Showing papers in "Journal of intelligent systems in 1997"
TL;DR: Electroencephalogram signals from one subject who performed three mental tasks are classified by neural networks using a sixth-order autoregressive (AR) model of half-second windows of six-channel EEG, and a classification accuracy of 89% on test data is achieved.
Abstract: Electroencephalogram, or EEG, signals are an important source of information for the study of underlying brain processes. Such studies now provide a framework for the development of a new modality of humancomputer interaction based on EEG. Current research in this area only detects a small number of mental states. In this article, EEG from one subject who performed three mental tasks are classified by neural networks. Using a sixth-order autoregressive (AR) model of half-second windows of six-channel EEG, a classification accuracy of 89% on test data is achieved. A cross-validation study of a variety of neural network topologies showed that a network with one hidden layer of 20 units produced the best performance. It was also found that averaging the output of the network over consecutive inputs improved performance. K-means clustering of the resulting neural networks' weights identified key components of the AR representation.
32 citations
TL;DR: The segmentation algorithm is used as a preprocessing technique to p repare raw training data for use with an Artificial Neural Network and has been successfully tested on real world handwritten postal addresses.
Abstract: Artificial Neural Networks (ANNs) have been success fully applied for pattern recognition, speech recognition, control and other real world problems. This paper presents a method for segmentation of printed and difficult ha ndwritten postal addresses. The segmentation algorithm is used as a preprocessing technique to p repare raw training data for use with an Artificial Neural Network. The C programming langua ge, the SP2 supercomputer and a SUN workstation were used for the experiments. The algo rithm has been successfully tested on real world handwritten postal addresses. Some experiment al results are presented in this paper.
22 citations
TL;DR: This work defines a valuable extension to previous definitions of this quantity, and exploits it to produce significant system performance enhancements and is also amenable to exploitation of diversity through averaging multiple versions.
Abstract: Data-defined problems are not restricted to any particular problem domain, and are common. Data-defined problems, as the name suggests, are defined by a set of input-output mappings, and for the problems of particular interest the full details of how the inputs are related to the outputs are unknown. Such problems present the traditional programmer, whether using AI techniques or not, with a difficult task. This is because programming requires a prerequisite understanding of the mechanisms relating input and output such that an algorithmic solution can be devised. Automatic induction techniques demand no such prerequisite information. Neural computing is one such induction technique. And, moreover, it is one that can outperform more traditional AI induction techniques, such as IF-THEN rule systems. Neural computing can do better because it can follow a welldefined path to an optimal result, and because multiple, alternative versions can be cheaply generated to permit exploitation of certain properties of the resultant version set in particular, 'diversity'. We define a valuable extension to previous definitions of this quantity, and exploit it to produce significant system performance enhancements. In addition, because our neural computing is an approximating technique, it is also amenable to exploitation of diversity through averaging multiple versions. An example of letter recognition is used to illustrate these ideas.
15 citations
TL;DR: A neural network paradigm capable of exhibiting behavior that is useful for developing a variety of artificial potential fields for evaluating paths for autonomous robots is presented.
Abstract: The realization that some forms of artificial potential fields (APF) for path planning can be developed by parallel distributed means has prompted research efforts in applying neural networks to the problem of generating these fields. This paper presents a neural network paradigm capable of exhibiting behavior that is useful for developing a variety of artificial potential fields for evaluating paths for autonomous robots. The neural network is called the Wave Expansion Neural Network or WENN. The WENN develops a topologically ordered neural map of the robot's environment, specifying the obstacles and the target configuration. The WENN neuron activity distribution is used as an APF to evaluate paths for the robot. The results are supported by computer simulations to illustrate the performance of the network.
13 citations
TL;DR: It is concluded that ANN is a fast alternative to classical statistical techniques for prediction and modeling of experimental data and two popular weight-adaptation algorithms, RPROP and Delta-Bar-Delta rules are compared.
Abstract: Artificial Neural Networks (ANNs) have been widely advocated as tools for solving many decision modeling problems. In this paper, we use ANNs for the prediction of coronary artery disease. Real data from four major international medical organizations are used in the training and testing of the ANN algorithm. To speed up the training time, we implemented the algorithm in parallel on an Intel Paragon parallel computer. We have achieved an accuracy of > 76%, a comparable performance to probabilistic and statistical techniques. Furthermore, with parallel implementation, we achieve the accuracy in < 5 minutes of training time. Compared with the statistical approach, such savings in time are substantial. We conclude, therefore, that ANN is a fast alternative to classical statistical techniques for prediction and modeling of experimental data. Two popular weight-adaptation algorithms, RPROP and Delta-Bar-Delta rules are compared. The effect of network architecture and how to treat missing values for these two algorithms are also investigated. In general, RPROP is more robust and less affected by choice of architecture, order of data presentation, and effect of missing values. This research was supported by the National Science Foundation under Grant No. ECS-9407363.
12 citations
TL;DR: The validity of this multi-layer perceptron diagnosis system has been verified by a comparison which shows that the multilayer perceptron significantly out-performs a Κ nearest-neighbour classifier trained and tested using the same data sets.
Abstract: This paper describes the successful application of a multi-layer perceptron diagnosis system to the classification of suspect cases of Creutzfeldt-Jakob disease referred to the UK National Surveillance Unit. Accurate clinical diagnosis of Creutzfeldt-Jakob disease is required as part of the surveillance of this disease. At present the clinical diagnosis is determined by the presence or absence of particular neurological features. Using data taken from the clinical case notes of 36 cases and 36 non-cases of this disease the network has been trained to recognise the clinical characteristics of this disease. Using the clinical features taken from an additional 64 referrals to the unit the trained network was able to classify 59 of this group of cases and non-cases correctly. The validity of this method of classification has been verified by a comparison which shows that the multilayer perceptron significantly out-performs a Κ nearest-neighbour classifier trained and tested using the same data sets. Analysis of the contribution of the different inputs to the network has shed new light on the relative importance of different clinical features used in the diagnosis of this disorder.
7 citations
TL;DR: Performance comparison to time-delay neural networks shows a two-fold increase in performance for the PRANN, a new recurrent neural network topology for the prediction of time series.
Abstract: A new recurrent neural network topology for the prediction of time series is developed. The back-propagation algorithm to train this network is derived. Such a network is called the Prediction Recurrent Artificial Neural Network (PRANN). The performance of the PRANN network is analyzed for linear and nonlinear time series. Performance comparison to time-delay neural networks shows a two-fold increase in performance for the PRANN.
7 citations
TL;DR: This paper shows how expert knowledge in terms of fuzzy rules can be incorporated into the architecture of a neural network using a gradient descent learning algorithm which is an adaptation of backpropagation.
Abstract: The aim of this paper is to contribute to a central issue in neural networks, that of combining expert knowledge and observations (data) for learning. It is generally known that neural networks, as other adaptive models, have good learning and generalization capabilities because of their statistical consistency. However, such consistency is theoretically valid only for large size training sets. To enhance learning with small size sets, it is naturally desirable to incorporate additional knowledge (herein referred to as expert knowledge) in the architecture of the network to allow a task-oriented rather than just a generic problem-free architecture. In this paper we show how expert knowledge in terms of fuzzy rules can be incorporated into the architecture of a neural network. The parameters of the network are therefore adapted from the available data using a gradient descent learning algorithm which is an adaptation of backpropagation. The neuro-fuzzy •This work was supported by the EEC ESPRIT-BRA program under the project MIX-9119.
7 citations
TL;DR: Five artificial neural network approaches for detecting and diagnosing four common plasma-etch fault conditions are examined and the best accuracy achieved is approximately 98.7%correct fault-detection for the four fault types, 100% correct fault classification, and a 2.3% false alarm rate.
Abstract: The plasma-etch process is one of many steps in the fabrication of semiconductor wafers. Currently, faultdetection/diagnosis for this process is done primarily by visual inspection of graphically displayed process data. By observing these data, experienced technicians can detect and classify many types of faults. The tediousness and intrinsic human unreliability of this method, as well as the high cost of mistakes, makes automation attractive. In this paper, five artificial neural network approaches for detecting and diagnosing four common plasma-etch fault conditions are examined. The data used for training and testing the networks were collected during a 162 day period, in which over 46,000 wafers were etched. The best accuracy achieved in this study is approximately 98.7% correct fault-detection for the four fault types, 100% correct fault classification, and a 2.3% false alarm rate. The five neural-based approaches are described in detail, and results are given for each approach.
5 citations
TL;DR: This paper surveys published work in active learning research with the purpose of providing a unified understanding of the area and identifies three major recognised approaches to the implementation of active learning goal-driven learning, reinforcement learning and querying.
Abstract: This paper surveys published work in active learning research with the purpose of providing a unified understanding of the area. A passive learning system relies entirely on pre-gathered information, whereas an active learning algorithm has the capability of interacting with its environment in order to collect information and/or to select learning policy. Active learning systems produce improved generalisation, reduce data costs and are most useful where data is expensive and computation is cheap. There are three major recognised approaches to the implementation of active learning goal-driven learning, reinforcement learning and querying. While the first is largely a meta-level symbolic approach, the second is more a class of problems employing a policy-based approach to learning in non-deterministic dynamic environments; the third is based upon gathering the most useful examples by asking 'intelligent' questions. Research in the area is still mostly at a theoretical level.
5 citations
TL;DR: The potential of simple weightless neural networks for fast, inexpensive, processing of tomographic data is described and the first report to describe direct estimates of these parameters without recourse to the time consuming process of image reconstruction is described.
Abstract: The potential of simple weightless neural networks for fast, inexpensive, processing of tomographic data is described. Image reconstruction and parameter estimation for two-component distributions in an electrical capacitance tomography system have been considered. Results concentrate on simulated data from finite element analysis but, importantly, these have been verified using direct measurements on a tomographic rig. Topically, for image reconstruction, 95% of the binary pixels can be classified correctly for previously unseen measurements. It can be expected that 97% of the estimates of component ratio are within 10% of the actual value. Similarly flow regimes are correctly classified for 92% of patterns. This is the first report to describe direct estimates of these parameters without recourse to the time consuming process of image reconstruction. A particularly interesting result is that the WNN approach demands only modest accuracy from the measurements and this has significant and beneficial implications for noise immunity. Proposed hardware implementation requires less than 4 Mbits of RAM in all cases. This will enable 1000 images per second to be reconstructed and process parameters to be estimated for 10,000 frames per second.
TL;DR: Two practical solutions to the problem of generating ZVT matrices are outlined, the first employs a Hopfield-style neural network and the second employs a heuristic-driven tree search.
Abstract: Due to their sensitivity to cognitive impairment, simplicity of administration and low cost, path-following tests, such as the HalsteadReitan Trail Making Test (TMT) and the Zahlen-Verbindungs-Test (ZVT), are among the most frequently used neuropsychological tests in the assessment of brain damage. Despite their usefulness, however, such tests have been constructed in an unprincipled manner which does not facilitate repeated testing, systematic investigation or theoretical interpretation. The reason for their unprincipled development is that the selective generation of the required pathways constitutes a computational problem that is NPcomplete. This paper outlines two practical solutions to the problem of generating ZVT matrices. The first employs a Hopfield-style neural network and the second employs a heuristic-driven tree search. The characteristics of the pathways generated by these two solutions are compared, and ways of manipulating the relative difficulty of alternative test forms are discussed.
TL;DR: A new approach based on dynamic recurrent neural networks (DRNN) to identify the relationship between the muscle electromyographic activities and the arm kinematics during the drawing of the figure eight using an extended arm and its generalization ability to draw unlearned movements is demonstrated.
Abstract: We propose a new approach based on dynamic recurrent neural networks (DRNN) to identify, in humans, the relationship between the muscle electromyographic (EMG) activities and the arm kinematics during the drawing of the figure eight using an extended arm. After learning, the DRNN simulations showed the efficiency of the model. We demonstrated its generalization ability to draw unlearned movements. We developed a test of its physiological plausibility by computing the error velocity vectors when small artificial lesions in the EMG signals were created. These lesion experiments demonstrated that the DRNN has identified the preferential direction of the physiological action of the studied muscles. The network also identified neural constraints such as the covariation between geometrical and kinematics parameters of the movement. This suggests that the information of raw EMG signals is largely representative of the kinematics stored in the central motor pattern. Moreover, the DRNN approach will allow one to dissociate the feedforward command (central motor pattern) and the feedback effects from muscles, skin and joints.
TL;DR: This neural based rainfall nowcasting system was capable of providing a reliable rain storm warning signal to the Hong Kong public in advance and was based on a recurrent Sigma-Pi network.
Abstract: This paper describes the development of a new approach of rainfall nowcasting (very short term forecasting) by using a neural network. This approach consisted of extracting the information from radar images and evaluating past rain gauge records to provide short term rainfall forecasting. All meteorology data were provided by the Royal Observatory of Hong Kong (ROHK). Pre-processing procedures were essential for this neural network rainfall nowcasting. The forecast of rainfall every half an hour is such that a storm warning signal can be delivered to the public in advance. The network architecture was based on a recurrent Sigma-Pi network. The results are very promising and this neural based rainfall nowcasting system was capable of providing a reliable rain storm warning signal to the Hong Kong public in advance.
TL;DR: The use of artificial neural networks to explore/generate dynamic form dance performance as a creative search for patterns of harmony which are not fixed but continuously remapped as the performance dynamically organises itself is discussed.
Abstract: Applications of artificial intelligence and other new computing technologies have hitherto been focused on engineering, scientific and business fields. This paper describes the novel application of the latest computing technologies to the often neglected arts field, in particular, dynamic form dance performance. Dynamic form dance performance can be seen as a creative search for patterns of harmony which are not fixed but continuously remapped as the performance dynamically organises itself. This paper discusses the use of artificial neural networks to explore/generate
TL;DR: A performance comparison of a neural network based on adaptive resonance theory (ART) against another network implemented with the backpropagation algorithm is presented.
Abstract: Today in the field of artificial intelligence, neural networks are being used to solve certain kinds of pattern recognition and pattern matching problems for which traditional artificial intelligence techniques have been unable to provide an adequate solution. In this paper, we present a performance comparison of a neural network based on adaptive resonance theory (ART) against another network implemented with the backpropagation algorithm. The backpropagation training algorithm attempts to minimize the difference between an input vector and its specified target vector, while ART models are generalizations of competitive learning
TL;DR: This article shows how pitch recognition and timbre discrimination for a string instrument are implemented using artificial neural networks, illustrating how related but increasingly difficult tasks are handled by networks of increasing complexity.
Abstract: The task of real time discrimination between greatly similar signals is an important and common one. In this article we show how pitch recognition and timbre discrimination for a string instrument are implemented using artificial neural networks. Pitch recognition, an easier task, is realized with a linear classifier, while timbre discrimination is achieved with a Multiple Layer Perceptron using gradient Back Propagation learning. This illustrates how related but increasingly difficult tasks are handled by networks of increasing complexity.