By Topic

Neural Networks, 1996., IEEE International Conference on

Date 3-6 Jun 1996

Go

Filter Results

Displaying Results 1 - 25 of 76
  • K-subspaces and time-delay autoassociators for phoneme recognition

    Page(s): 1871 - 1876 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (496 KB)  

    This paper presents a new approach using time-delay autoassociators (TDAA) to perform phoneme recognition. The time-delay autoassociator combines the time-delay design for phoneme recognition and the technique of multilayer perceptron autoassociators. Each time-delay autoassociator is constructed and trained to model one and only one phoneme using data belonging to that phoneme category. This non-classification training procedure provides a method with high recognition performance to avoid the drawback encountered in most conventional speech recognition neural networks that the network output values do not represent candidate likelihoods. The approach with the proposed architecture, K-subspaces with linear time-delay autoassociators, in which each phoneme is modelled by K linear TDAAs, has yielded a high recognition performance compared to that of a time delay neural net and a shift-tolerant LVQ trained by classification learning procedures, over the three difficult phonemes “B”, “D” and “G”. It has also been observed that the nonlinear time-delay autoassociators could perform better than linear ones View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Compact modular neural networks in a hybrid speaker-independent speech recognition system

    Page(s): 1895 - 1899 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (316 KB)  

    In recent years, the computational effort for novel speech recognition systems has increased much more than the resulting recognition rates. Therefore, we present an approach for overcoming this drawback by using a modular phoneme recognition system which is included in a hybrid system with a discrete hidden Markov model (HMM). The development of a suitable topology for the modular architecture and the determination of relevant input parameters for the modules are the essential aspects of this paper. The main idea of the proposed system is the distribution of the complexity for the classification task on a set of modules with a higher degree of specialization. Therewith, acoustic-phonetic knowledge can be selectively incorporated because of the higher number of interfaces. Important system features are module-specific selection of input parameters according to the decision tasks and the utilization of time delay neural networks (TDNNs) as well as static neural networks without time processing. The intention is to improve the recognition rate for speaker-independent phoneme recognition and, at the same time, to reduce the necessary effort for simulating the system after the initial learning phase View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Massive neural video compression with temporal subsampling

    Page(s): 1963 - 1968 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (648 KB)  

    The large amounts of information involved in the transmission and storage of video data requires massive compression for efficient storage and transmission. However state-of-the-art video compression techniques do not achieve compression levels which are large enough to transmit video data on low band-width links. Thus compression levels need to be improved by a factor of 2 to 10 before they will become useful on the kinds of links which may be encountered in personal communications and cellular telephony. In this paper we present a new video compression technique which makes use of temporal subsampling and reconstruction of frames. When used in conjunction with our adaptive neural video compression (ANVC) technique, this new method leads to compression ratios as high as 500:1 for gray scale sequences with little loss of video quality View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Speech classification using a modified focused gamma network

    Page(s): 1877 - 1882 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (360 KB)  

    A modified version of the focused gamma network is proposed and applied to speech classification. Each input neuron of the network is equipped with a pair of gamma memories which have an adjustable time-scale parameter. In addition, each channel of speech features is equally divided into several segments, and each segment is clocked into an individual input neuron. With these modifications, the network is able to develop a proper time scale for each segment of patterns and handle speech patterns of different lengths as well View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Quasi-optimum detection results using a neural network

    Page(s): 1929 - 1932 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (244 KB)  

    We study some particularities for the application of a neural network to binary detection. Using a modeled input (the classical J. Marcum model for pulsed radar detection), we optimize the design of the network, evaluating its performance by Monte Carlo trials. After comparing the detection curves with the theoretical optimum ones, it is found that the number of pulses integrated for each detection is critical for a quasi-optimum performance of the neural network View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A neural network approach to nondestructive evaluation of complex structures, with application to highway bridges

    Page(s): 2154 - 2159 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (540 KB)  

    Many methods have been proposed for nondestructive evaluation (NDE) of structures such as highway bridges, skyscrapers, and pipelines. The analysis of acoustic emission (AE) signals produced during cracking in concrete or steel is a promising approach for nondestructive monitoring to detect degradation in the integrity of a structure. Because of their central role in the highway infrastructure, bridge analysis is a particularly important application area for NDE. We discuss the advantages and disadvantages of AE testing, and describe some of the difficulties in applying classical signal processing (deconvolution) techniques to AE analysis of a bridge. We present instead a neural network approach that has the potential to overcome many of these difficulties View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Improved time series segmentation using gated experts with simulated annealing

    Page(s): 1883 - 1888 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (340 KB)  

    Many real-world time series are multi-stationary, where the underlying data generating process switches between different stationary subprocesses, or modes of operation. An important problem in modeling such systems is to discover the underlying switching process, which entails identifying the number of subprocesses and the dynamics of each subprocess. For many time series, this problem is ill-defined, since there are often no obvious means to distinguish the different subprocesses. We discuss the use of nonlinear gated experts to perform the segmentation and system identification of the time series. Unlike standard gated experts methods, however, we modify the training algorithm to enhance the segmentation for high-noise problems where only a few experts are required View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Neural network commissioning of a PI controller for a rigidly coupled motor/mechanical system

    Page(s): 2049 - 2054 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (432 KB)  

    A neural network based tuning scheme for a motion control system is described. A continuous-time tuning rule is developed. This provides guaranteed system performance, based on both frequency domain (bandwidth) and time domain (overshoot) criteria. The neural net is trained entirely from simulation experiments and its pattern recognition capabilities are utilised to determine optimum controller gain values from experimental test data. It is found that the finite bandwidth of the current loop amplifier controlling the motor current can lead to undesirable effects on the demanded closed loop velocity performance. A shift factor is introduced to the neural net selected gains to counter the overshoot and bandwidth error introduced by the non-ideality of this loop. A nonlinear sampling technique is introduced which allows sufficiently accurate tuning over a larger workspace of parameter variation. The subsequent on-line performance of the neural net-tuned servo-system is tested through experimental results. The neural net topology and training algorithm are also detailed View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Stock market trend prediction using ARIMA-based neural networks

    Page(s): 2160 - 2165 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (304 KB)  

    We develop a prediction system useful in forecasting mid-term price trend in Taiwan stock market (Taiwan stock exchange weighted stock index, abbreviated as TSEWSI). The system is based on a recurrent neural network trained by using features extracted from ARIMA analyses. By differencing the raw data of the TSEWSI series and then examining the autocorrelation and partial autocorrelation function plots, the series can be identified as a nonlinear version of ARIMA(1,2,1). Neural networks trained by using second difference data are shown to give better predictions than otherwise trained by using raw data. During backpropagation training, in addition to the traditional error modification term, we also feedback the difference of two successive predictions in order to adjust the connection weights. Empirical results shows that the networks trained using 4-year weekly data is capable of predicting up to 6 weeks market trend with acceptable accuracy View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Real time recurrent neural networks for time series prediction and confidence estimation

    Page(s): 1889 - 1894 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (328 KB)  

    This paper explores two established techniques for doing time series modeling and prediction of mean and variance. The first method is an explicit method used to establish the embedding dimension of the time series, define the sensitivity of each variable and come up with a systematic decision of the delay of inputs for future prediction. The second method makes use of recurrent networks to implicitly derive models with “adaptive” time delays for the mean and variance predictions of a given time series. The recurrent system gives better prediction performance on artificial chaotic signals as well as real world exchange rate data in terms of mean squared error criterion and requires no laborious determination of the number of inputs View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • The automatic button-color matching system using Kohonen's self-organizing feature maps in the textile field

    Page(s): 2055 - 2059 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (260 KB)  

    We introduce the automatic button-color matching system using Kohonen's self-organizing feature maps (SOMs). The system consisted of two processes: (1) self-organizing feature mapping and (2) SOM analysis. The recognition test of the system was performed using an actual data set. From the results, the total recognition rate of 78% as almost the same as LVQ (80%) was obtained. Furthermore, ranking information of “best button”, “next best button”, ... can be obtained View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Adaptive system identification using multilayer neural networks and Gaussian potential function networks

    Page(s): 2261 - 2265 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (304 KB)  

    This paper deals with the characteristics of multilayer neural networks and radial basis function networks, and provides their hybridization by considering their advantages and disadvantages. The hybrid networks show their effectiveness in system identification as well as alleviate problems of error backpropagation algorithm such as local minima, slow speed, and size of structure by adopting other networks effectively. Potential performance improvement is demonstrated by computer simulation for two general problems of identification: static and dynamical system identification View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Reduction of neural network models for identification and control of nonlinear systems

    Page(s): 2250 - 2254 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (400 KB)  

    Structural learning is a proven pruning technique which induces decay of redundant weights. This paper introduces a method to significantly reduce the size of multilayer feedforward neural networks used as plant models and controllers. Initially oversized models are reduced during training thereby eliminating the need for a priori model order selection. A modification of structural learning is used to train the networks. Several examples nonlinear identification and control are presented. Order reduction can be performed both off-line and online. The reduced neural models and controllers lessen the computational load and thus benefit real time applications View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Stock market indices in Santiago de Chile: forecasting using neural networks

    Page(s): 2172 - 2175 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (248 KB)  

    Artificial neural networks (ANN) were used to predict the general index of share prices at the Santiago de Chile stock market. Time series with daily values of the index and of total amount of transactions were used to train the ANN. Input data was standardized and normalized shifting mean value to zero, variance to one and maximum values to one. A combined ANN produced better results than simple architectured ANN. The network managed TIS and I-delay memories in parallel. A time delay of ten labor days were sufficient to forecast. Results shows adequate performance of ANN in comparison with other methods used at INDECSA View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A neural network for grey level and color correction used in photofinishing

    Page(s): 2166 - 2171 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (244 KB)  

    The application of a multilayer perceptron for color and gray level correction in the field of photofinishing is presented. It is shown, that a neural network can improve the overall performance of a state of the art photo printer. The improved correction ability will reduce the number of unsalable pictures and thus lowers the production costs for the photo laboratory. The training experiments were carried out on a database of 30,000 photos using the MUSIC parallel supercomputer. The MUSIC system made it possible, for the first time, to process this large database in a reasonable time View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Spatio-temporal self-organizing feature maps

    Page(s): 1900 - 1905 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (448 KB)  

    Thus far, the success of capturing and classifying temporal information with neural networks has been limited. Our methodology adds a spatio-temporal coupling to the self-organizing feature map (SOFM) which creates temporally and spatially localized neighborhoods in the map. The spatio-temporal coupling is based on traveling waves of activity which start at each winning node and are naturally attenuated over time. When these traveling waves reinforce each other, temporal activity wavefronts are created which are then used to enhance a node's possibility of winning the next competition. The spatiotemporal coupling is easily implemented with only local connectivity and calculations. Once trained, the spatio-temporal SOFM can be used for detection or for partial pattern recall. The methodology gracefully handles time-warping and multiple patterns with overlapping input vectors View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Device-independent color correction for multimedia applications using neural networks and abductive modeling approaches

    Page(s): 2176 - 2181 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (468 KB)  

    Although color has appeal for developers and consumers alike, color reproduction poses a major problem in many computer based applications inducting multimedia and desktop publishing. The problem arises because of device independence of color, and the way each device processes color. In order to control the error in porting color, different traditional techniques have been applied. In this paper the utilization of artificial neural networks as well as abductive modelling approaches to color error reduction are introduced from an RGB (red/green/blue) color model perspective. Analysis of the results and ongoing research issues are discussed View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A neural architecture to predict pollution in industrial areas

    Page(s): 2107 - 2112 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (492 KB)  

    In this paper a novel approach, based on a neural network structure, is introduced in order to face the problem of pollutant estimation in an industrial area. In particular a short-term prediction (six hours ahead) of the SO2 pollutant mean value has been performed. A neural architecture, based essentially on a suitable number of MLPs devoted to predict alarm situations and to estimate the mean value of the pollutant, has been implemented. The strategy employed has been revealed to be particularly suitable, as it is shown in the results reported in the paper View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A neural demodulator for quadrature amplitude modulation signals

    Page(s): 1933 - 1938 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (340 KB)  

    A neural demodulator is proposed for quadrature amplitude modulation (QAM) signals. It has several important features compared with conventional linear methods. First, necessary functions for the QAM demodulation, including wide-band noise rejection, pulse waveform shaping, and decoding, can be embedded in a single neural network. This means that these functions are not separately designed but are unified in a learning process. Second, these functions can be self-organized through the learning. Supervised learning algorithms, such as the back-propagation algorithm, can be applied for this purpose. Finally, both wide-band noise rejection and a very sharp waveform response can be simultaneously achieved. It is very difficult to be done by linear filtering. Computer simulation demonstrates efficiency of the proposed method View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Learning rate functions in CMAC neural network based control for torque ripple reduction of switched reluctance motors

    Page(s): 2078 - 2083 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (412 KB)  

    This paper presents a novel approach to adapting the weights of a CMAC neural network-based controllers for torque ripple reduction in switched reluctance motors. The proposed method modifies the conventional LMS algorithm using a varying learning rate which, for the present application, is defined as a function of the rotor angle of the motor under control. Simulation results demonstrate that developing CMAC network based adaptive controllers following this approach affords lower torque ripple with high power efficiency, whilst offering rapid learning convergence in system adaptation View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Adaptive system identification by nonadaptively trained neural networks

    Page(s): 2066 - 2071 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (448 KB)  

    This paper proposes a novel adaptive neural system (ANS), which minimizes computation, focuses on learning about and adapting to the unknown environmental parameter, and eliminates (or reduces) poor local minima of the performance surface during the operation of the ANS. The idea is illustrated by its application to adaptive system identification. The adjustable weights of the ANS are divided into nonadaptively and adaptively adjustable weights. The former are determined by a nonadaptive training, using a priori information. Only the latter are adapted in operation. If they consist of linear weights of the ANS, the fast algorithms for adaptive linear filters are applicable for adaptation View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Robust design of flow control in ATM network switching by using Gaussian neural networks

    Page(s): 2194 - 2201 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (340 KB)  

    This paper deals with a design problem of flow patterns control at the level of asynchronous transfer mode (ATM) switching nodes. The bursty character of traffic patterns in ATM networks is modelled as multicommodity flows for multiple origin-destination networks and since the usual Markovian assumptions are unrealistic for ATM flow patterns, general distributions for the traffic streams and service times have to be considered. The Gaussian neural networks are used to approximate the dynamic map between the probability space of the general distributions (describing the flow clusters) and a finite set of parameters controlling the quality of service in the ATM network View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Two point training for complex plane transformation

    Page(s): 2266 - 2269 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (184 KB)  

    The complex neural network whose weights, threshold values, inputs and outputs signals are all complex variables is studied in this paper. It is shown that using a training set of only two points in the complex plane, the complex neural network can realize complex plane transformation for all points in the plane View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Application of artificial neural networks to effective bandwidth estimation in ATM networks

    Page(s): 1951 - 1956 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (444 KB)  

    A prime instrument for controlling congestion in ATM networks is admission control, which limits calls and guarantees a grade of service (GOS) determined by delay and loss probability in the multiplexer. It is essential for an admission control scheme to characterize, for a given GOS, the effective bandwidth requirement of the aggregate bandwidth usage of multiplexed connections. In this paper, an accurate and computationally efficient approach is proposed to estimate the effective bandwidth of multiplexed connections. In this method, a feedforward neural network is employed to model the complex relationship between the effective bandwidth and the traffic situations and a GOS measure. It is trained and tested via a large number of patterns generated by the accurate fluid flow model. Due to the neural network's adaptive learning, high computation rate and generalization features, this method can increase the link utilization and is suitable for real-time network traffic control applications View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A binary neural network approach for net assignment problems

    Page(s): 2188 - 2193 vol.4
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (412 KB)  

    This paper presents a binary neural network approach for the net assignment problem in the over-the-cell routing model in VLSI layout design. The goal of the problem for one cell row and two adjacent channels is to assign a subset of nets in the cell row such that the total channel density is minimized without violating the capacity constraint, the selection constraint, and the routing constraint. The neural network is composed of NM binary neurons for the N-net-M-track problem. Unlike the existing algorithm, our neural network finds an assignment of all the nets either in the cell row or in two channels simultaneously. The performance is verified through simulations in seven benchmark problems for channel routing. With the help of four heuristic methods, the neural network can find near-optimum solutions on the synchronous parallel computation View full abstract»

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