(Return to ICNN97 Homepage) (Return to ICNN'97 Agenda)
ICNN'97 FINAL ABSTRACTS
SPECIAL SESSIONS
Adaptive Critic Designs .......... Biomedical Applications of Neural Networks .......... Intelligent Control Theory and Applications .......... Knowledge-based Methods in Neural Networks .......... Linguistic Rule Extraction .......... Neural Networks Applications for Monitoring of Complex Systems .......... Neuro-Fuzzy Integration .......... Sensors and Biosensors .......... Visual System Models & Prostheses
ADAPTIVE CRITIC DESIGNS (Return to Top)
ICNN97 Special Session on Adaptive Critic Designs Session: SS1A Paper Number: 709 Oral
Primitive adaptive critics
Danil Prokhorov and Lee A. Feldkamp
Keywords: adaptive critics recurrent neural networks feedback controllers
Abstract:
We propose a simple framework for critic-based training of recurrent neural networks and feedback controllers. We term the critics that are used primitive adaptive critics, since we represent them with the simplest possible architecture (bias weight only). We derive this framework from two main premises. The first of these is a natural similarity between a form of approximate dynamic programming, called Dual Heuristic Programming (DHP), and backpropagation through time (BPTT), as we will discuss. The second premise is our emphasis on a development of a truly on-line critic-based training procedure competitive in performance and computational cost to truncated BPTT. Three examples illustrate the main features of the framework proposed.
_____
ICNN97 Special Session on Adaptive Critic Designs Session: SS1B Paper Number: 710 Oral
Unified formulation for training recurrent networks with derivative adaptive critics
L. A. Feldkamp, G. V. Puskorius and D. V. Prokhorov
Keywords: training recurrent networks derivative adaptive critics
Abstract:
We present a procedure for obtaining derivatives used in training a recurrent network that combines in a unified framework the techniques of backpropagation through time and derivative adaptive critics. The resulting formulation is consistent with previous descriptions, but has the advantage of allowing the mentioned techniques to be used together in a proportion that is appropriate to a given problem.
_____
ICNN97 Special Session on Adaptive Critic Designs Session: SS1C Paper Number: 708 Oral
Asymptotic dynamic programming: Preliminary concepts and results
R. Saeks, C. Cox and A. Maren
Keywords: Asymptotic dynamic programming Adaptive critic nonlinear dynamical systems
Abstract:
A formal setting for the development of adaptive critic techniques is established in a nonlinear dynamical systems setting and some preliminary stability and suboptimality theorems are developed. Two alternative versions of the theory are developed, one with a known plant and one with an initial unknown plant which must be identified on-line.
_____
ICNN97 Special Session on Adaptive Critic Designs Session: SS1D Paper Number: 707 Oral
The need for improved reinforcement learning techniques in intelligent agents
Donald C. Wunsch II
Keywords: reinforcement learning intelligent agents neural networks
Abstract:
Reinforcement learning is an integral part of intelligent agent research. The development of this field, however, has been largely independent of the latest developments in neural networks. As a result, the most popular designs for intelligent agents utilize neural network architectures from several years ago. This article recommends newer, proven designs for reinforcement learning. The recommended designs share historical roots with the most popular architectures in place today, allowing improved performance without radical redesign of existing agents.
_____
ICNN97 Special Session on Adaptive Critic Designs Session: SS1E Paper Number: 706 Oral
Immunized adaptive critics
K. KrishnaKumar and J. Neidhoefer
Keywords: Adaptive critic Immunized computational systems intelligent control
Abstract:
Immunized Computational Systems combine a priori knowledge with the adapting capabilities of immune systems to provide a powerful alternative to currently available techniques for intelligent control. In this paper, we present our perspective on various levels of intelligent control and relate them to similar functioning in human immune systems. A technique for implementing immunized computational systems as adaptive critics is presented and the technique is then applied to a control problem to document the effectiveness of this approach.
_____
ICNN97 Special Session on Adaptive Critic Designs Session: SS1F Paper Number: 705 Oral
Adaptive critic based neurocontroller for autolanding of aircrafts with varying glideslopes
Gaurav Saini and S.N. Balakrishnan
Keywords: neurocontroller autolanding of aircrafts varying glideslopes Adaptive critic
Abstract:
VISUAL SYSTEM MODELS & PROSTHESES (Return to Top)
ICNN97 Visual System Models & Prostheses Session: SS2.1A Paper Number: 732 Oral
Top-down and bottom-up image processing
Lawrence W. Stark and Claudio Privitera
Keywords: Top-down image processing bottom-up image processing scanpath theory
Abstract:
We introduce the problem of fitting a picture to the eye, by referring to the scanpath theory for top-down vision in humans and discuss how even such early processes as segmentation are likely controlled top-down by internal cognitive-spatial models. Experimental evidence is adduced for this view of human vision. The internal model must approximate the external world, since our species has survived. Alternately, certain bottom-up image processing algorithms may inversely serve to identify approximate sub-features or regions of interest of the internal model. Our results indicate, however, that these algorithms can not predict the sequential ordering of the subfeatures used by a person (L). Applications of schemata exploiting the above notions are suggested for image compression; guided viewing of computer screens for increased efficiency; and compensating distortions for such eye lesions as scotomas.
_____
ICNN97 Visual System Models & Prostheses Session: SS2.1B Paper Number: 731 Oral
Introducing visual latencies into spin-lattice models for image segmentation: a neuromorphic approach to a computer vision problem
Ralf Opara and Florentin Woergoetter
Keywords: visual latencies spin-lattice models image segmentation computer vision
Abstract:
In this study we will show how an algorithmic principle which might play a role in information processing in the brain of higher vertebrates - the so called "visual latencies" - can be transferred with high efficiency to a model system which is better suited for implementation on conventional computer hardware. To this end we assign luminance dependent temporal delays (latencies) to the individual pixels of an image. This temporal structure of the input data stream then accelerates and improves the relaxation of a spin-lattice labeling algorithm for scene segmentation.
_____
ICNN97 Visual System Models & Prostheses Session: SS2.1C Paper Number: 730 Oral
Artificial retina chips
Kazuo Kyuma and Yasunari Miyake
Keywords: Artificial retina chips interactive games
Abstract:
_____
ICNN97 Visual System Models & Prostheses Session: SS2.1D Paper Number: 729 Oral
Physiological engineering model of the outer retina
Shiro Usui, Yoshimi Kamiyama, Toshihiko Ogura, Akito Ishihara and Tomoya Hamada
Keywords: Physiological engineering outer retina ionic current
Abstract:
In the vertebrate retina, the outer plexiform layer which consists of synaptic connections among photoreceptors, horizontal cells and bipolar cells, plays a fundamental role in color- and spatial-information processings. Developing a quantitative model of the outer retina is essential toward better understanding of the outer plexiform layer. Here we propose ionic current models of photoreceptor, horizontal cell and bipolar cells based on the published experimental data. The models provide a better understanding of the functional role of the ionic currents in outer retinal cells in generating the electrical responses. We also discuss how the function of the retinal outer plexiform layer may be studied by constructing a network model from the ionic current based single cell models.
_____
ICNN97 Visual System Models & Prostheses Session: SS2.2A Paper Number: 735 Oral
Dialog concepts for learning retina encoders
R. Eckmiller, M. Becker and R. Hunermann
Keywords: Dialog concepts retina encoders blindness retina stimulator
Abstract:
_____
ICNN97 Visual System Models & Prostheses Session: SS2.2B Paper Number: 733 Oral
The physiological connection: stimulating the human and amphibian retina
Gislin Dagnelie, Mark Humayun, Robert Greenberg and Eugene de Juan Jr.
Keywords: human retina amphibian retina blindness visual impairment
Abstract:
_____
ICNN97 Visual System Models & Prostheses Session: SS2.2C Paper Number: 734 Oral
The cellular nonlinear network as a retinal camera for visual prosthesis
Frank S. Werblin, Adam Jacobs and Tama Roska
Keywords: cellular nonlinear network retinal camera visual prosthesis
Abstract:
A retinal camera-on-a-chip is proposed. The chip is used for two purposes: 1) It provides a computationally powerful camera to drive a visual prosthetic and 2) it serves as a "hypothesis generator" for proving the patterns of activity that must be generated by the camera. The design of the chip is based on CNN technology a massively parallel analog processor of enormous power. The patterns are based upon physiological studies of the massively parallel output patterns measured in living retinas. The device will be capable of generating a series of simultaneous retina-like output patterns that can be introduced either at the optic nerve or visual cortex as an interactive visual prosthetic camera, controlled by both the user and the visual environment.
_____
ICNN97 Visual System Models & Prostheses Session: SS2.2D Paper Number: 728 Oral
Color and intensity information representations by a network of turtle retinal ganglion cells
Richard A. Normann, Josef Ammermuller, Shy Shoham and Almut Branner
Keywords: Color and intensity information representatio neural network turtle retinal ganglion cells
Abstract:
NEURO-FUZZY INTEGRATION (Return to Top)
ICNN'97 Neuro-fuzzy Integration Session: SS3A Paper Number: 747 Oral
A fuzzy neural hybrid system modeling
I. Burhan Turksen
Keywords: fuzzy neural integration unsupervised learning fuzzy sets fuzzy clustering
Abstract:
_____
ICNN'97 Neuro-fuzzy Integration Session: SS3B Paper Number: 714 Oral
Fuzzy inference networks: an introduction
W. Pedrycz and M. H. Smith
Keywords: Fuzzy inference networks fuzzy sets neural networks fuzzy inference
Abstract:
One of the problems of existing fuzzy-neural approaches is that the logic nature of the structure is often lost, i.e., what is being processed by the neural networks becomes irrelevant. To retain this logic content while benefiting from the advantage of integrating fuzzy set and neural network approaches, we propose in this paper a fuzzy neural network which supports fuzzy inference mechanisms by being based exclusively on logic implication neurons. A supervised learning method involving an equality index performance measure and an on-line update delta rule (gradient-based) learning procedure is used. An experimental study involving Wolfer's sunspot numbers is carried out, demonstrating faster convergence accompanied by explicit format of the inference network.
_____
ICNN'97 Neuro-fuzzy Integration Session: SS3C Paper Number: 711 Oral
Neural nets can be universal approximators for fuzzy functions
J. J. Buckley and Yoichi Hayashi
Keywords: Neural nets universal approximators fuzzy functions
Abstract:
We first argue that the extension principle is too computationally involved to be an efficient way for a computer to evaluate fuzzy functions. We then suggest using alpha-cuts and interval arithmetic to compute the value of fuzzy functions. Using this method of computing fuzzy functions, we then show that neural nets are universal approximators for (computable) fuzzy functions, when we only input non-negative, or non-positive, fuzzy numbers.
_____
ICNN'97 Neuro-fuzzy Integration Session: SS3D Paper Number: 713 Oral
Speaker recognition with a self-configuring neural network
Jie lei and Lawrence O. Hall
Keywords: Speaker recognition self-configuring neural network LBG clustering algorithm
Abstract:
This paper discusses preliminary work on a promising method for recognizing speakers. A self-configuring neural network is trained to recognize sentences that have been compressed by the LBG clustering algorithm. The bias weights of the trained neural networks are adjusted to minimize the false positive percentage. Recognition results from the TIMIT speech database of greater than 90% correct are obtained with no false positives. The results presented here provide a basis for the generation of secure speaker recognition systems which use neural networks.
_____
ICNN'97 Neuro-fuzzy Integration Session: SS3E Paper Number: 761 Oral
Neural net approximations to solutions of systems of fuzzy linear equations
J.J. Buckley and Yoichi Hayashi
Keywords: fuzzy linear equations neural net approximations sign restrictions
Abstract:
This paper continues our research into using neural nets to solve fuzzy problems. We show how to train neural nets, with certain sign constraints on their weights, using genetic algoithms, to approximate solutions to systems of fuzzy linear equations. This paper presents a new application of layed, feedforward, neural nets with sign restrictions on their weights.
_____
ICNN'97 Neuro-fuzzy Integration Session: SS3F Paper Number: 712 Oral
On tool wear estimation through neural networks
Guenter Luetzig, Manuel Sanchez-Castillo and Reza Langari
Keywords: intelligent monitoring sensor data fusion tool wear monitoring recurrent neural networks
Abstract:
During metal cutting operations as machining progresses and the tool wears out, the surface quality and the dimensional accuracy of the product degrade. Estimates of tool life based on experience or past tool wear data are usually very conservative, and hence tools are generally underutilized. Due to the very complex nature of tool wear, it is not possible to achieve a reliable and consistent tool wear monitoring scheme by using information from a single sensor. In our research we are developing, for end milling operations, a neural network based indirect method for on-line continuous estimation of tool flank wear.
In this paper we present the simulation results of a performance driven design study, of the recurrent part of the network proposed by Kamarthi et al. (1995). The aim of this study is to improve the performance of the data fusion algorithm in order to generate more accurate final flank wear estimates. Testing of the recurrent network has proved its ability to properly integrate the first level flank wear estimates into a reliable final flank wear estimate. Using an architecture with one delayed output and one additional delayed input vector improves the performance of the network.
LINGUISTIC RULE EXTRACTION (Return to Top)
ICNN97 Linguistic Rule Extraction Session: SS4A Paper Number: 719 Oral
A study in experimental evaluation of neural network and genetic algorithm techniques for knowledge acquisition in fuzzy classification systems
Ilona Jagielska, Chris Matthews and Tim Whitfort
Keywords: neural network genetic algorithm knowledge acquisition fuzzy classification systems
Abstract:
This paper addresses the issue of appropriate evaluation criteria for knowledge acquisition techniques for fuzzy classification systems. It describes an empirical study in which two different systems, one based on neural networks, and the other based on genetic algorithms were developed, applied to three classification problems and evaluated. Comparison of the approaches with the C4.5 inductive algorithm was also carried out.
_____
ICNN97 Linguistic Rule Extraction Session: SS4B Paper Number: 720 Oral
Acquisition of fuzzy rules using fuzzy neural networks with forgetting
Motohide Umano, Shiro Fukunaka, Itsuo Hatono and Hiroyuki Tamura
Keywords: fuzzy rule Acquisition fuzzy neural networks forgetting
Abstract:
We acquire fuzzy rules from data using a fuzzy neural network. First, we generate an initial fuzzy neural network of the specified number of fuzzy rules that have the less number of good membership functions generated using a self-organization algorithm by T.~Kohonen. Then, we tune and prune fuzzy rules based on a structural leaning algorithm with forgetting by M.~Ishikawa, where the numerals in the consequent part and the center values and widths of membership functions in the antecedent part are tuned and forgotten a little, and thus redundant rules and variables are pruned to acquire simpler, general rules. We apply the method to the iris classification problem by R.A.~Fisher and have a very good result.
_____
ICNN97 Linguistic Rule Extraction Session: SS4C Paper Number: 718 Oral
Knowledge extraction from hierarchical fuzzy model obtained by fuzzy neural networks and genetic algorithm
Takeshi Furuhashi, Seiichi Matsushita, Hiroaki Tsutsui, Yoshiki Uchikawa
Keywords: Knowledge extraction hierarchical fuzzy model fuzzy neural networks genetic algorithm
Abstract:
Knowledge extraction from trained artificial neural network has been studied by many researchers. Modeling of nonlinear systems using Fuzzy Neural Networks (FNN) is a promising approach to the knowledge acquisition, and FNN is specially designed for knowledge extraction. The authors have proposed a hierarchical fuzzy modeling method using FNNs and Genetic Algorithm (GA). This method can identify fuzzy models of nonlinear objects with strong nonlinearity. The disadvantage of the method is that the training of FNN is time consuming. This paper presents a quick method for rough search for proper structures in the antecedent of fuzzy models. The fine tuning of the acquired rough model is done by the FNN. This modeling method is quite efficient to identify precise fuzzy models of systems with strong nonlinearities. A simulation is done to show the effectiveness of the proposed method.
_____
ICNN97 Linguistic Rule Extraction Session: SS4D Paper Number: 717 Oral
Fuzzy rule extraction, reasoning and rules adaptation in fuzzy neural networks
Nikola K. Kasabov
Keywords: Fuzzy rule extraction Fuzzy reasoning rules adaptation fuzzy neural networks
Abstract:
_____
ICNN97 Linguistic Rule Extraction Session: SS4E Paper Number: 716 Oral
Extraction of crisp logical rules using constrained backpropagation networks
Wlodzilslaw Duch, Rafal Adamczak, and Krzysztof Grabczewski
Keywords: crisp logical rules constrained backpropagation networks rule extraction
Abstract:
_____
ICNN97 Linguistic Rule Extraction Session: SS4F Paper Number: 715 Oral
Linguistic rule extraction from neural networks and genetic-algorithm-based rule selection
Hisao Ishibuchi, Manabu Nii and Tadahiko Murata
Keywords: rule extraction rule selection fuzzy reasoning genetic algorithm
Abstract:
In this paper, we propose a hybrid approach to the design of a compact fuzzy rule-based classification system with a small number of linguistic rules. The proposed approach consists of two procedures: rule extraction from a trained neural network and rule selection by a genetic algorithm. First, we describe how linguistic rules can be extracted from a multilayer feedforward neural network that has been already trained for a classification problem with many continuous attributes. Next we explain how a genetic algorithm can be utilized for selecting a small number of significant linguistic rules from a large number of extracted rules. Our rule selection problem has two objectives: to minimize the number of selected linguistic rules and to maximize the number of correctly classified patterns by the selected linguistic rules. Finally we illustrate our hybrid approach by computer simulations on real-world test problems.
INTELLIGENT CONTROL, THEORY & APPLICATIONS (Return to Top)
ICNN97 Intelligent Control, Theory & Applications Session: SS5A Paper Number: 725 Oral
Robust NLq neural control theory
J.A.K. Suykens, B. De Moor and J. Vandewalle
Keywords: multilayer recurrent neural networks neural control NLq systems real parametric uncertainty
Abstract:
In this paper we present sufficient conditions for global asymptotic stability and I/O stability (finite L2-gain) of multilayer recurrent neural networks, with parametric uncertainties upon the interconnection matrices. This is done by considering perturbed NLq systems.
NLq's are discrete time nonlinear dynamical systems in state space form, containing q layers with alternating linear and static nonlinear operators that satisfy a sector condition. Within the present framework, Linear Fractional Transformations with real diagonal uncertainty block are interpreted as perturbed NLq's with q=1.
While in mu robust control theory uncertainties upon nominal linear models are investigated, uncertainties upon nominal nonlinear models can be studied in NLq neural control theory.
It is shown how the state space upper bound test of mu theory, related to the case of a real diagonal uncertainty block, follows as a special case from the derived Theorems for q=1 (one layer).
Keywords. multilayer recurrent neural networks, neural control, NLq systems, real parametric uncertainty, (robust) global asymptotic stability, (robust) I/O stability, finite L2-gain, LFTs, mu theory.
_____
ICNN97 Intelligent Control, Theory & Applications Session: SS5B Paper Number: 723 Oral
Feedback-error-learning control with considering smoothness of unknown nonlinearities
Yasuaki Kuroe, Hidehisa Inayoshi and Takehiro Mori
Keywords: Feedback-error-learning control inverse models forward models
Abstract:
In recent years, learning control of nonlinear systems by using neural networks has been widely studied. Among them the feedback-error-learning control proposed by M. Kawato et al. has been recognized to be an excellent learning method because of the fact that this method makes it possible to realize inverse models of unknown nonlinear controlled objects on neural networks. Since forward or inverse models of controlled objects, in general, are expressed by nonlinear smooth functions, taking account of the smoothness of forward or inverse models in the learning control would improve its performance considerably. In this paper the feedback-error-learning control is extended so as to being able to treat the smoothness of unknown nonlinearity of controlled objects. The proposed method makes it possible to realize inverse models more accurately and to attain more precise control.
_____
ICNN97 Intelligent Control, Theory & Applications Session: SS5C Paper Number: 726 Oral
Fuzzy logic controllers generated by pseudo-bacterial genetic algorithm with adaptive operator
Tomonori Hashiyama, Norberto Eiji Nawa, Takeshi Furuhashi and Yoshiki Uchikawa
Keywords: fuzzy modeling genetic algorithm bacterial genetics hybrid systems
Abstract:
This paper presents a new genetic operator called adaptive operator to improve local portions of chromosomes. This new operator is implemented into Pseudo-Bacterial Genetic Algorithm (PBGA). The PBGA was proposed by the authors as a new approach combining a genetic algorithm (GA) with a local improvement mechanism inspired by a process in bacterial genetics. The PBGA was applied for the acquisition of fuzzy rules. The aim of the newly introduced adaptive operator is to improve the quality of the generated rules of the fuzzy models, producing blocks of effective rules and more compact models.
The new operator adaptively decides the division points of each chromosome for the bacterial mutation and the cutting points for the crossover, according to the distribution of degrees of truth values of the rules. In this paper, first, results obtained when using the PBGA with the adaptive operator for a simple fuzzy modeling problem are presented. Second, the PBGA with adaptive operator is used in the design of a fuzzy logic controller for a semi-active suspension system. The results show the benefits obtained with this operator.
_____
ICNN97 Intelligent Control, Theory & Applications Session: SS5D Paper Number: 722 Oral
A sliding mode controller with neural network and fuzzy logic
Minho Lee
Keywords: sliding mode controller neural network fuzzy logic
Abstract:
A sliding mode controller with a neural network and a fuzzy boundary layer is proposed. A multilayer neural network is used for constructing the inverse identifier which is an observer of the uncertainties of a system. Also, fuzzy boundary layer is introduced to make the continuous control input of sliding mode controller combined with the neural inverse identifier. The proposed control scheme not only reduces an effort for finding an unknown dynamics of a system but also alleviates the chattering problems of the control input. Computer simulation reveals that the proposed approach is effective to alleviate the chattering problem of the control input.
_____
ICNN97 Intelligent Control, Theory & Applications Session: SS5E Paper Number: 721 Oral
Neuro-approach for intelligent systems development
Sigeru Omatu and Mitchifumi Yoshioka
Keywords: neuro-control PID control intelligent control inverted pendulum
Abstract:
In this paper, we propose a method to use the neural networks to tune the PID(proportional plus integral plus derivative) gains such that human operators tune the gains adaptively according to the environmental condition and systems specification. The tuning method is based on the error back-propagation method, which is abbreviated by the BP method and hence, it may be trapped in a local minimum. In order to avoid the local minimum problem, we use the genetic algorithm to find the initial values of the connection weights of the neural network and initial values of PID gains. The experimental results show the effectiveness of the present approach.
_____
ICNN97 Intelligent Control, Theory & Applications Session: SS5F Paper Number: 724 Oral
Adaptive control and stability analysis of nonlinear systems using neural networks
Osamu Yamanaka, Naoto Yoshizawa, Hiromitsu Ohmori and Akira Sano
Keywords: Adaptive control stability analysis nonlinear systems neural networks
Abstract:
This paper is concerned with new neural-network (NN)-based adaptive control schemes for a class of nonlinear system which includes a finite Volterra series system and a Wiener system. First, introducing a new kind of dynamic neural network which consists of Laguerre filters and memoryless nonlinear elements, a model reference adaptive control (MRAC) scheme is presented for the nonlinear systems. In the proposed MRAC system adopting overparameterization and a robust adaptive algorithm, the boundedness of the estimated parameters is assured under some conditions. Second, an adaptive linearization scheme for Wiener systems with nonlinearity in their output part is realized by using a kind of functional-link network. It is shown that the obtained controller has a structure similar to the MRAC and then the boundedness of the estimated parameters as well as that of all the signals in the closed loop are also ensured. Finally, the effectiveness of the proposed schemes is illustrated through numerical simulations.
_____
ICNN97 Intelligent Control, Theory & Applications Session: SS5G Paper Number: 727 Oral
temperature control in a batch process by neural networks
Takeshi Iwasa, Noboru Morizumi, and Sigeru Omatu
Keywords: PID control neural network chemical plant batch process
Abstract:
NEURAL NETWORKS APPLICATIONS FOR MONITORING OF COMPLEX SYSTEMS (Return to Top)
ICNN97 NN Applications for Monitoring of Complex Systems Session: SS6A Paper Number: 703 Oral
An ANS based helicopter transmission diagnostic system
Xiaoshu Xu, Hans Vanderveldt and Robert Allen
Keywords: Artificial Neural System helicopter transmission diagnostic system mechanical diagnostics
Abstract:
In 1994 the Office of Naval Research (ONR) proposed a provocative question: Can ANS technology recognize sounds of impending failure in a high speed gearbox in real time. ONR possessed sound sensor readings from a CH-46 helicopter gearbox. The data was from 30 minutes of test stand flight activity and contained each of the six different fault types. AJI was provided the data in its "raw" form with no preprocessing. The answer is a resounding yes, ANS technology can in real time recognize impending mechanical faults in a high speed, high noise helicopter gearbox. The achievements of AJI are best summarized by Dr. Thomas McKenna, ONR Program Officer, "Initially, we were skeptical that you [AJI] could achieve robust and accurate fault identification using only the raw time series, since nearly every other approach uses spectral feature representation. Your success will undoubtedly encourage other investigators to reexamine the value of the time series inputs to neural systems. Your approach may lead to less complex and more cost effective hardware implementations of mechanical diagnostic systems."
_____
ICNN97 NN Applications for Monitoring of Complex Systems Session: SS6B Paper Number: 701 Oral
A neural architecture for fuzzy classification with application to complex system tracking
Patrick A. Stadter and Amulya K. Garga
Keywords: neural architecture fuzzy classification complex system tracking Voronoi diagram
Abstract:
The application of a new architecture for fuzzy pattern classification is described to address the problem fo tracking complex, discrete event driven systems. The classifier relies upon the integration of fuzzy logic techniques with an artificial neural architecture to produce an efficient mechanism for classifying input patterns. In addition, the fuzzy classifier provides a method for quantifying and handling ambiguity near the decision surfaces.
The proposed application consists of classifying input feature patterns as events which drive complex, dynamic systems modeled as discrete event systems.
______
ICNN97 NN Applications for Monitoring of Complex Systems Session: SS6C Paper Number: 704 Oral
Learning the correlation of beam position monitors for Pohang synchrotron light source using neural networks
Jae Woo Lee and Sungzoon Cho
Keywords: Learning correlation of beam position monitors synchrotron light source neural networks
Abstract:
Neural networks are used to learn the correlation of the beam position monitors (BPM) which trace the electron beam orbit in the storage ring of the Pohang Synchrotron Light Source. Since a beam in the storage ring passes through many BPMs, there is a correlation among the measurements of those monitors. A perceptron is trained to predict one BPM's measurement given other BPMs' measurements. If the predicted value of a perceptron is different from the actual measurement, the corresponding BPM can be considered to have a fault. Test results indicate that the neural network approach has a potential for actual fault diagnosis of BPM. Compared to the current diagnosis methods, the neural network approach is more economical and less disruptive. It is shown to perform better than a numerical approach.
_____
ICNN97 NN Applications for Monitoring of Complex Systems Session: SS6D Paper Number: 702 Oral
Intelligent hybrid multi-agent architecture for engineering complex systems
R. Khosla and T. Dillon
Keywords: Intelligent hybrid multi-agent architecture engineering complex systems real time alarm processing
Abstract:
BIOMEDICAL APPLICATIONS (Return to Top)
ICNN97 Biomedical Applications Session: SS7.1A Paper Number: 743 Oral
Neural network based segmentation using a priori image models
S. Sanjay Gopal, Berkman Sahiner, Heang-Ping Chan and Nicholas Petrick
Keywords: Neural network segmentation a priori image models
Abstract:
We examine image segmentation using a Hopfield neural network. Image segmentation is posed as an optimization problem, and is correlated with the energy function of the neural network. By carefully designing the optimization criterion for segmentation, it is possible to identify the bias inputs and the interconnection weights of the corresponding neural network. We provide a general framework for the design of the optimization criterion, which consists of a component based on the observed image, and another component based on an a priori image model. As an application, we consider a smoothness constraint for the segmented image as our a priori information, and solve a gray-level based segmentation problem. The feasibility of using the neural network architecture based on this optimization criterion for the segmentation of masses in mammograms is demonstrated.
_____
ICNN97 Biomedical Applications Session: SS7.1B Paper Number: 739 Oral
Feature selection and classification for the computerized detection of mass lesions in digital mammography
Matthew A. Kupinski and Maryellen L. Giger
Keywords: Feature classification Feature selection mass lesions digital mammography
Abstract:
We have investigated various methods of feature selection for two different data classifiers used in the computerized detection of mass lesions in digital mammograms. Numerous features were extracted from abnormal and normal breast regions from a database consisting of 210 individual mammograms. A stepwise method, a genetic algorithm and individual feature analysis were employed to select a subset of features to be used with linear discriminants. Similar techniques were also employed for an artificial neural network classifier. In both tests the genetic algorithm was able to either outperform or equal the performance of other methods.
_____
ICNN97 Biomedical Applications Session: SS7.1C Paper Number: 740 Oral
An artificial neural network to predict mortality in patients who undergo percutaneous coronary interventions
Georgia D. Tourassi and Nicholas P. Xenopoulos
Keywords: artificial neural network mortality prediction percutaneous coronary interventions
Abstract:
The objective of this study was to develop a method for identifying patients at increased risk for mortality after percutaneous coronary interventions (PCI). Although the mortality rate after PCI is low (1-2%), the ability to predict the patients with increased risk of mortality can alter the preferred medical strategy and potentially improve the outcome of the patient. We developed a feed-forward artificial neural network (ANN) which predicts mortality using 24 variables. The study was based on 812 consecutive patients who underwent PCI between 1.1.95 and 6.30.95 at the Jewish Hospital Heart and Lung Center, Louisville, KY. The predictive power of the network was compared to that of linear discriminant analysis (LDA) using Receiver Operating characteristics (ROC) methodology. Our study showed that the performance of the network strongly depended on the choice of the criterion function. Specifically, a modified cross-entropy function worked the best for the network resulting in an ROC area index of Az(ANN) = 0.84 ± 0.07 compared to Az(LDA) = 0.64 ± 0.12.
_____
ICNN97 Biomedical Applications Session: SS7.1D Paper Number: 742 Oral
Neural network design for optimization of the partial area under the receiver operating characteristic curve
Berkman Sahiner, Heang-Ping Chan, Nicholas Petrick, S. Sanjay Gopal and Mitchell M. Goodsitt
Keywords: neural network design optimization receiver operating characteristic curve
Abstract:
A new backpropagation training algorithm was developed for the maximization of the area under the Receiver Operating Characteristic (ROC) curve between two user-specified true-positive fraction thresholds. The algorithm was used to design a neural network classifier with high specificity at the high-sensitivity region of the ROC curve, which is of particular interest for computer-aided diagnosis applications. The effectiveness of the algorithm was demonstrated with a simulation study.
_____
ICNN97 Biomedical Applications Session: SS7.1E Paper Number: 741 Oral
Self-organizing maps for analyzing mammographic findings
Joseph Y. Lo and Carey E. Floyd, Jr.
Keywords: Self-organizing maps mammographic findings analysis neural networks
Abstract:
The purpose of this study is to analyze mammographic findings using self-organizing map (SOM) artificial neural networks. Using two findings of patient age and mass margin extracted by radiologists, self-organizing maps were developed to analyze both the distribution and topology of the input findings. These results can help to explain the underlying nature of mammographic findings data, which may in turn help radiologists to improve breast cancer diagnosis and assist in the development of other neural networks.
_____
ICNN97 Biomedical Applications Session: SS7.2A Paper Number: 737 Oral
Neurometric assessment of adequacy of intraoperative anesthetic
Lars J. Kangas, Paul E. Keller, Carlton M. Cadwell, Rick Webber, Polly Pierce and Harvey L. Edmonds
Keywords: intraoperative anesthetic EEG signals supervised artificial neural networks
Abstract:
_____
ICNN97 Biomedical Applications Session: SS7.2B Paper Number: 738 Oral
A three dimensional neural architecture
Evangelia Micheli-Tzanakou and Timothy J. Dasey
Keywords: three dimensional neural architecture lateral connections feedback connections
Abstract:
A new neural network architecture is presented which encompasses functions similar to those in a biological brain, such as lateral and feedback connections and neurons. Neurons are randomly distributed on 2D-planes. Each neuron on each plane can connect to a neighborhood of neurons at the next layer (plane), as well as receive feedback from neurons on that layer, or any other layer in the immediate or distant vicinity. In addition, lateral inhibitory connectivity within a layer adds to the flexibility and generalization abilities of the neural network.
_____
ICNN97 Biomedical Applications Session: SS7.2C Paper Number: 736 Oral
Visual ophtalmologist: an automated system for classification of retinal damage
Sergey Aleynikov and Evangelia Micheli-Tzanakou
Keywords: Visual ophtalmologist image classification retinal damage
Abstract:
The objective of this research is to provide an ophthalmologist with a helpful system, capable of classifying a degree of patients' retinal hemorrhage. The system is composed of four modules: a) data acquisition module, b) image Database module, c) image processing module, d) image classification module. The system was trained with a modular neural network on a set of 25 images, and tested on a set of 160 images. A training performance of greater than 95% was achieved. The classifying part of the system showed 79% recognition accuracy. Since the testing images were taken from independent sources, we assume that the system should also provide an accurate classification of other image types.
_____
ICNN97 Biomedical Applications Session: SS7.2D Paper Number: 744 Oral
Genetic function approximation in the molecular pharmacology of cancer
Leming M. Shi, Yi Fan, Timothy G. Myers and John N. Weinstein
Keywords: Genetic function approximation molecular pharmacology of cancer cancer breast cancer
Abstract:
The National Cancer Institute's Developmental Therapeutics Program screens more than 10,000 compounds per year for their ability to inhibit growth of 60 human cancer cell lines. Using a combination of cross-validated back-propagation neural networks and multivariate statistical methods, we found that a compound's mechanism of action could be predicted with considerable accuracy solely on the basis of its pattern of growth inhibitory activity against the 60 cell lines (Weinstein, et al., Science 258: 447, 1992, Weinstein, et al., Science 275: 343, 1997). Over the last several years, the developments, in terms of different mathamatical approaches, led to formulation of a general "information-intensive" strategy for drug discovery that integrates data on a compound's molecular structure, pattern of growth inhibitory activity, and possible molecular targets in the cell. Here we will summarize our recent investigations of a new approach to the regression problem, "genetic function approximation" (GFA).
Keywords: genetic function approximation, cancer, molecular pharmacology, QSAR
_____
ICNN97 Biomedical Applications Session: SS7.2E Paper Number: 752 Oral
Evaluating artificial neural networks for medical application
Harry B. Burke
Keywords: artificial neural networks model accuracy significance testing
Abstract:
The validity and usefulness of an artificial neural network depends on whether an appropriate measure is used to assess its accuracy and whether the artificial neural network is significantly more accurate than traditional statistical models for the medical task. We discuss a method for determining whether an artificial neural network is necessary for the task, how to select the most appropriate measure of model accuracy, and the importance of significance testing.
SENSORS AND BIOSENSORS (Return to Top)
ICNN97 Sensors and Biosensors Session: SS8A Paper Number: 762 Oral
Neural network assisted drug detection in x-ray images
Gregg D. Wilensky, Narbik Manukian, John L. Kirkwood, and Jung-Chou Chang
Keywords: drug detection x-ray images neural networks
Abstract:
______
ICNN97 Sensors and Biosensors Session: SS8B Paper Number: 748 Oral
Detection and discrimination using X radiation
David G. Brown and Robert J. Jennings
Keywords: X radiation X-ray imaging reconstruction imaging multiple-energy imaging
Abstract:
X-ray imaging modalities such as projection radiography and computed tomography are critical members of the miraculous armamentarium of modern diagnostic medical imaging. They are also of great importance for nondestructive testing in other fields. This paper examines the fundamental principles underlying x-ray imaging, describes several key aspects of x-ray imaging systems, and indicates areas in which neural network methods may be expected to be important. Special attention is paid to relatively sophisticated techniques such as reconstruction imaging and multiple-energy imaging, and several concepts from imaging science are described. New modalities such as phase imaging and coherent scatter imaging are introduced. Neural network methods have significant potential utility for the operation and evaluation of x-ray imaging systems. In addition they should prove vitally important for the analysis of the tremendous quantities of data which these systems generate. Of course, these conclusions are equally applicable to other types of imaging, including nuclear tracer, magnetic resonance, and ultrasound.
_____
ICNN97 Sensors and Biosensors Session: SS8C Paper Number: 749 Oral
Pulse coupled neural networks (PCNN) and wavelets: Biosensor applications
Mary Lou Padgett and John L. Johnson
Keywords: pulse coupled neural network wavelets biosensor factor classification
Abstract:
This paper focuses on the novel approaches to chemosensor signal analysis. 1) forming image patterns from the time sequences, 2) PCNN Factor formation, and 3) Factor classification using wavelets.
_____
ICNN97 Sensors and Biosensors Session: SS8D Paper Number: 750 Oral
Blind De-mixing with unknown sources
Harold Szu and Charles Hsu
Keywords: multisensor data processing multispectral data processing Lagrangian-Hopfield neural network Hopfield energy function
Abstract:
KNOWLEDGE-BASED METHODS IN NEURAL NETWORKS (Return to Top)
ICNN97 Knowledge-based Methods in NN Session: SS9.1A Paper Number: 754 Oral
On the conscious and subconscious components of knowledge representation in neural networks
Robert Kozma
Keywords: neural networks knowledge representation learning
Abstract:
Principles of learning and knowledge representation are studied in complex neural networks with a large number of parameters. Our neural networks incorporate both deep and shallow knowledge representations.
In the case of stationary environment, the neural nets can develop a deep understanding of the problem and working in a properly chosen, narrow subspace of the system variables. In this functional mode, a small number of parameters dominate the operation of the network on the surface (conscious components), while the overwhelming part of the network operates at a the subconscious level. The components realizing the conscious operational task are quite stable, nevertheless, they might change in time if the external environment of the analyzed system varies.
_____
ICNN97 Knowledge-based Methods in NN Session: SS9.1B Paper Number: 759 Oral
Some issues in system identification using clustering
Nikhil R. Pal, Kuhu Pal, James C. Bezdek and Thomas A. Runkler
Keywords: system identification clustering fuzzy reasoning system
Abstract:
This article surveys the use of clustering for identification of various parameters of fuzzy systems. Issues discussed include the proper domain for clustering, the clustering algorithm used, validation of clusters, and system validation.
_____
ICNN97 Knowledge-based Methods in NN Session: SS9.1C Paper Number: 753 Oral
The truth is in there: Current issues in extracting rules from trained artificial neural networks
Alan B. Tickle, Mostefa Golea, Ross Hayward and Joachim Diedrich
Keywords: artificial neural networks training rule extraction
Abstract:
A recognized impediment to the more widespread utilization of Artificial Neural Networks (ANNs) is the absence of a capability to explain, in a human-comprehensible form, either the process by which a trained ANN arrives at a specific decision/result or, in general, the totality of knowledge embedded therein. Recently there has been a proliferation of techniques aimed at redressing this situation and, in particular, for extracting the knowledge embedded in trained feedforward ANNs as sets of symbolic rules. However, if the dissemination of ideas in the field of ANN rule extraction is to proceed in a systematic manner, then it is essential that a rigorous taxonomy exists for categorizing the plethora of techniques being developed. This paper shows how one of the proposed schemas for categorizing ANN rule extraction techniques is able to accommodate such recent developments in the field. In addition attention is drawn to what are seen to be some of the key challenges in the area including the identification of factors which appear to limit what is actually achievable through the rule extraction process.
_____
ICNN97 Knowledge-based Methods in NN Session: SS9.1D Paper Number: 758 Oral
Asynchronous translations with recurrent neural nets
Ramon P. Neco and Mikel L. Forcada
Keywords: asynchronous translation recurrent neural networks finite-state machines finite-state recognizers
Abstract:
In recent years, many researchers have explored the relation between discrete-time recurrent neural networks (DTRNN) and finite-state machines (FSMs) either by showing their computational equivalence or by training them to perform as finite-state recognizers from examples. Most of this work has focussed on the simplest class of deterministic state machines, that is deterministic finite automata and Mealy (or Moore) machines. The class of translations these machines can perform is very limited, mainly because these machines output symbols at the same rate as they input symbols, and therefore, the input and the translation have the same length; one may call these translations synchronous. Real-life translations are more complex: word reorderings, deletions, and insertions are common in natural-language translations; or, in speech-to-phoneme conversion, the number of frames corresponding to each phoneme is different and depends on the particular speaker or word. There are, however, simple deterministic, finite-state machines (extensions of Mealy machines) that may perform these classes of ``asynchronous'' or ``time-warped'' translations. A simple DTRNN model with input and output control lines inspired on this class of machines is presented and successfully applied to simple asynchronous translation tasks with interesting results regarding generalization. Training of these nets from input-output pairs is complicated by the fact that the time alignment between the target output sequence and the input sequence is unknown and has to be learned: we propose a new error function to tackle this problem. This approach to the induction of asynchronous translators is discussed in connection with other approaches.
_____
ICNN97 Knowledge-based Methods in NN Session: SS9.2A Paper Number: 756 Oral
Possibilistic reasoning in a computational neural network
Andreas Kanstein, Marc Thomas and Karl Goser
Keywords: possibilistic reasoning computational neural network classification system fuzzy logic system
Abstract:
Possibilistic reasoning is implemented in a computational neural network for the formulation of a new classification system. The possibilistic classification is derived in analogy to the reasoning used in Bayesian classifiers. A principle of relational consistency is introduced to establish a connection of possibility and probability distributions. It is shown that possibilistic classification is suitable if distributions of very small classes like system failure data tend to be covered by distributions of large clusters. The classification system is also a paradigm for the implementation of a fuzzy logic system in a neural network architecture.
____
ICNN97 Knowledge-based Methods in NN Session: SS9.2B Paper Number: 757 Oral
Genetic algorithms for structural optimization, dynamic adaptation and automated design of fuzzy neural networks
Nikola K. Kasabov and Michael J. Watts
Keywords: genetic algorithms structural optimization dynamic adaptation fuzzy neural networks
Abstract:
_____
ICNN97 Knowledge-based Methods in NN Session: SS9.2C Paper Number: 760 Oral
An overview on supervised neural networks for structures
Alessandro Sperduti
Keywords: supervised neural networks structures hybrid systems
Abstract:
In recent years neural networks for the representation and processing of structures have been developed. These kind of networks are of paramount importance for the development of hybrid systems, since they allow the treatment of structured information very naturally and, in several cases, very efficiently. We present the basic concepts underpinning this class of networks and discuss computational and complexity issues.
_____
ICNN97 Knowledge-based Methods in NN Session: SS9.2D Paper Number: 755 Oral
BEXA: Set covering vs. neural network knowledge acquisition - a comparative review
Ian Cloete
Keywords: set covering neural network knowledge acquisition machine learning
Abstract:
Machine learning approaches to knowledge acquisition usually employ a symbolic method based on search, heuristically guided through the concept space to avoid the combinatorial explosion of possible concept descriptions to be examined. Neural networks, on the other hand, usually employ gradient based minimization of a cost function to acquire classificational knowledge. This paper presents a new symbolic set covering algorithm for rule induction, reviews five learning paradigms and compare that to knowledge acquisition by a neural network classifier.
Keywords: Symbolic learning Knowledge aquisition