Tuning rules for adaptive neural networks have featured Lyapunov-based approaches in recent years. Although these have some desirable qualities they have led to complex tuning procedures. In order to take more advantage of the power of adaptive neural networks less complex and computationally expensive tuning rules are desirable. In addition, tuning rules should be simple and provide for rapid, reliable convergence. In this paper a proportional-integral approach to adaptive neural network tuning rules is studied. Simulation on a nonlinear system is used for a demonstration.

Neural networks (NNs) represent a class of systems that do not fit into the current paradigms of software development and certification. Instead of being programmed, a learning algorithm "teaches" a NN using a set of data. Often, because of the non-deterministic result of the adaptation, the NN is considered a "black box" and its response may not be predictable. Testing the NN with similar data as that used in the training set is one of the few methods used to verify that the network has adequately learned the input domain. In most instances, such traditional testing techniques prove adequate for the acceptance of a neural network system. However, in more complex, safety- and mission-critical systems, the standard NN training-testing approach is insufficient to provide a reliable method for their certification. Verifying correct operation of NNs within NASA projects, such as autonomous mission control agents and adaptive flight controllers, and within nuclear engineering applications, such as safety assessors and reactor controllers, requires as rigorous an approach as those applied to common programming techniques. This verification and validation (V&V) challenge is further compounded by adaptive neural network systems; ones that modify themselves, or "learn," during operation. These systems continue to evolve during operation, for better or for worse. Traditional software assurance methods fail to account for systems that change after deployment. Several experimental NN V&V approaches are beginning to emerge, but no single approach has established itself as a dominant technique. This paper describes several of these current trends and assesses their compatibility with traditional V&V techniques.

Conventional model-based or statistical analysis methods for functional MRI (fMRI) are easy to implement, and are effective in analyzing data with simple paradigms. However, they are not applicable in situations in which patterns of neural response are complicated and when fMRI response is unknown. In this paper the "neural gas" network is adapted and rigorously studied for analyzing fMRI data. The algorithm supports spatial connectivity aiding in the identification of activation sites in functional brain imaging. A comparison of this new method with Kohonen's self-organizing map and with a minimal free energy vector quantizer is done in a systematic fMRI study showing comparative quantitative evaluations. The most important findings in this paper are: (1) the "neural gas" network outperforms the other two methods in terms of detecting small activation areas, and (2) computed reference function several that the "neural gas" network outperforms the other two methods. The applicability of the new algorithm is demonstrated on experimental data.

This paper describes our current work in developing a vision-based tracking and trajectory prediction system for an aerial robot based on low-cost digital cameras, image processing techniques, and a filtering and prediction algorithm. The system determines the pose (location and orientation) of a miniature airship, online during indoor flight, and will be used in a development framework for a future autonomous flight control system. Object localization is achieved by tracking an infra-red target array mounted to a model airship. Its pose in three-dimensional space can be computed from corresponding points in the images of two cameras which are calibrated in a global coordinate system. The calibration procedure and the localization, as well as some aspects of the measurement accuracy achieved, are discussed. Real-world applications provide an uncertain static or dynamic environment which complicates the tracking of a target. To overcome problems due to noisy data or even failed target detection in image frames, a filtering procedure is applied for estimating the airship's pose. In a first step, points in the two-dimensional image planes are directly tracked and propagated forward to the vehicle pose. In a second step, an adaptive noise Kalman filter is applied for estimating and predicting the flight trajectory. Its state is propagated back to points in the image planes to guide the detection algorithm by defining regions of confidence. Both approaches are combined in a tracking algorithm. In-flight measurements are used to validate the parameters of the adaption procedure. Some experimental results are shown.

Due to the increased threats of chemical and biological weapons of mass destruction (WMD) by international terrorist organizations, a significant effort is underway to develop tools that can be used to detect and effectively combat biochemical warfare. Furthermore, recent events have highlighted awareness that chemical and biological agents (CBAs) may become the preferred, cheap alternative WMD, because these agents can effectively attack large populations while leaving infrastructures intact. Despite the availability of numerous sensing devices, intelligent hybrid sensors that can detect and degrade CBAs are virtually nonexistent. This paper reports the integration of multi-array sensors with Support Vector Machines (SVMs) for the detection of organophosphates nerve agents using parathion and dichlorvos as model stimulants compounds. SVMs were used for the design and evaluation of new and more accurate data extraction, preprocessing and classification. Experimental results for the paradigms developed using Structural Risk Minimization, show a significant increase in classification accuracy when compared to the existing AromaScan baseline system. Specifically, the results of this research has demonstrated that, for the Parathion versus Dichlorvos pair, when compared to the AromaScan baseline system: (1) a 23% improvement in the overall ROC Az index using the S2000 kernel, with similar improvements with the Gaussian and polynomial (of degree 2) kernels, (2) a significant 173% improvement in specificity with the S2000 kernel. This means that the number of false negative errors were reduced by 173%, while making no false positive errors, when compared to the AromaScan base line performance. (3) The Gaussian and polynomial kernels demonstrated similar specificity at 100% sensitivity. All SVM classifiers provided essentially perfect classification performance for the Dichlorvos versus Trichlorfon pair. For the most difficult classification task, the Parathion versus Paraoxon pair, the following results were achieved (using the three SVM kernels: (1) ROC Az indices from approximately 93% to greater than 99%, (2) partial Az values from ≈79% to 93%, (3) specificities from 76% to ≈84% at 100 and 98% sensitivity, and (4) PPVs from 73% to ≈84% at 100% and 98% sensitivities. These are excellent results, considering only one atom differentiates these nerve agents.

In this paper, we propose an Avatar-based intelligent agent technique for 3D Web based Virtual Communities based on distributed artificial intelligence, intelligent agent techniques, and databases and knowledge bases in a digital library. One of the goals of this joint NSF (IIS-9980130) and ACM SIGGRAPH Education Committee (ASEC) project is to create a virtual community of educators and students who have a common interest in comptuer graphics, visualization, and interactive techniqeus. In this virtual community (ASEC World) Avatars will represent the educators, students, and other visitors to the world. Intelligent agents represented as specially dressed Avatars will be available to assist the visitors to ASEC World. The basic Web client-server architecture of the intelligent knowledge-based avatars is given. Importantly, the intelligent Web agent software system for the 3D virtual community is implemented successfully.

In this paper, a smart agent is proposed to search image/graphics based database for students and an instructor. The seach procedure is based on content-based image retrieval (CBIR). In addition, the smart agent can also give help in seaching image/graphics data, answering simple questions and optimizing educational activities graphically.

The heterogeneity and the lack of structure that permeates much of the ever expanding information sources on the WWW makes it difficult for the user to properly and efficiently access different web pages. Different users have different needs from the same web page. It is necessary to train the system to understand the needs and demands of the users. In other words there is a need for efficient and proper web mining. In this paper issues and possible ways of training the system and providing high level of organization for semi structured data available on the web is discussed. Web pages can be evolved based on history of query searches, browsing, links traversed and observation of the user behavior like book marking and time spent on viewing. Fuzzy clustering techniques help in grouping natural users and groups, neural networks, association rules and web traversals patterns help in efficient sequential anaysis based on previous searches and queries by the user. In this paper we analyze web server logs using above mentioned techniques to know more about user interactions. Analyzing these web server logs help to closely understand the user behavior and his/her web access pattern.

The new generation of instruments to measure the atmosphere gives a large amount of information whose content is not immediately obvious. For this reason the information content (in the Shannon sense) and the number of degrees of freedom for signal related to parameters to retrieve, are crucial to carry out, for instance, the selection of microwindows or subset of channels for retrieval. This seems to be even true for the miniaturized devices like those obtained by MEOMS technologies that have high flexibility because are highly integrated. In this paper we present the methodologies to obtain the information content related to wavelengths selection from satellite with limb and nadir view with the MEOMS technology in mind.

To mitigate the problem of intrusion attacks by malicious nodes in mobile ad hoc networks (MANETs), security attributes and quantifiable trust levels, unique to the MANET's transient, self-organizing topology, augment or replace traditional protocol metrics of throughput, packet delay and hop-count in the ad hoc route discovery procedures. The new features are unique to the candidate security protocol, which views security as a quality metric to improve the relevance of the routes discovered by established reactive ad hoc routing protocols. Attributes of a secure route are identified in order to define the appropriate metrics to quantify the "level of security" associated with the protocol messaging and the detection of malicious activities by some intrusive nodes. A state vector of features and metrics based on the published Secure Routing Protocol (SRP) for MANETs is constructed to encode network security characteristics. This route discovery protocol mitigates the detrimental effects of various malicious behaviors to provide trustworthy connectivity information. The protocol ensures that fabricated, compromised, or replayed route replies would either be rejected or never reach the querying source node. In this paper, the pattern of values, taken by the state vector of the SRP features in the route request, discovery and reply operations, are analyzed to detect evidence of intrusion attacks by malicious nodes that could lead to denial of service and network shutdown. The pattern analysis applies a technique based on negative selection found in natural immune systems that can detect extraneous patterns in the (nonself) space that is the complement of vector values associated with correct route discovery and route maintenance. The immune system is well-suited to the distributed nature of the MANET. It does not rely on a central controller, but instead uses a distributed detection and response mechanism in order to respond to foreign invaders, mirroring the operation of the route discovery and selection process in the presence of intrusive or malicious nodes. Furthermore, this pattern detection approach is suitable for the difficult problem of passive or hidden security threats. Based on the SRP features of the state vector, an artificial immune system (AIS) is constructed as a hierarchy of rules to detect different types of intrusive activity within the MANET. The pattern detection rules in the complement (nonself) space are generated in an evolutionary manner using a genetic search algorithm. The effect of the genetic search is to discern the varying levels of abnormal behavior in the MANET protocol transactions. The efficacy of the AIS technique is compared to the positive characterization method based on nearest-neighbor classification. Initial evaluations of the detection scheme are performed to validate the AIS-based method using training and test data sets, generated from intrusion scenarios simulated from various threat models and security-aware modifications to reactive MANET routing protocols. These results are reported along with a performance analysis comparing the AIS approach with competing techniques. Conclusions about the AIS application to MANETs using the SRP are discussed.

In this paper, we present an evolution of the classical Barrow-Wheeler Transform (BWT) algorithm applied to the sorting procedure of interferometric data, in view of their lossless compression: the Dynamic Segmentation and Sorting (DSS) algorithm. This algorithm is based on the application of the Dynamic Perceptron (DP) neural algorithm. It allows a fast and computationally efficient dynamic segmentation into homogeneous blocks of the coefficients of the Fourier transform. In this way, a limited order sorting procedure of such coefficients can be allowed, optimizing the BWT sorting procedure, characterized as such by an unlimited order. This new method has been specifically studied in order to its hardware implementation for integrating, as a fast compression module, the Micro-Electro-Optical-Mach Zender-Sensor (MEOMS) -- a integrated optical micro-sensor of the MEOS class, based on an array of Mach-Zender type interferometers. This device has been recently designed and constructed in Italy, at the IMM and ISAC Institutes of the Italian National Research Council (CNR) in Bologna. Because of its characteristics, the complete system is suitable for being installed onboard on satellites, for the continuous monitoring of the earth atmosphere. Some encouraging previous results of our lossless compression module of the interferograms outputted by the MEOMS are presented at the end of this paper.

A challenge for intelligent computing is translating the skills of innovation into mathematical theory and persistent learning algorithms. Computational intelligence differs from artificial intelligence in that artificial intelligence reasons over symbols while computational intelligence reasons over sub-symbolic data and information. Natural symbos arise from shared human experiences. The creative quality of human interaction suggests symbol generation involves a collection of cooperative agents capable of representing relative experience, negotiating innovation, and---finally---building consensus. As hybrids of sub-symbolic and symbolic reasoning become the norm, it is necessary to formalize the design and evaluation of artificial symbols. In this paper, we delineate the difference between sub-symbolic patterns and symbolic experience. Further, we propose fundamental theory supporting the autonomous construction of artificial symbols which---we assert---is the ultimate culmination of an intelligent computation. We apply this theory to model selection among neural networks.

Given a classifier trained on two-class data one wishes to determine how well the classifier will perform on new, unseen data. To do this task one typically uses the data to estimate a distribution of the data, generate new data from this distribution, and then test the data. Also, hold-out methods are used including cross-validation. We propose a new method that uses computational geometry techniques that produces a partial ordering on subsets in feature space and measures how well the classifier will perform on these subsets. There are some conditions on the classifier that must be satisfied in order that this measure, in fact, exists. We give the details for these conditions as well as the results concerning this special collection of classifiers. We derive the measure that quantifies the generalization capability for the special collection classifier.

The applications of non-standard logic device are increasing fast in the industry. Many of these applications require high speed, low power, functionality and flexibility, which cannot be obtained by standard logic device. These special logic cells can be constructed by the topology design strategy automatically or manually. However, the need arises for the topology design verification. The layout versus schematic (LVS) analysis is an essential part of topology design verification, and subcircuit extraction is one of the operations in the LVS testing. In this paper, we first provided an efficient decision tree approach to the graph isomorphism problem, and then effectively applied it to the subcircuit extraction problem based on the solution to the graph isomorphism problem. To evaluate its performance, we compare it with the neural networks based heuristic dynamic programming algorithm (SubHDP) which is by far one of the fastest algorithms for subcircuit extraction problem.

At its substratum, brain/mind organization requires both synaptic firings and non-synaptic events. Synaptic firings organize the pattern of non-synaptic events. Non-synaptic events organize the pattern of synaptic firings. The processes are related in a bizarre hierarchy. Comparing these processes to electric circuits, it is as if we have two circuits that each continuously and simultaneously update the topology, and consequently, the dynamical laws of the other. Since either can be seen to be rebuilding the other, from its own perspective each process appears higher than the other in a hierarchy. This same kind of hierarchy is found in a hyperset structure. Interpreted as a directed graph, the nodes in a hyperset form a hierarchy in which, from the perspective of any node in the hierarchy, that node is at the top. This organizational structure violates the Foundation Axiom. Algorithmic computation strictly complies with the Foundation Axiom. Thus, an algorithm organized like a hyperset is a contradiction in terms. Does this contradiction mean are we precluded forever from implementing brain-like activities artificially? Not at all! An algorithm is incapable of doing the job, but nothing prevents us from constructing interacting analog processes that update each other's dynamical laws on the fly.

Techniques based on neural networks can provide efficient solutions to a wide variety of problems in computer science. Routing in computer networks is to schedule messages and select communication links so that messages can be transferred efficiently between source and destination processors. Finding an optimal solution to many routing problems usually reqrueis exponential time and is impractical in reality. Hence, many heuristic algorithms have been designed to find sub-optimal solutions. In this research we use neural networks with a set of constraints to capture various collisions in multistage interconnection networks (MINs). Our simulation results have indicated that the Hopfield neural network can be used to routing to avoid link collisions in electronic MINs and crosstalks in optical MINs.

Image registration is formulated as a nonlinear optimization problem of finding an affine transformation minimizing the difference between images. A particular scheme of the hybrid evolutionary algorithm is used to solve the problem. The reproduction phase of the algorithm is enhanced with a two-phase operation of local search and correction performed on the subset of the best chromosomes in the reproduction pool. In order to reduce the computational cost of the correction, a mechanism of the adaptive control of the local search is designed, based on a micro model of the local image response. The mechanism correlates the step size of the search with the local properties of the gray level surface at different points of an image. To reduce the number of evaluations of the local image response required by the control mechanism, all participating image points are evaluated and classified at once in the preprocessing stage. A self-organizing neural network is employed to classify different points according to their response values, and to build an adaptive, compact response map of the entire image. During the execution of the main algorithm, this map is used as a lookup table, to retrieve the appropriate response values for the points participating in the local search.

Many-class learning is the problem of training a classifier to discriminate among a large number of target classes. Together with the problem of dealing with high-dimensional patterns (i.e. a high-dimensional input space), the many class problem (i.e. a high-dimensional output space) is a major obstacle to be faced when scaling-up classifier systems and algorithms from small pilot applications to large full-scale applications. The Autonomous Recursive Task Decomposition (ARTD) algorithm is here proposed as a solution to the problem of many-class learning. Example applications of ARTD to neural classifier training are also presented. In these examples, improvements in training time are shown to range from 4-fold to more than 30-fold in pattern classification tasks of both static and dynamic character.

In this paper, a new parallel hardware architecture dedicated to compute the Gaussian Potential Function is proposed. This function is commonly utilized in neural radial basis classifiers for pattern recognition as described by Lee; Girosi and Poggio; and Musavi et al. Attention to a simplified Gaussian Potential Function which processes uncorrelated features is confined. Operations of most interest included by the Gaussian potential function are the exponential and the square function. Our hardware computes the exponential function and its exponent at the same time. The contributions of all features to the exponent are computed in parallel. This parallelism reduces computational delay in the output function. The duration does not depend on the number of features processed. Software and hardware case studies are presented to evaluate the new CORDIC.

Most animals use color and shape to recognize things in their worlds. Shapes are properties of objects inferable from their 2D images. Colors are spectral discriminants computed by their brains using data sensed in 2-4 broad, spectrally overlapping bands. So color is not an object property. Both shape and color combine to make discrimination work well. I will describe Artificial Color and then ways to recognize an Artificially colored image.

Vision is a part of a larger information system that converts visual information into knowledge structures. These structures drive vision process, resolve ambiguity and uncertainty via feedback projections, and provide image understanding that is an interpretation of visual information in terms of such knowledge models. The ability of human brain to emulate knowledge structures in the form of networks-symbolic models is found. And that means an important shift of paradigm in our knowledge about brain from neural networks to "cortical software". Symbols, predicates and grammars naturally emerge in such active multilevel hierarchical networks, and logic is simply a way of restructuring such models. Brain analyzes an image as a graph-type decision structure created via multilevel hierarchical compression of visual information. Mid-level vision processes like clustering, perceptual grouping, separation of figure from ground, are special kinds of graph/network transformations. They convert low-level image structure into the set of more abstract ones, which represent objects and visual scene, making them easy for analysis by higher-level knowledge structures. Higher-level vision phenomena are results of such analysis. Composition of network-symbolic models works similar to frames and agents, combines learning, classification, analogy together with higher-level model-based reasoning into a single framework. Such models do not require supercomputers. Based on such principles, and using methods of Computational intelligence, an Image Understanding system can convert images into the network-symbolic knowledge models, and effectively resolve uncertainty and ambiguity, providing unifying representation for perception and cognition. That allows creating new intelligent computer vision systems for robotic and defense industries.

This paper is concerned with the detection of physical flaws on pipe walls in gas pipelines. The sensor technology is EMAT, a non-contact ultrasonic technology. One EMAT is used as a transmitter, exciting an ultrasonic impulse into the pipe wall. Another EMAT located a few inches away from the first is used as a receiving transducer. This paper reports on the identification of flaw signatures in the receiver output. The first step in flaw characterization is to perform wavelet analysis of the signature. Being non-shift-invariant, an array of coefficients of a discrete wavelet transfor of a signal is not directly suitable as a pattern recognition feature. However, comparing composite properties of the signal on different scales is useful, because the more conversion caused by a flaw, changes the composite properties of the signal in wavelet space. For EMAT data, the useful information projects onto five mutually orthogonal wavelet scales. This paper reports onteh use of a robust 17-dimensional feature vector that mutually orthogonal wavelet scales. This paper reports on the use of a robust 17-dimensional featuer vector that consistently distinguishes "flaw" signatures from "no-flaw" signatures in a substantial collection of experimental data.

The column optical thickness of the atmospheric aerosol load is routinely retrieved using space borne measurements of the natural levels of upwelling shortwave radiation at the top of the atmosphere. The information content of the measurements includes contributions from several components of the earth-atmosphere system: the scattering and absorption processes of the molecular atmosphere (i.e., Rayleigh scattering and gas absorption), the scattering and absorption of aerosol particles, as well as the reflection from clouds and the underlying surface. The relative contribution from each component is a strong function of the wavelength of the incident radiation. Generally, the aerosol contribution is a very small fraction of the measurement. Therefore, as a part of the retrieval process, corrections for non-aerosol effects need to be applied to isolate the net aerosol effect.

In order to retrieve accurate information about surface physical and biogeochemical properties from space-borne sensors removal of the atmospheric signal is required. On the other hand, the retrieval of aerosol properties from space is hampered by the necessity to remove the signal originating from the underlying surface. Instead of retrieving surface (aerosol) properties in a two-step procedure in which an "atmospheric (surface) correction" is followed by retrieval of the desired quantity, it may be advantageous to attempt a simultaneous retrieval of atmospheric and surface parameters. The forward and inverse modeling strategies associated with this approach are discussed. This approach is particularly useful over bright surfaces, and can be used for simultaneous retrieval of snow and aerosol properties.