By Topic

Artificial Neural Networks, 1991., Second International Conference on

Date 18-20 Nov 1991

Filter Results

Displaying Results 1 - 25 of 82
  • A back propagation network as a decision aid in flexible welding system design

    Page(s): 271 - 275
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (304 KB)  

    A neural network has been applied to a new classification and coding system, a sub-set of Group Technology. The new code matches welding processing requirements of components with the features of a flexible welding cell. This work forms part of the development of a new method of flexible manufacturing systems design. The emphasis of the paper is coding the inputs and outputs of the network, and the effects of varying network parameters, during training, on the performance of the network View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Further developments of a neural network speech fundamental period estimation algorithm

    Page(s): 340 - 344
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (300 KB)  

    This work describes a speech fundamental period estimation algorithm that estimates the time of excitation of the vocal tract using a pattern classifier, the multi-layer perceptron (MLP). The pattern classifier was trained using speech semi-automatically labelled by means of an algorithm that makes use of the output from a Laryngograph. Various issues arising in the training of the system were explored. Three basic configurations of the system were compared using different pre-processing strategies. lt was found that processing the sampled speech time-waveform directly with the pattern classifier gave better results than using one of two filterbanks. The performance of the algorithm was evaluated against that of a simple peak-picking algorithm and the well known cepstrum algorithm using quantitative frequency contour comparisons. The performance of the new algorithm on a difficult set of test data was shown to be better than the peak-picker and comparable to the cepstrum algorithm. The advantage of the scheme is that fundamental period estimates are made on a period-by-period basis, thus preserving the irregularity in the speech excitation that is lost by techniques that produce an average period estimate View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • EEG analysis using self-organisation

    Page(s): 210 - 213
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (216 KB)  

    The electro-encephalogram (EEG) has formed the basis for the classification of sleep into several stages. The authors propose a method of sleep analysis which requires no pre-defined application of rules, and aims to give some indication of the dynamics of sleep in humans. The authors show that the use of a self-organising feature map has enabled clustering of feature vectors in a high dimensional space, from a highly complex signal, about which little prior knowledge is known. They also demonstrated that the transition trajectories between the main cluster sites are representative of three competing dynamic processes which govern the gross structure of the EEG during sleep. They are in a position to apply this method to clinical situations for which it has hitherto been impossible to analyse the sleep EEG View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Complexity reduction in Volterra connectionist networks using a self-structuring LMS algorithm

    Page(s): 44 - 48
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (272 KB)  

    This paper describes the development of an algorithm for structure optimisation in linear weight neural networks which although maintaining a unimodal error surface adaptively optimises network structure. The methods developed may be applied to any network which is linear in its weights, for example the radial basis function (RBF) networks and Volterra networks. These linear weight networks (LWNs) are important as their error surfaces are unimodal allowing high speed single run learning. By use of the optimal output mapper they may also be shown to have lighter computational loads in general than hidden layer back propagation (HLBP) networks View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Invariant digit recognition by Zernike moments and third-order neural networks

    Page(s): 82 - 85
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (204 KB)  

    The classification of hand-written digits with invariance under translations, rotations and scaling using neural networks is discussed. Two approaches are considered. First, Zernike moment expansions are used to produce invariant representations of the image. Secondly, the image is coded using triplets of pixels grouped into similarity classes of triangles. Both types of coding form the input into a multi-layered perceptron classifier. Methods of reducing the dimensionality of the ensuing image representations are discussed, and the performances of both coding methods are assessed and compared. Third-order networks result in a generalisation success rate of 79% under all transformations combined View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Second International Conference on Artificial Neural Networks (Conf. Publ. No.349)

    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (24 KB)  

    The following topics were dealt with: artificial neural networks theory; implementations; images; engineering applications; dynamical systems; control and robotics; hybrids; speech and natural language; medical applications; and character recognition View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Logical neural nets and distributed implementations of weighted regular languages

    Page(s): 158 - 162
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (296 KB)  

    A logical neural network, Aleksander (1), is a finite state machine then it is only possible to recognise regular grammars with these networks. When extra memory is associated with the nodes of these networks, the computational power of the model is increased and now weighted regular grammars, Salomaa (14), can be recognised. Through a constructive method based on the complexity of the production rules of the grammar, a logical network can be built to recognise any weighted regular language. The network generated by the constructive method is a distributed implementation of the weighted regular language View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • An analysis of self-organising networks based on goal-seeking neurons

    Page(s): 257 - 261
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (272 KB)  

    The principle involved in applying self-organising architectures to pattern recognition problems is that patterns which share common features are clustered together, with each cluster representing one and only one class. One architecture that follows such a principle is the GSN self-organising architecture (GSN8) a Boolean neural network proposed by Filho, Fairhurst and Bisset (1990, 1991). In creating new clusters to store unfamiliar patterns and grouping similar patterns in the same cluster, the GSN8 exhibits the typical behaviour of a self-organising architecture. The authors present an analysis of the computational mechanisms of the GSN8 architecture, and modifications will be suggested with the aim of improving its performance in practical processing tasks View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Self-supervised training of hierarchical vector quantisers

    Page(s): 5 - 9
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (316 KB)  

    The author has previously developed a hierarchical vector quantisation (VQ) model which successfully applied to time series and image compression respectively. The paper derives an extension to this model, in which the author backpropagates signals from higher to lower layers of the hierarchy to self-supervise the training of the VQ. He reviews the basic properties of his VQ model and its relationship to neural network methods. He extends the model to an ensemble of VQs, and derives its properties in the limit of a large codebook size (i.e. the continuum limit). Finally, he demonstrates how self-supervision emerges naturally in this type of model View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Encoding temporal structure in probabilistic RAM nets

    Page(s): 369 - 372
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (272 KB)  

    The authors show how both gradient descent and reinforcement training can be used to teach a recurrent probabilistic RAM (pRAM) net to classify binary strings of arbitrary length, and how a form of reinforcement training can be used with `leaky integrator' pRAM nodes to enable them to store complex temporal sequences View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Realtime tracking and identification of multiple objects in a cluttered environment

    Page(s): 74 - 78
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (348 KB)  

    The authors have developed a distributed multi-module system for realtime tracking of objects, and online training of neural network classifiers. The aim of the project was to build a system which is able to locate objects and to follow them and learn to recognise them with minimal supervision. The problem domain chosen was the tracking and recognition of three species of tropical fish swimming in an aquarium View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Motion perception and recognition using moving light displays

    Page(s): 91 - 94
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (192 KB)  

    Despite the limited amount of information that is available to the observer of a moving light display, humans are able to infer detailed information about the activity and identity of the actors within the film. The development of a system, which uses such displays to recognise different types of human motion, is described using a spatio-temporal neural network to classify different temporal sequences of movement. ART architectures are employed to adapt and maintain the learned codes of the network providing real-time learning and stability against insignificant events View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A high order feedback net (HOFNET) with variable non-linearity

    Page(s): 59 - 63
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (196 KB)  

    Most neural networks proposed for pattern recognition sample the incoming image at one instant and then analyse it. This means that the data to be analysed is limited to that containing the noise present at one instant. Time independent noise is therefore, captured but only one sample of time dependent noise is included in the analysis. If however, the incoming image is sampled at several instants, or continuously, then in the subsequent analysis the time dependent noise can be averaged out. This, of course, assumes that sufficient samples can be taken before the object being imaged, has moved an appreciable distance in the field of view. High speed sampling requires parallel image input and is most conveniently carried out by optoelectronic neural network image analysis systems. Optical technology is particularly good at performing certain operations, such as Fourier Transforms, correlations and convolutions while others such as subtraction are difficult. So for an optical net it is best to choose an architecture based on convenient operations such as the high order neural networks View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Nonlinear time series prediction

    Page(s): 354 - 358
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (196 KB)  

    There is considerable interest in the use of nonlinear techniques to perform prediction of naturally occurring time series, e.g. medical signals and signals from seismic returns. The motivation in considering these techniques lies in the fact that may of the underlying generation mechanisms are nonlinear. There has been growing interest in the use of neural network architectures for such applications. This has been coupled with the research carried out in the field of nonlinear dynamical systems, especially so-called chaotic systems. This paper reports an investigation of the prediction of a chaotic time series arising from a well-known differential equation. A series approach is applied to the estimation of such a series as well as the more regular and well-understood periodic case. First of all, the radial basis function technique is applied, using a approach based on the Wiener theory. A brief comparison is made with the same sort of approach applied to the Volterra series predictor (Nisbet et al., 1991) View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • The application of neural networks to cognitive phonetic modelling

    Page(s): 280 - 284
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (176 KB)  

    A neural network is used to generate control parameters for a parallel formant speech synthesizer, corresponding to a sequence of allophonic tokens. Training is to be accomplished using formant data obtained from both natural and synthetic speech. It is intended that theories of cognitive phonetics, currently being developed in the Department of Language and Linguistics at the University of Essex will be used in order to improve the modelling of coarticulation View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Applications of artificial neural networks to reverse software engineering

    Page(s): 163 - 169
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (468 KB)  

    In this paper a system has been described that has been used to explore the application of ANN based technology to reverse software engineering (RSE). The demonstrator system has been trained to identify complex sort algorithms within large `real world' COBOL source listings. The ease with which the system can be extended, by training further neural networks to identify new algorithmic structures, offers the possibility of rapidly extending and customising the functionality of such a system. This research also suggests the potential for constructing a larger system that utilises recursively applied neural networks to automatically generate abstract descriptions of a source program View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Isolated word recognition with the radial basis function classifier

    Page(s): 345 - 349
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (288 KB)  

    The paper introduces a method of optimising the radial basis function classifier representing a compromise between fast heuristic approaches and a fully adaptive approach. The method is especially suitable when the size and complexity of the classification problem are such that large numbers of kernel functions are required to maximise generalisation (minimise error rate on test data). The problem examined is the speaker-independent recognition of isolated utterances of the letters of the alphabet, which is a difficult and useful task. Results are presented to illustrate the optimisation process, following which the performance of the optimised classifier is compared with several other neural and traditional techniques View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Neural classification of chest pain symptoms: a comparative study

    Page(s): 200 - 204
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (320 KB)  

    The authors demonstrate the effectiveness of neural networks in the diagnosis of heart attacks (acute myocardial infarction). Two neural network classifiers are compared. The multi-layered Perceptron is found to perform well but the probabilistic interpretation of its output is not well defined. The Boltzmann Perceptron Classifier is found to have comparable performance and has the advantage that it estimates a posteriori class probabilities directly, making no a priori assumptions about the class probabilities View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Automatic signalized point recognition with feed-forward neural network

    Page(s): 359 - 363
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (500 KB)  

    The recognition and accurate location of specific patterns, such as of special targets or signalized points in digital images, is an important step in photogrammetric measurement procedures. This paper explores the capability of the feed-forward neural network using a version of back-propagation training for the recognition of targets that appear in digitized images of aerial photographs. These targets commonly appear with differing orientations, backgrounds, scales, and suffer from varying shape distortions. Thus, for the network to establish an appropriate representation it must be trained with a very large number of cases that adequately reflect the variations of the target and non-target patterns. In order to eliminate redundancy and minimize the size of the training set, an iterative training scheme for the selection of such a set was developed. After two iterations of training promising results were reached View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Universal architectures for logical neural nets

    Page(s): 262 - 266
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (284 KB)  

    A universal architecture of logical neural nets is proposed which includes the conventional N-tuple and pyramid architectures as its extremes. The discrimination function of the architecture can be adjusted conveniently via the structure parameters and proper spreading operation. This flexibility enables tailored discriminator design in a practical environment. The technique of spreading with a multilayered net is studied and the authors conclude that spreading is only efficient for the first layer. The discussion in this report is concentrated on understanding the different aspects of the neural net as a discriminator. A practical classifier containing several discriminators can be easily designed based on the results here and some global considerations View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A direct control method for a class of nonlinear systems using neural networks

    Page(s): 134 - 138
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (304 KB)  

    Presents a direct control scheme for a class of continuous time nonlinear systems that are linear in the control variable. The objective of control is to track a desired reference signal. This type of system is encountered in many applications, e.g. rigid link robot manipulator control. The advantage of restricting attention to these systems is that the control theory of these systems is well developed. The control method proposed here does not assume knowledge of plant nonlinearities View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Adaptive equalization using the lp back propagation algorithm

    Page(s): 10 - 13
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (164 KB)  

    This paper discusses the performance of adaptive equalization using lp, 1<p⩽2, back propagation algorithm. The results indicate that as p decreases, the convergence time tends to reduce roughly linearly. Considerable improvement in the rate of convergence and bit error rate performance for 1<p<2 over p=2, has been shown to be feasible View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Experience in using neural networks for electronic diagnosis

    Page(s): 115 - 118
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (244 KB)  

    British Telecommunication plc (BT) has an interest in developing fast, efficient diagnostic systems especially for high volume circuit boards as found in today's digital telephone exchanges. Previous work to produce a diagnostic system for line cards has shown that a model-based, expert system shell can be most beneficial in assisting in the diagnosis and subsequent repair of these complex, mixed-signal cards. Expert systems, however successful, can take a long time to develop in terms of knowledge acquisition, model building and rule development. The re-emergence of neural networks stimulated the authors to develop a system that would diagnose common faults found on line cards by training a network using historical test data View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • On the iterative inversion of RBF networks: a statistical interpretation

    Page(s): 29 - 33
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (316 KB)  

    The method of the inversion of arbitrary continuous multilayer networks is extended to the class of radial basis function networks. The problem centres around how to obtain an input pattern (or set of patterns) which has an associated network output which is close (in a least squares sense) to a desired target pattern. Why this nonlinear inverse problem is of interest is discussed and the action of inverting a radial basis function network is motivated from the point of view of statistical semiparametric density function estimation View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Error recovery behaviour of feedback RAM-networks

    Page(s): 304 - 308
    Save to Project icon | Click to expandQuick Abstract | PDF file iconPDF (252 KB)  

    Presents an analysis of the dynamics involved in the process of error recovery in Boolean neural nets with feedback loops. The main objective of the work is to show the conditions under which a randomly generated RAM-network recovers from input and state errors. RAM-nets of the type mentioned tend to have some inherent stability with respect to its input sequence. The approach adopted here is a probabilistic one based on the theory of discrete-time stationary Markov chains though the net is in fact deterministic. The reason for adopting such approach is due to the fact that the RAM nodes connections and logical functions are randomly selected in order to make the analysis as general as possible. Also, besides the vast number of states the net can enter, both the encoding of the input patterns and the initial state of the net are chosen at random View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.