallel processing hardware is discussed as well as alternative hardware and software approaches. Finally we give some comparisons between multi-processor computer architecture, Pentium class microcomputers and custom hardware.

A compact VLSI neural computer integrated with an active pixel sensor has been under development to mimic what is inherent in biological vision systems. This electronic eye- brain computer is targeted for real-time machine vision applications which require both high-bandwidth communication and high-performance computing for data sensing, synergy of multiple types of sensory information, feature extraction, target detection, target recognition, and control functions. The neural computer is based on a composite structure which combines Annealing Cellular Neural Network (ACNN) and Hierarchical Self-Organization Neural Network (HSONN). The ACNN architecture is a programmable and scalable multi- dimensional array of annealing neurons which are locally connected with their local neurons. Meanwhile, the HSONN adopts a hierarchical structure with nonlinear basis functions. The ACNN+HSONN neural computer is effectively designed to perform programmable functions for machine vision processing in all levels with its embedded host processor. It provides a two order-of-magnitude increase in computation power over the state-of-the-art microcomputer and DSP microelectronics. A compact current-mode VLSI design feasibility of the ACNN+HSONN neural computer is demonstrated by a 3D 16X8X9-cube neural processor chip design in a 2-micrometers CMOS technology. Integration of this neural computer as one slice of a 4'X4' multichip module into the 3D MCM based avionics architecture for NASA's New Millennium Program is also described.

After a frequency-domain-preprocessing of any standard 2D images, the standard patterns to be learned can generally be presented by N-D analog vectors. When these vectors satisfy the most general separability condition, the positive linear independency (or PLI) condition, these vectors can definitely be learned by a one-layered, hard-limited perceptron in a noniterative manner. From this noniterative learning scheme, an automatic feature extraction and automatic feature competition method can be derived which will enhance significantly the robustness in the recognition mode. This paper reports the theoretical analysis of this novel feature extraction scheme and the experimental result demonstrating the robustness of recognition and the speed of learning of this scheme. Comparison of this novel feature extraction scheme with Karhunen-Loeve scheme and the SVD (singular value decomposition) scheme is discussed in detail.

A derivative of the Fukashima Neocognitron was trained with a set of preprocessed target image chips for the purpose of target classification. The Neocognitron was chosen because of robust performance on handwritten characters that contain contours and complex line shapes not unlike the processed target images used in our multi-sensor/multi-look information fusion system. Advantages of the Neocognitron include translation and distortion invariance, which are desirable properties of any classifier. This paper represents our initial results and conclusions from experiments and data analysis using the Neocognitron on FLIR target sets of ground vehicles.

The use of neural networks in information retrieval and text analysis has primarily suffered from the issues of adequate document representation, the ability to scale to very large collections, dynamism in the face of new information and the practical difficulties of basing the design on the use of supervised training sets. Perhaps the most important approach to begin solving these problems is the use of `intermediate entities' which reduce the dimensionality of document representations and the size of documents collections to manageable levels coupled with the use of unsupervised neural network paradigms. This paper describes the issues, a fully configured neural network-based text analysis system--dataHARVEST--aimed at data mining text collections which begins this process, along with the remaining difficulties and potential ways forward.

A neural network methodology was developed and evaluated for performing vehicle classification with multisensor data collected from simulated unattended ground based sensors. A high fidelity computer simulation, called Sensor Evaluation Model (SENSEM), was used to create a battlefield scenario realistically simulating sensor-vehicle interactions consisting of three sensor types and four vehicle types. Data generated from sensor-vehicle interactions were preprocessed and used to train and test three types of neural networks. Backpropagation, probabilistic, and radial basis function networks were evaluated to determine differential performance effectiveness under various training conditions. Optimal network designs were experimentally determined for each network type for each training condition. An overall comparative analysis of classification accuracy and computational efficiency was then conducted to evaluate the performance between the network types' optimal network designs for each training condition. Results showed extremely high accuracy and rapid training and test times for all optimal network designs. Results also showed optimal network performance varied as a function of training conditions facilitating specification of the most effective neural network paradigm under certain conditions. The results of this research support the conclusion that neural network techniques can be applied successfully as part of an analysis subsystem for classifying vehicles with multisensor data obtained from unattended ground based sensors.

Bottlenose dolphins produce individually distinctive `signature whistles'. This paper considers the feasibility of using artificial neural networks to identify individual dolphins using their signature whistles. The proposed scheme uses wavelet analysis to extract salient features from the signature whistles which are then classified by a neural network. Experimental results suggest that wavelet analysis followed by classification by neural networks may provide a reliable means of identifying dolphin signature whistles.

In this paper we propose a neural network based method for the automatic detection of ground targets in airborne infrared linescan imagery. Instead of using a dedicated feature extraction stage followed by a classification procedure, we propose the following three step scheme: In the first step of the recognition process, the input image is decomposed into its pyramid representation, thus obtaining a multiresolution signal representation. At the lowest three levels of the Laplacian pyramid a neural network filter of moderate size is trained to indicate the target location. The last step consists of a fusion process of the several neural network filters to obtain the final result. To perform this fusion we use a belief network to combine the various filter outputs in a statistical meaningful way. In addition, the belief network allows the integration of further knowledge about the image domain. By applying this multiresolution recognition scheme, we obtain a nearly scale- and rotational invariant target recognition with a significantly decreased false alarm rate compared with a single resolution target recognition scheme.

Frequency-resolved optical grating (FROG) is a technique for measuring the intensity and phase of ultrashort laser pulses. In FROG, a spectrogram of the pulse is produced from which the intensity and phase of the pulse's electric field is then retrieved using an iterative algorithm. This iterative algorithm performs well for all types of pulses, but it sometimes requires more than a minute to converge, and faster retrieval is desired for many applications. As a faster alternative, we therefore employed a neural network to invert the function that relates the pulse intensity and phase to its FROG trace. In previous work, we showed that a neural network can retrieve simple pulses, described by four or six parameters, rapidly and directly. In this contribution, we discuss our latest attempts to train an artificial neural network for more complex pulse shapes.

We have developed a robust Partial Least-Squares Regression (PLSR) neural network approach to statistical calibration model development. Generalized neural network learning rules derived from a weighted statistical representation error criterion that grows less than quadratically are presented. This optimization criterion allows for higher-order statistics associated with the inputs to be taken into account and also serves to robustify the results when the empirical data contains impulsive and colored noise and outliers. The learning rules presented are considered generalized because they can be used to implement several specialized cases including: robust PLSR, linear PLSR, weighted least-squares, and variance scaling. The same learning rules also implement steepest descent or Newton's method. Newton's method can be used to formulate an adaptive learning rate for training the network.

A kind of fuzzy max-product associative memory network and its learning algorithm are presented in this paper. The connection weight matrix for fuzzy max-product auto- associative memory is determined by the generalized fuzzy solution. Each initial state pattern will be converged another state of the network via the connection weight matrix at one iteration. Fuzzy max-product associative memory network possess strong ability of error-tolerance. The computer simulations show the better performance of the fuzzy max-product associative memory network and its learning algorithm.

Target identification decisions must be `positive'--accurate with high confidence. To achieve this, a classifier must build a model (train) with some data and then generalized its decisions to new data sets. The ability to accurately estimate the true classifier error (or predict classification performance) is contingent on the amount of data available. The more data, the better the error estimate and the more confident the resulting decisions. Ideally, one needs infinite data for modeling and assessing classifier performance; however, this is rarely the case. This paper investigates techniques for improving audio target identification accuracy and confidence with a Multilayer Perceptron (MLP). The first technique, bagging, is a combination of bootstrapping and (classifier) aggregation and the second, decision fusing, combines the decisions of multiple classifiers. The `bagged' identification performance for a subset of the Rome Laboratory Greenflag database is compared to the MLP performance without bagging and that of MLP whose decisions are combined with those of our other classifiers. Both techniques improved identification accuracy, albeit bagging did so only slightly. More importantly, the confidence of the identification decisions were significantly improved by the pooling of evidence inherent in both the bagging and decision fusion processes.

This paper proposes an evolutionary based method to design the feedforward Artificial Neural Networks (ANN). The Evolutionary Programming method evolves concurrently the topology and weight distribution of ANN, accelerates the rate of convergence of the solution obtained, and approaches the global optimal solution, then BP is used to train the exact weight values. The experiments demonstrate the method is promising.

In this paper, we propose the approximate transformable technique, which includes the direct transformation and indirect transformation, to obtain a CPB unified model neural networks for feedforward/recurrent neural networks via Chebyshev polynomials approximation. Based on this approximate transformable technique, we have derived the relationship between the single-layer neural networks and multilayer perceptron neural networks. It is shown that the CPB unified model neural networks can be represented as a functional link networks that are based on Chebyshev polynomials, and these networks use the recursive least squares method with forgetting factor as learning algorithm. It turns out that the CPB unified model neural networks not only has the same capability of universal approximator, but also has faster learning speed than conventional feedforward/recurrent neural networks. Computer simulations show that the proposed method does have the capability of universal approximator in some functional approximation with considerable reduction in learning time.

In a number of applications the approximation, interpolation or nonlinear extrapolation of certain weakly (when every subsequent term of power series expansion is much less than previous one) nonlinear dependencies d(x), where x an arbitrary signal in time, is demanded. The problem of cancellation of nonlinear distortions of a signal in high precision analog engineering can be an example. In such cases it seems to be reasonable to use polynomial-based devices. In this paper the neural network based devices able to perform the operations of approximation, interpolation and nonlinear extrapolation are described. The schemes and working characteristics of a breadboard, based on analog radio components, are presented. Legendre polynomials were offered as basis functions for significant increasing of the speed of the approximator training. The scheme of analog synthesizer of Legendre polynomials was also suggested.

We use the polytomous Boltzmann Machine for time series forecasting and prediction applications. We will simulate the polytomous Boltzmann Machine by using the binary Boltzmann Machine. We will show our software (1) can learn and predict the behavior of deterministic dynamic systems. Given enough training, the software can predict with 100% accuracy; (2) can learn and predict the behavior of the Markov chains (or probabilistic finite state machines). Given enough training, the software can predict with 100% accuracy.

Various forms of neural networks have been applied to identification of non-linear dynamical systems. In most of these methods, the network architecture is set prior to training. In this paper, a method that evolves a symbolic solution for plant models is described. This method uses a evolutionary program to manipulate collections of parse trees expressed in a task specific language. Experiments performed on two unknown plants show this method is competitive with those that train neural networks for similar problems.

The oscillatory model of neuronal networks with dynamical chaos has been examined. The neurons in this model are shown to be considered as the coupled nonlinear oscillators. The instanton-like solution obtained is most applicable for information transmittance due to its small dissipation.

In this paper, we describe a stochastic algorithm which can be used for evaluation of energy function of Hopfield neural networks designed for solving optimization problems. To demonstrate this capability we used the Hopfield neural network described by T. P. Troudet to optimize the operation of a crossbar packet switch. The evaluation of energy function of a Hopfield neural network requires determination of synaptic weights from the energy function which represents the optimization problem under consideration. This process involves several iterations of adjusting the coefficients of the energy function and calculating the synaptic weights from the coefficients of the energy function. The proposed stochastic algorithm allows the designer of a Hopfield network to directly evaluate the energy function and adjust the coefficients without going through the step to determine the synaptic weights. In this paper, we also present the simulation results for evaluation of the energy function described by T. P. Troudet using the stochastic algorithm.

The problem of speech recognition becomes more and more important now. And one of the difficult tasks in that problem is the transformation of an initial signal to set of parameters to be recognized. The optimum orthogonal transform which is the best preserve information is given by the augmented eigenvectors of the data domain covariance matrix. Such transformation, commonly called the Karhunen- Loeve transformation (KLT) is too complex to be installed into speech processing system. By imposing of some restrictions to the input signal, it is possible to simplify calculation of KLT by its reduction to another transformation, which is easily calculated. Moreover, these restrictions allow to use fast algorithms for calculation. Such technique was applied to reach speech signal and correctness of imposed restrictions was proved. The technique of transformation of speech signal into sequence of parameters has been developed and experimentally proved.

Nowadays, neural networks have been applied to various fields for pattern recognitions, controls, speech recognition, etc. But most of them are only concerned with improving properties and performances of those neural networks. And only few works are related with real usages of neural networks, like neural knowledge base and semantic and logic fields. So, on this paper, a knowledge base for neural networks will be studied in a use of obstacle avoidance. By using the back propagation neural networks, it is quite possible to handle semantic and logic sentences. Which means that it is possible to make a certain sense of knowledge base by neural networks. But with this stage, this knowledge base is not flexible enough to treat a problem as our human does. It cannot create a new knowledge base which changes its structure by itself. And in order to have a system which can change its structure by itself, here introduces the ideas of Adoption Factor and Satisfaction Factor. And as the result of this, very interesting behaviors have come out and it has been observed that the networks processes for understanding the atmosphere of obstacles have been improved one repetition by one repetition.

The purpose of autonomous robot is to solve various tasks while adapting its behavior to the variable environment, expecting it is able to navigate much like a human would, including handling uncertain and unexpected obstacles. To achieve this the robot has to be able to find solution to unknown situations, to learn experienced knowledge, that means action procedure together with corresponding knowledge on the work space structure, and to recognize working environment. The planning of the intelligent robot behavior presented in this paper implements the reinforcement learning based on strategic and random attempts for finding solution and neural network approach for memorizing and recognizing work space structure (structural assignment problem). Some of the well known neural networks based on unsupervised learning are considered with regard to the structural assignment problem. The adaptive fuzzy shadowed neural network is developed. It has the additional shadowed hidden layer, specific learning rule and initialization phase. The developed neural network combines advantages of networks based on the Adaptive Resonance Theory and using shadowed hidden layer provides ability to recognize lightly translated or rotated obstacles in any direction.

The World Wide Web has brought wide attention from all over the world. It integrates the information from a distributed environment into a global information universe. As the number of web servers and users continues to grow, it is increasingly difficult to find the desired information even though there are many search tools available. In this study, an Intelligent Agent called Cyber Agent is proposed as a data miner to help users search and organize information. The agent attempts to dig out the information from the distributed environment, the World Wide Web, according to the special preferences of each individual. It contains two major subsystems, namely, WebTracer and WebOrganizer. It has the capability of performing tasks such as object decomposition, object representation, approximate string matching, and data clustering. These capabilities enable the agent to dig out and analyze an individual's preferences in surfing the web and then explore more related information for this individual.

Comparing to the progress accomplished in the area of digital circuits and systems, the analogue circuits fault diagnosis field is still in its infancy. Recently, some approaches to analog circuit's fault diagnosis have been proposed using pattern recognition capability of artificial neural networks. However, the major of these papers have analyzed linear analog circuits including resistors exclusively. In this paper, we present several neural network based approaches to analog circuits fault diagnosis using Back-Propagation, Learning Vector Quantization and Radial Basis Function neural models. The interest of our approaches is related to the fact that we use competitive multi-neural network architecture. Case study, simulation results and experimental validation of presented techniques have been reported.

The use of artificial neural networks (NNs) to address the channel assignment problem (CAP) for cellular time-division multiple access and code-division multiple access networks has previously been investigated by this author and many others. The investigations to date have been based on a hexagonal cell structure established by omnidirectional antennas at the base stations. No account was taken of the use of spatial isolation enabled by directional antennas to reduce interference between mobiles. Any reduction in interference translates into increased capacity and consequently alters the performance of the NNs. Previous studies have sought to improve the performance of Hopfield- Tank network algorithms and self-organizing feature map algorithms applied primarily to static channel assignment (SCA) for cellular networks that handle uniformly distributed, stationary traffic in each cell for a single type of service. The resulting algorithms minimize energy functions representing interference constraint and ad hoc conditions that promote convergence to optimal solutions. While the structures of the derived neural network algorithms (NNAs) offer the potential advantages of inherent parallelism and adaptability to changing system conditions, this potential has yet to be fulfilled the CAP for emerging mobile networks. The next-generation communication infrastructures must accommodate dynamic operating conditions. Macrocell topologies are being refined to microcells and picocells that can be dynamically sectored by adaptively controlled, directional antennas and programmable transceivers. These networks must support the time-varying demands for personal communication services (PCS) that simultaneously carry voice, data and video and, thus, require new dynamic channel assignment (DCA) algorithms. This paper examines the impact of dynamic cell sectoring and geometric conditioning on NNAs developed for SCA in omnicell networks with stationary traffic to improve the metrics of convergence rate and call blocking. Genetic algorithms (GAs) are also considered in PCS networks as a means to overcome the known weakness of Hopfield NNAs in determining global minima. The resulting GAs for DCA in PCS networks are compared to improved DCA algorithms based on Hopfield NNs for stationary cellular networks. Algorithm performance is compared on the basis of rate of convergence, blocking probability, analytic complexity, and parametric sensitivity to transient traffic demands and channel interference.

Popularized by the images from weather satellites and other Earth observing satellites, remote sensing from space has already become a household term. Airborne remote sensing, however, still holds its important place in the development of the remote sensing technology and in many applications. Prototype, proof-of-concept instruments are flown on aircraft before their improved versions are deployed on space shuttles or satellites. Airborne remote sensing is also more practical for regional applications. Since an aircraft flies in the Earth's atmosphere, factors contributing to geometric distortion are less systematic and more random. Substantial amount of effort is usually required to rectify the measurements. In this study, a scanner model is developed to generate simulated aircraft measurements. A backpropagation network and other variations are used to map the measurement space to the physical space. For measurements conducted over extensive area, techniques of anchoring the training data is developed such that geometric rectification can be performed in segments. Advantages of the neural network methods over the traditional method, and the need of constrained optimization are discussed.

In the 1980's, neurobiologist suggested a simple mechanism in primate visual cortex for maintaining a stable and invariant representation of a moving object. The receptive field of visual neurons has real-time transforms in response to motion, to maintain a stable representation. When the visual stimulus is changed due to motion, the geometric transform of the stimulus triggers a dual transform of the receptive field. This dual transform in the receptive fields compensates geometric variation in the stimulus. This process can be modelled using a Lie group method. The massive array of affine parameter sensing circuits will function as a smart sensor tightly coupled to the passive imaging sensor (retina). Neural geometric engine is a neuromorphic computing device simulating our Lie group model of spatial perception of primate's primal visual cortex. We have developed the computer simulation and experimented on realistic and synthetic image data, and performed a preliminary research of using analog VLSI technology for implementation of the neural geometric engine. We have benchmark tested on DMA's terrain data with their result and have built an analog integrated circuit to verify the computational structure of the engine. When fully implemented on ANALOG VLSI chip, we will be able to accurately reconstruct a 3D terrain surface in real-time from stereoscopic imagery.

The objective of this research is to automate the classification of clouds from satellite images providing a method for studying their properties over time. Analysis was applied to the International Satellite Cloud Climatology Project (ISCCP) low resolution (2.5 degrees per pixel) database for January 1987. Our approach differs from earlier studies by taking advantage of cloud top pressure and optical thickness from the ISCCP database, providing more accurate measures of cloud height with less dependency on the sun's angle of illumination. A total of 365 regions of interest (ROI), each classified Storm or Non Storm were used in the analysis. The algorithms used were Backpropagation Artificial Neural Network and Nearest Neighbor Pattern Classification. Each ROI was assigned on identification number between 1 and 365. One third of the ROIs were randomly selected for testing using a random number generator and the remaining ROIs were assigned to be training set. This process was repeated 29 times resulting in a mean classification error of 5.76% for the nearest neighbor algorithm and 3.97% for the backpropagation neural network.

The research presented in this paper describes a comparison of the use of Multi-Layer Perceptron Neural Network and the Radial Basis Functions Neural Networks in predicting the physical structure of the ocean up to 20 km ahead of a sea going vessel. The data used in this experiment was acquired on board a vessel transecting the Atlantic Ocean from 50 degree(s) South (the Falkland Island) to 50 degree(s) North (the UK). The research is part of the Strategic Tactical Environmental Bubble project which aims to develop a methodology for predicting the physical structures in three dimensions, around a sea going vessel from data acquired in situ and historical records. A Sobel filter has been applied to pre- process the raw data, enhancing features and smoothing stable areas. Some characteristics of both types of Neural Networks are outlined, as well as an exhaustive comparison of both methods. The advantages and disadvantages of both to this particular problem are discussed. A final system is presented together with a critical evaluation of its performance. For this purpose a real-time simulator has been built and its results are presented here.

The NASA Scatterometer, launched in August 1996, is designed to measure wind vectors over ice-free oceans. To prevent contamination of the wind measurements, by the presence of sea ice, algorithms based on neural network technology have been developed to classify ice-free ocean surfaces. Neural networks trained using polarized alone and polarized plus multi-azimuth `look' Ku-band backscatter are described. Algorithm skill in locating the sea ice edge around Antarctica is experimentally evaluated using backscatter data from the Seasat-A Satellite Scatterometer that operated in 1978. Comparisons between the algorithms demonstrate a slight advantage of combined polarization and multi-look over using co-polarized backscatter alone. Classification skill is evaluated by comparisons with surface truth (sea ice maps), subjective ice classification, and independent over lapping scatterometer measurements (consecutive revolutions).

The problem of Direction of Arrival (DOA) estimation of users in mobile communication systems using linear antenna arrays is addressed. Superresolution algorithms, such as Multiple Signal Classification (MUSIC), are used to locate desired as well as cochannel mobile users. However these algorithms require extensive computation and are difficult to implement in real-time. In this paper, the DOA problem is approached as a mapping problem which can be modeled using a suitable artificial neural network trained with input output pairs. A study of a three-layer Radial Basis Function Neural Network (RBFNN) which can learn multiple source direction finding with a six-element array is conducted. RBFNNs were used due to their ability for data interpolation in higher dimensions. The network weights are modified using the normalized cumulative delta rule. The performance of this network is compared to that of the MUSIC algorithm for both uncorrelated and corrected signals. It was found that networks implementing these functions were indeed successful in performing the required task and their performance approached that of the MUSIC algorithm. It is also shown that the RBFNN substantially reduced the CPU time for the DOA estimation computations.

Auditory perception neurons, also called inner hair cells (IHCs), because of their physical shape, transform the mechanical movements of the basilar membrane into electrical impulses. The impulse coding of the IHC is the main information carrier in the auditory process and is the basis for improvements of cochlea implants as well as for low rate, high quality speech processing and compression. This paper compares biologically motivated models (MEDDIS, COOKE, BRACHMAN-PAYTON) with a newly developed model which is transfer function oriented. The new model has only three reservoirs and the parameters can be controlled through five small ROM tables. We compare this model with the often used MEDDIS model in terms of accuracy, system parameter flexibility, and hardware effort in an FPGA implementation.

Geophysical Model Functions (GMF) describing the relationship between the scatterometer normalized radar cross section (sigma-0) and useful geophysical parameters such as sea-surface wind vectors, wave heights, and sea- surface temperatures have been undergoing extensive research and development during the last decade. In this study, we investigate the use of two feed-forward neural networks, Multilayer Perceptron and Radial Basis Functions, for developing a useful and accurate representation of the C- band GMF. Collocated radar sigma-0 cells with global wind vector models were used as the database of the study. The resulting well-known biharmonic relationship between the sigma-0 and the relative azimuth angle between the scatterometer antenna beam azimuth and wind direction shows the excellent agreement between the neural network and previous results. The applicability of the neural techniques in this application are clearly presented and the potential for possible enhancement over previous approaches are discussed.

Many techniques have been developed in recent years in order to speed up the training process of Multilayer Perceptrons (MLPs), such as accelerated versions of gradient descent, second order methods, Kalman related algorithms, and block learning methods; among them, the latter offer a reasonable way of reducing training effort and getting good results with MLPs. However, they usually find suboptimal solutions and, sometimes, the values obtained for the weights are excessively large. In this paper, we analyze the drawbacks of these block methods using a very simple test model, we propose several modifications by discarding samples (DS- LSB), using controlled perturbation (CP-LSB), and by means of a Reduced Sensitivity version (RS-LSB) and we also extend their forms to include some especial characteristics: we propose a robust training algorithm relying on Total Least Squares (TLS) minimizations (the so-called RS-TLS algorithm), useful for noisy training patterns, and we also present a method for training MLPs with general output cost function (the `RS-WLSB' algorithm). The advantages of these methods with respect to related algorithms are illustrated using several test problems. Finally, we extract some conclusions and propose, as a further work, the development of recursive implementations, able to learn on-line and to deal with non-stationary problems, and the application of block training methods to recurrent networks.

In this paper we introduce a new technique to build a Gaussian mixture classifier. It is based on the selection of the number and location of nodes dedicated to every class by means of discriminative rules. This feature allows us to make a fair comparison with MLP networks for terrain classification in remote sense applications, a field where non-parametric techniques usually outperform classical ML Gaussian classifiers. The main characteristic of the architecture proposed is the ability to select the proper number of Gaussian nodes per class attending to discriminative rules. The growth control is imposed by the use of an information theoretic criterion that prevents the network from becoming extremely complex, thus loosing generalization capabilities. After the growing phase is finished, a mutual information criterion is maximized to bias the parameters to a more discriminative configuration. We report a comparative study on terrain classification over a Landsat-TM image, using this technique and MLP classifiers with one hidden layer.

We establish robustness stability results for a large class of artificial neural networks for associative memories under parameter perturbations and determine conditions that ensure the existence of asymptotically stable equilibria of the perturbed neural system that are near the asymptotically stable equilibria of the original unperturbed neural network. The proposed stability analysis tool is the sliding mode control and it facilitates the analysis by considering only a reduced-order system instead of the original one and time-dependent external stimuli.

In this paper, a new control strategy is presented that combines sliding mode control theory with a neural network. Sliding mode control theory requires the complete knowledge of the dynamics of the controlled system. However, in practice, this could be a serious limitation on the practical usefulness of sliding mode control theory. A multilayer neural network with a back-propagation learning algorithm is employed to solve this kind of problem. The neural network serves as a compensator without a priori knowledge about the system. The robustness against parameter uncertainty and nonlinearity, and the effectiveness of the proposed algorithm is verified by simulation results.

This paper summarizes theoretical findings and applications of artificial neural networks to modeling of complex engineered system response in the abnormal environments. The thermal fire impact on the industrial container for waste and fissile materials was investigated using model and experimental data. Solutions for the direct problem show that the generalization properties of neural network based model are significantly better than those for standard interpolation methods. Minimal amount of data required for good prediction of system response is estimated in computer experiments with MLP network. It was shown that Kohonen's self-organizing map with counterpropagation may also estimate local accuracy of regularized solution for inverse and combined problems. Feature space regions of partial correctness of the inverse model can be automatically extracted using adaptive clustering. Practical findings include time strategy recommendations for fire-safe services when industrial or transport accidents occur.

Based on neural net and Hopfield neural network, this paper introduces the method of direction of arrival estimation, which is based on neural network and shows the simulated method and results. It gives a new and effective way for the realization of real time accurate location of the target.

The radial basis function network (RBFN) is analyzed and the fuzzy radial basis function network (FRBFN) which is more suitable for the radar target recognition is proposed in this paper. Here both of the two networks are used as classifiers. This FRBFN utilize fuzzy clustering method to determine the structure of the net. The generalization property of the two networks are discussed. It is shown from the theoretical analysis and experiment that the FRBFN has better generalization property. The Doppler echoes of the targets gotten from a current surveillance radar are used in the experiment. The experimental results shows that the classification rate of the FRBFN is higher than that of the RBFN. The network proposed in this paper is promising in the application of radar target recognition.

In this paper we develop techniques that are suitable for the parallel implementation of Fuzzy ARTMAP networks. Speedup and learning performance results are provided for execution on a DECmpp/Sx-1208 parallel processor consisting of a DEC RISC Workstation Front-End and MasPar MP-1 Back-End with 8,192 processors. Experiments of the parallel implementation were conducted on the Letters benchmark database developed by Frey and Slate. The results indicate a speedup on the order of 1000-fold which allows combined training and testing time of under four minutes.

The main purpose of this paper is to investigate theoretically and experimentally the use of family of Polynomial Powers of the Sigmoid (PPS) Function Networks applied in speech signal representation and function approximation. This paper carries out practical investigations in terms of approximation fitness (LSE), time consuming (CPU Time), computational complexity (FLOP) and representation power (Number of Activation Function) for different PPS activation functions. We expected that different activation functions can provide performance variations and further investigations will guide us towards a class of mappings associating the best activation function to solve a class of problems under certain criteria.

Non-equilibrium dynamics of ensemble consisting of interacting Josephson-like elements with lateral excitatory and inhibitory couplings is studied analytically and by computer simulation. For this model an extremely rich and unusual dynamic behavior (long-term and short-term metastable states, modulated limit cycles and strange attractors) is observed. Phase synchronization of activity patterns in ensembles with different types of couplings between oscillators are found.

In this paper, we will discuss two applications of the Boltzmann machines. We will first briefly review the theory of the Boltzmann machine; then, we will report the results of using a software simulation of the Boltzmann machine. The software can learn 1,000 characters (19 X 19) in less than 10 seconds, and learn 4,000 characters (19 X 19) in less than one minute; it can also recognize one of the 4,000 characters (19 X 19) in less than 0.5 seconds. We will also report our results of image retrieval.

This paper presents a new method with Neural Network on the branch matching in the recognition process of Chinese Character described by Attributed Relational Graphs. The experimental result is shown in the end of the paper.

An algorithm for the circle recognition using deformable templates was carried out and its performances were studied. The developed technique was explored in a number of different situations ranging from sets of simulated circles to real data obtained in a lead-gold interactions from the RICH detectors used in CERES/NA45 experiment. A special study was devoted to track finding problem for a vertex detector. Our technique with reasonable modification allows to recognize tracks and estimate their parameters on high multiplicity and background conditions. Results show the satisfactory robustness of our algorithm to background contaminations. It can be used in many data handling problems that appear in high energy physics like Cherenkov ring recognition and track reconstruction.

Hispars is an European EUCLID investigation projected devoted to evaluation of Artificial Neural Networks for defense pattern applications. Three demonstrators representing three military operational contexts (Air-to- Ground, Ground Battlefield, Naval Threat Evaluation) have been defined and developed. A set of operational processing chains have been selected, and for each of them, ANN methods have been proposed and evaluated on real data set at each level of processing, in comparison to those classical techniques used in existing equipment.

This paper demonstrates the usefulness of neural networks in classifying environmental samples from compound mixture data. This problem was solved by careful determination of neural network learning parameters and forward sequential selection of input features. Finally, the fundamental limit of any classifier on this data was determined using Bayes error bounding.

This paper summarizes the work that was done to explore the use a system based approach to generate test vectors for electronic systems. The process that was developed takes advantage of an unsupervised neural network algorithm Adaptive Resonance Theory (ART) and a Genetic Algorithm (GA) that is combined to form an optimal control system. The GA generates a population of test patterns (individuals). Each individual is provided as timed inputs to a set of behavior based simulations representing good and faulty circuits. The response of each model is recombined in the form of an image matrix with each row representing a signature of each of the different circuits. FuzzyART provides a method of image recognition, extracting those images that are distinctly different from any other. Each individual generated by the GA is evaluated and a fitness is provided by FuzzyART by the number of neuron clusters formed. New test sequences evolve with increasing fault isolation and detection. The process is repeated until a maximum number of models have been identified and separated. A selective breading algorithm was included to reduce the need for large populations, thus increasing the speed to converge to the `best test.'

Based on linear plan neural network, this paper provides the method of detecting space direction by forward/backward linear prediction. The first step of this method is to turn the solution in complex field of forward/backward linear prediction coefficients to the solution in real field. Then we come to the realization by use of linear plan neural network which has been supplied by Tank and Hopfield. Because this network can come to a dynamic balance within several hundreds ns, it provides a new way for the realization of real time detecting direction in forward/backward linear forecast space. It is demonstrated by computer simulation that this method is better than traditional Burg algorithm, so it is practicable.

Human visual system can perceive 3D illusory object from the visual stimuli of disparity given by some inducing objects arranged with suitable relations. In this paper, we will introduce the visual effect in two kinds of arrangement with occlusion and sustaining relationship between the illusory object and inducing objects. In the former case, we name the inducing objects which provide the stimuli as occlusion cues, and classify them in two types: contour occlusion cues and bulky occlusion cues. In the later case, those inducing objects are named sustaining cues. A 3D transparent illusory object is perceived, just like we can imagine the scenes from the actions and positions of the pantomimists, so we call this visual effect as Pantomime Effect. According to the positions of sustaining cues, we found that sustaining cues played different actions in pantomime effect perception, and classified them into three types: front sustaining cues, side sustaining cues and back sustaining cues. It is also found that this effect was closely related to occlusion cues; especially pantomime effect by front sustaining cues can be regarded as a kind of bulky occlusion in opaque object perception.

Application of the neural networks methods to problems in astrophysics has gained popularity in recent years. We discuss here the application of standard multilayered feedforward neural networks to classification problems in balloon-borne cosmic ray experiments.

The primordial antimatter search in primary cosmic radiation is a very difficult task. The main reasons are: the low signal to background ratio expected and the limited geometrical factor available onboard of space facilities. Moreover the data transmission system to send the information from space to the ground control room is also a critical factor. As a consequence, an high performance onboard recognition system, based on a multi-level decision making strategy has to be taken into account. The study of a L2 trigger based on neural networks classifiers for a cosmic ray space experiment on a polar orbit satellite is addressed.

Identification of star constellations with an onboard star tracker provides the highest precision of all attitude determination techniques for spacecraft. A method for identification of star constellations inspired by neural network (NNW) techniques is presented. It compares feature vectors derived from histograms of distances to multiple stars around the unknown star. The NNW method appears most robust with respect to position noise and would require a smaller database than conventional methods, especially for small fields of view. The neural network method is quite slow when performed on a sequential (serial) processor, but would provide very high speed if implemented in special hardware. Such hardware solutions could also yield lower low weight and low power consumption, both important features for small satellites.

Adaptive Resonance Theory provides a model of pattern classification that addresses the plasticity--stability dilemma and allows a neural network to detect when to construct a new category without the assistance of a supervisor. We show that Adaptive Resonance Theory can be applied to the study of natural concept development. Specifically, a model is presented which is able to categorize different usages of a common noun and group the polysemous senses appropriately.

In this paper we discuss a first parallel implementation of our new compression algorithm for lossy and lossless image compression. The compression algorithm is based on a new method of `dynamic quantization' of the coefficients of the wavelet transform applied on the data. The parallel implementation of the algorithm is on the Italian QUADRICS parallel supercomputer of the Alenia-Space Co. This implementation follows the principles of our `dynamic perceptron' neural algorithm we presented elsewhere. The results of the comparison between the performances of our compression prototype and the usual algorithms for image compression will also be discussed. The tests were made on standard SAR images of the European Space Agency. The results suggest this parallel implementation is able to perform a real time efficient compression of SAR image (15 Mb/sec).

This paper illustrates and discusses the relative merits of three methods--k-fold Cross Validation, Error Bounds, and Incremental Halting Test--to estimate the accuracy of a supervised learning algorithm. For each of the three methods we point out the problem they address, some of the important assumptions that they are based on, and illustrate them through an example. Finally, we discuss the relative advantages and disadvantages of each method.