(Return to ICNN97 Homepage) (Return to ICNN'97 Agenda)



ICNN'97 FINAL ABSTRACTS


TS: SPEECH PROCESSING, TIME SERIES & FILTERING


ICNN97 Speech Processing, Time Series & Filtering Session: TS1A Paper Number: 559 Oral

A hierarchical Bayes approach to nonlinear time series prediction with neural nets

T. Matsumoto, H. Hamagishi and Y. Chonan

Keywords: hierarchical Bayes nonlinear time series prediction neural networks

Abstract:

An attempt is made to solve a class of nonlinear time series prediction problems with a hierarchical Bayes approach using neural nets.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS1B Paper Number: 467 Oral

Sequential network construction for time series prediction

Tomasz J. Cholewo and Jacek M. Zurada

Keywords: Time series prediction Nonlinear Cross Validation Sequential network

Abstract:

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS1C Paper Number: 356 Oral

Markov Gated Experts for time series analysis beyond regression

Shanming Shi and Andreas S. Weigend

Keywords: Time series Regression Markovian architecture

Abstract:

Most traditional time series models are based on local methods (in time), which means assuming that the time series can be fully and locally (in time) characterized with a finite embedding space. There are many situations in which simple regression can not help find the temporal structural in time series. In this research, a Markovian architecture: Markov Gated Experts, has been developed based on nonlinearly gated experts. This paper discusses the statistical framework and compares the performance of Markov gated experts to gated experts on both computer generated time series and real world data. Compared with the original method, Markov gated experts are more powerful in finding the underlying temporal structure, and are therefore a more powerful analytical and forecasting model for non-stationary and structurally changing time series.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS1D Paper Number: 650 Oral

Improving the accuracy of financial time series prediction using ensemble networks and high order statistics

Roy Schwarzel and Bruce Rosen

Keywords: time series analysis ensemble networks higher order statistics backpropagation

Abstract:

We apply neural network ensembles to the task of forecasting financial time series and explore the use of higher order statistical information as part of network inputs. We show that the prediction accuracy of the time series can be significanlty improved utilizing this methodology.

Since prediction accuracy is only an estimate for the profitability on the financial market, we report good and profitable results using a profit/loss metric based on market simulations.

Our simulations show an improvement of between 1.3 and 12.4% over a simple buy and hold trading strategy, and an improvement of between 6.5 and 20.9% over trading strategy using linear autoregressive models.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS1E Paper Number: 3 Oral

Neural networks based traffic prediction for cell discarding policy

Hsiou-Ping Lin and Yen-Chieh Ouyang

Keywords: Neural networks traffic prediction cell discarding policy

Abstract:

Traditional cell discarding policies have some limitations. They are either difficult to implement or lack of flexibility. In this paper, we proposed a new cell discarding policy that is based on the traffic load prediction by time-delayed neural networks. We use the finite-duration impulse response (FIR) filter in the multilayer neural networks to determine which cells will be discarded when the network buffer is going to overflow. The simulation uses ten different sources to generate cells according to their respective characteristic. The number of learning iterations, the normalized squared sum prediction error of the multilayer neural network are measured. The goodput is used to evaluate the performance of the proposed cell discarding policy. From the simulation result, the proposed cell discarding policy can achieve high goodput value that is near optimal.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS2A Paper Number: 184 Oral

A neural network solver for basic pursuit and its application to time-frequency analysis of biomedical signals

Z.S. Wang, Y.S. Xia, W.H. Li, Z. Y. He and J. D.Z. Chen

Keywords: neural network basic pursuit time-frequency analysis of biomedical signals

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS2B Paper Number: 348 Oral

Broadband noise cancellation using a functional link ANN based nonlinear filter

Ganapati Panda and Taposhi Chatterjee

Keywords: Broadband noise cancellation functional link ANN nonlinear filter

Abstract:

In many signal processing applications sharp edged signals are encountered which are contaminated with broadband noise. It is often required to recover these signals by suitable filtering technique. In image, speech signal processing and telecommunication, the linear filter performs poorly when signal having an envelope of steep edges is corrupted by random noise. Further, when these signals are processed, removal of additive random noise requires a higher order digital linear filter which involves more complexity and hence, higher cost. In a recent paper a nonlinear delayed N-path adaptive FIR (NDNAFIR) digital filter has been proposed which outperforms the linear adaptive FIR filter and the nonlinear median filter in canceling broadband noise. But it is observed that the NDNAFIR filter is quite complex in structure as well as its noise rejection capability is not commendable. In this paper, we have proposed functional link artificial neural network based nonlinear adaptive filter (FLANAF) structure with two schemes of learning. In the first the delayed version of the input signal is used whereas in the second the noise free desired signal is used for training. It is shown that in all compartments FLANAF is superior to those of the previously reported structures.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS2C Paper Number: 399 Oral

Estimation of unmeasured inputs using recurrent neural networks and the extended Kalman filter

Ressom Habtom and Lothar Litz

Keywords: Recurrent neural networks Kalman Filter estimation

Abstract:

A multiple-input-multiple-output dynamic system, whose some inputs could not be measured on-line but data for the training of a neural network could be made available, is considered. A method of estimating those on-line immeasurable inputs is proposed based on a recurrent neural network model of the system and using the extended Kalman filter (EKF). A stirred tank heater (a laboratory experimental setup) and a simulation model of a drying process are employed to demonstrate the effectiveness of the method. The two systems are modeled using recurrent neural networks and the estimation of an unmeasured valve position in the case of the stirred tank heater and an unmeasured moisture content for the drying process is carried out using the EKF algorithm. The experimental and the simulation results substantiate the practical effects of the proposed method for a class of dynamic systems.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS2D Paper Number: 471 Oral

Adaptive target detection in UWB images using recurrent networks

Li-Kang Yen, Jose C. Principe, John Fisher III

Keywords: Target detection Recurrent networks UWB images SAR images

Abstract:

The new ultra wide band (UWB) synthetic aperture radar (SAR) promises foliage penetration capabilities due to the different phenomenology of the reflection of the target in UWB SAR. Therefore, the automated target detection algorithms must be redesigned to take advantage of the new phenomenology. Besides, the new algorithms should also work robustly in the nonstationary environment in UWB SAR. In this paper, a new adaptive GLRT is proposed, which can detect the metallic object robustly using the resonance response in the UWB SAR scenario. The adaptive GLRT is formulated based on the linear transform of the resonance response. For practical on-line applications, the applied linear transform must be chosen to achieve a better representation of the signal without too much complexity. In this case, Laguerre recurrent networks is proposed to implement the linear transform,so that the on-line GLRT become feasible.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS2E Paper Number: 480 Oral

Multiple target tracking using recurrent neural networks

Gilles P. Mauroy, Edward W. Kamen

Keywords: Multiple target tracking Kalman filter Set estimation recurrent Neural network

Abstract:

Multiple Target Tracking (MTT) encounters the Data Association Problem (DAP) when the target-measurement associations are uncertain because of the measurement noises, the targets' proximity, and the initial condition uncertainty.

Standard approaches to MTT rely upon evaluations of association probabilities between targets and measurements whereas the SME filter developed by Kamen relies upon the choice of particular symmetric measurements and the Extended Kalman Filter (EKF) as a nonlinear filter. This paper centers on improvements of this latter strategy by using recurrent neural networks instead of the EKF.

We argue that too much uncertainty in the initial condition prevents even the optimal filter from having an acceptable performance. To overcome this problem, we use the concept of set estimation and present comparative performances of several strategies.

Key words: Set estimation, recurrent neural network, data association problem, multiple target tracking, optimal nonlinear filter.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS2F Paper Number: 605 Oral

Performance of two neural receiver structures in the presence of co-channel interference

Kimmo Raivio, Ari Hämäläinen , Jukka Henriksson and Olli Simula

Keywords: neural receiver structures co-channel interference self-organizing map RBF

Abstract:

Real communication channels with multipath propagation, interference and possible nonlinearities pose a difficult problem to the detecting receiver. This paper deals with neural approaches to solve those difficulties. Two types of neural networks, self-organizing map and radial basis functions have been studied. The results show that, while there are no actual benefits in using neural receivers in simple white noise Gaussian channels, the performance in nonlinear channels is much better with these new approaches than with the traditional ones.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS3A Paper Number: 25 Oral

Speaker verification and identification using gamma neural networks

Chuan Wang, Dongxin Xu and Jose C. Principe

Keywords: Speaker verification Speaker identification gamma neural networks

Abstract:

Gamma neural networks are used for speaker verification and identification in this paper. When input features are cepstral coefficients, the memory depth of the gamma networks can be adjusted on-line and the outputs of the gamma memory may be viewed as the combination of cepstral and delta cepstral coefficients with adaptable weights. So, gamma networks are very suitable to grasp the dynamics of speech. Simulation results show that the gamma networks outperform other neural approaches for speaker identification and verification in TIMIT database experiments.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS3B Paper Number: 586 Oral

Context modeling in a hybrid HMM neural net speech recognition system

Horacio Franco, Mitch Weintraub and Michael Cohen

Keywords: Context modeling hybrid HMM neural net speech recognition

Abstract:

We compare two methods for modeling context in the framework of a hybrid hidden Markov model (HMM) / multi layer perceptron (MLP) speaker-independent continuous speech recognition system. The first method for modeling context is based on the computation of HMM context-dependent observation probabilities using a Bayesian factorization in terms of scaled posterior phone probabilities that are computed with a set of MLPs, one for every relevant context. The second way of context modeling is based on the use of input features composed of extended multiframe windows of acoustic vectors that include the acoustic information of the current phone as well as various degrees of the acoustic information of the adjacent left and right phones.

Experimental results using a hybrid HMM-MLP speaker-independent continuous speech recognition system show that the first approach, based on connectionist context-dependent estimation of observation probabilities, is more efficient in the use of parameters for the same level of recognition performance.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS3C Paper Number: 558 Oral

Parallel neural networks for speech recognition

Byoung Jik Lee

Keywords: Parallel neural networks speech recognition confidence machine learning

Abstract:

This paper presents the PNNC (Parallel Neural Networks by Confidence) and PNNS (Parallel Neural Networks by Success/Failure), which generate and integrate parallel neural networks to achieve high performance on the test problem of letter recognition from string of phonemes. Our approach provides a way to create subproblems for a complex problem by partitioning the data, thus each neural network adapts to each subproblem more efficiently. Each neural network is iteratively trained on the training data which the previous neural networks could not guarantee or produce proper results. Each network works by filtering out unsatisfactory instances to pass to the next sub network to handle. This approach provides a way, by exploring different search spaces, to handle the local minima problem without complex computations via the use of neural networks working in parallel. Experimental results show that our approach achieves improvement over the general multi-layered neural network on the speech recognition problem which converts strings of phonemes to strings of letters taken from the 934 most commonly used English words. Particularly, the fact that error on vowel recognition is reduced significantly by the subsequent neural network indicates the usefulness and effectiveness of our approach.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS3D Paper Number: 366 Oral

Text-independent speaker identification using a hybrid neural network and conformity approach

Atanas P. Ouzounov

Keywords: Kohonen Map speaker identification hybrid Network

Abstract:

A hybrid neural network consisting of a Kohonen map and a multilayer perceptron is described. This hybrid network is designed for the text-independent speaker identification. Each speaker known to the system has a personalized hybrid neural network. In the hybrid network the multilayer perceptron works in a histogram feature space created by the Kohonen map. The map is partitioned into a set of overlapping fields. The modified field histograms are obtained and they are directly fed to the hidden layer neurons. To obtain the histograms the input vectors in supra-segments are used. The hybrid system is evaluated with a 10 speakers database of telephone speech. The experimental results have shown the validity of this kind of cooperation of a Kohonen map and a multilayer perceptron.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS3E Paper Number: 577 Oral

Speaker identification based on a modified Kohonen network

Karina Vieira, Bogdan Wilamowski and Robert Kubichek

Keywords: Speaker identification modified Kohonen network human information processing

Abstract:

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS3F Paper Number: 397 Oral

Continuous Speech recognition with neural networks and stationary-transitional acoustic Units

Roberto Gemello, Darion Albesano, Franco Mana

Keywords: Markov model Speech recognition stationary-Transitional Acoustic units

Abstract:

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS4A Paper Number: 515 Oral

An efficient stock market forecasting model using neural networks

Amir Atiya, Noha Talaat and Samir Shaheen

Keywords: stock market forecasting neural networks financial markets

Abstract:

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS4B Paper Number: 333 Oral

Prediction of system marginal price in the UK power pool using neural networks

A Wang and B Ramsay

Keywords: system marginal price UK power pool neural networks

There is an increasing interest in the prediction of System Marginal Price (SMP) in the Power Pool since electricity industry Vesting in England and Wales in 1990. The prediction of SMP improves the financial performance of an Independent Power Produc (incomplete abstract received)

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS4C Paper Number: 301 Oral

Application of artificial neural networks in sales forcasting

Devil H.F. Yip, E.L. Hines and William W.H. Yu

Keywords: sales forcasting artificial neural networks backpropagation genetic algorithms

Abstract:

The aim of the work presented in this paper is to forecast sales volumes as accurately as possible and as far into the future as possible. The choice of network topology was Silva's adaptive back-propagation algorithm and the network architectures were selected by Genetic Algorithms (GAs). The networks were trained to forecast from 1 month to 6 months in advance and the performance of the network was tested after training. The test results of artificial neural networks (ANNs) are compared with the time series smoothing methods of forecasting using several measures of accuracy. The outcome of the comparison proved that the ANNs generally perform better than the time series smoothing methods of forecasting. Further recommendations resulting from this paper are presented after the conclusions._____

ICNN97 Speech Processing, Time Series & Filtering Session: TS4D Paper Number: 171 Oral

Sales forecasting using neural networks

Frank M. Thiesing and Oliver Vornberger

Keywords: Sales forecasting neural networks backpropagation

Abstract:

Neural networks trained with the back-propagation algorithm are applied to predict the future values of time series that consist of the weekly demand on items in a supermarket. The influencing indicators of prices, advertising campaigns and holidays are taken into consideration. The design and implementation of a neural network forecasting system is described that has been developed as a prototype for the headquarters of a German supermarket company to support the management in the process of determining the expected sale figures.

The performance of the networks is evaluated by comparing them to two prediction techniques used in the supermarket now. The comparison shows that neural nets outperform the conventional techniques with regard to the prediction quality.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS5A Paper Number: 259 Oral

Blind source separation of real world signals

Te-Won Lee and Reinhold Orglmeister

Keywords: Blind source separation real world signals FIR polynomial filter

Abstract:

We present a method to separate and deconvolve sources which have been recorded in real environments.

The use of noncausal FIR filters allows us to deal with nonminimum mixing systems. The learning rules can be derived from different viewpoints such as information maximization, maximum likelihood and negentropy which result in similar rules for the weight update. We transform the learning rule into the frequency domain where the convolution and deconvolution property becomes a multiplication and division operation. In particular, the FIR polynomial algebra techniques as used by Lambert present an efficient tool to solve true phase inverse systems allowing a simple implementation of noncausal filters.

The significance of the methods is shown by the successful separation of two voices and separating a voice that has been recorded with loud music in the background. The recognition rate of an automatic speech recognition system is increased after separating the speech signals.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS5B Paper Number: 565 Oral

Dual cascade networks for blind signal extraction

Andrzej Cichocki, Ruck Thawonmas and Shun-ichi Amari

Keywords: Dual cascade networks blind signal extraction linear mixture

Abstract:

Keywords: Unsupervised Learning, Adaptive Algorithms, PCA and Prewhithening

A new neural-network approach is presented for extracting independent source signals one-by-one from a linear mixture of them when the number of noisy mixed signals is equal to or larger than the number of sources. In this approach, two types of cascade neural networks, having similar structures, are employed. The first cascade network performs prewhitening (preprocessing) of the mixed signals by sequentially extracting principal components. From the normalized (to unit variance) prewhitened signals, the second network, then, sequentially extracts the original source signals in order according to their stochastic properties, namely, in decreasing order of absolute values of normalized kurtosis. Extensive computer simulations confirm the validity and high performance of our approach.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS5C Paper Number: 566 Oral

Information back-propagation for blind separation of sources from non-linear mixture

Howard H. Yang, Shun-ichi Amari and Andrzej Cichocki

Keywords: Information back-propagation blind separation of sources non-linear mixture

Abstract:

The linear mixture model is assumed in most of the papers devoted to independent component analysis. A more realistic model for mixture should be non-linear. In this paper, a two layer perceptron is used as a de-mixing system to extract sources in non-linear mixture. The learning algorithms for the de-mixing system are derived by two approaches: maximum entropy and minimum mutual information. The algorithms derived from the two approaches have a common structure. The new learning equations for the hidden layer are different from our previous learning equations for the output layer. The natural gradient descent method is applied in maximizing entropy and minimizing mutual information. The information (entropy or mutual information) back-propagation method is proposed to derive the learning equations for the hidden layer.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS5D Paper Number: 513 Oral

Blind source separation and tracking using nonlinear PCA criterion: a least-squares approach

Juha Karhunen and Petteri Pajunen

Keywords: Blind source separation Blind source tracking nonlinear PCA criterion least-squares

Abstract:

In standard blind source separation, one tries to extract unknown source signals from their instantaneous linear mixtures by using a minimum of a priori information. We have recently shown that certain nonlinear PCA type neural learning rules can be successfully applied to this problem. In this paper, we introduce computationally efficient least-squares type algorithms for the basic blind source separation problem.

The proposed algorithms can still be regarded neural, and they have a close relationship to our previous algorithms. The new algorithms converge clearly faster and provide more accurate final results than our previous instantaneous stochastic gradient type learning algorithms. We also consider blind tracking of sources from nonstationary mixtures.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS5E Paper Number: 342 Oral

Single-sensor active noise cancellation using recurrent neural network predictors

Kyungmin Na and Soo-Ik Chae

Keywords: Single-sensor active noise cancellation recurrent neural network predictors

Abstract:

In this paper, we propose a recurrent neural network (RNN) predictor with an application to a single-sensor active noise cancellation (ANC) system. The proposed RNN predictor has one hidden layer whose neurons are classified into two categories, recurrent hidden neurons and non-recurrent hidden neurons. Due to the RNN's ability of modeling time-varying signals such as acoustic noises, the proposed RNN may be more suitable than the LMS-type digital filters and multilayer perceptrons (MLP). Moreover, the number of non-recurrent hidden neurons can be arbitrary increased according to the complexity of a given problem with a relatively little increase in computation during training. In the simulation on the noise data from a moisture-removing machine, about 22.35 dB attenuation was obtained with the proposed approach while 20.83 dB attenuation with the MLP-based approach, and 14.35 dB with a filtered-x LMS algorithm.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS5F Paper Number: 490 Oral

Elimination of multiple reflections in marine seismograms using neural networks

Robert Essenreiter, Martin Karrenbach, Sven Treitel

Keywords: marine seismograms neural networks deconvolution

Abstract:

We train an artificial neural network to perform deconvolution of seismic data and thereby recognize and remove multiple arrivals in reflection seismic data. Basis for the learning process is a well log that is typical for the area in which the data were gathered. Modeling data from this well log and comparing it to real recorded data allows deduce relations between the subsurface model in the recorded data. In contrast to conventional geophysical data processing techniques, the neural network does not depend on any assumptions concerning the underlying model. It is adaptive and able to learn highly non-linear interrelations in the data, should they exist. A further advantage of neural nets is that it is possible to make extensive use of a-priori knowledge by using information from existing well-logs. Preliminary tests with synthetic data show the ability of the neural net to extract the desired information.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS6A Paper Number: 481 Oral

Adaptive fuzzy filtering for audio applications using neuro-fuzzy modelization

Maddalena Di Giura, Nadia Serina, and Gianguido Rizzotto

Keywords: Fuzzy filtering modelization denoising

Abstract:

The paper describes a new denoising technique particulary suited for audio signals affected by white noise. The filtering algorithm is based on adaptive fuzzy rules taking into consideration the local temporal signal characteristics in order to estimate the noise components and consequently eliminate them. For a correct initial setting of the membership functions parameters describing the variables involved in the fuzzy processing, a pre-processing phase based on a neuro-fuzzy network has been implemented. The results of this non-linear approach compared with classical filtering techniques are found to be attractive especially for non-stationary signals.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS6B Paper Number: 37 Oral

A novel algorithm for wideband DOA estimates based on neural networks

Zhong Ding Lei, Xiu Kun Huang and Shu Jing Zhang

Keywords: wideband DOA estimates neural networks coherent signal subspace method

Abstract:

In this Paper, a novel class of focusing matrices for coherent signal subspace method(CSM) is proposed. These matrices are formed based on the concept of uniform focusing transformation of the direction-of-arrival (DOA). A focusning matrix upon the DOA's (q1, q2, ..., qn) is right a uniform focusing matrix upon each sub-set of { q1, q2, ..., qn}. If there exists a uniform focusing matrix upon an angle range, the uniform focusing matrix can be employed as the focusing matrix upon any DOA's belong to the angle range. In this paper, a three-layer BP network has been designed to approximate the uniform focusing matrix. In Conjunction with the CSM method, DOA estimation of super-high resolution can be obtained. The proposed DOA estimating procedure is efficient without prior knowledge of the DOA's and can be employed for an arbitrary geometry array.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS6C Paper Number: 561 Oral

Neural speech enhancement using dual extended Kalman filtering

Alex T. Nelson and Eric A. Wan

Keywords: Neural speech enhancement dual extended Kalman filtering cellular communications

Abstract:

Keywords: Kalman filtering, speech enhancement, speech processing, noise removal, robust estimation, state-space estimation

The removal of noise from speech signals has applications ranging from speech enhancement for cellular communications, to front ends for speech recognition systems. Spectral techniques are commonly used in these applications, but frequently result in audible distortion of the signal. A nonlinear time-domain method called dual extended Kalman filtering (DEKF) is presented that demonstrates significant advantages for removing nonstationary and colored noise from speech.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS6D Paper Number: 560 Oral

Combining fossil and sunspot data: committee predictions

Eric A. Wan

Keywords: fossil data sunspot data committee predictions

Abstract:

It is hypothesized that 680 million years ago solar magnetic storms producing ultraviolet and X-radiation affected the earths ozone layer, which in turn influenced the variations in the silt deposition from glacial run-off. Preserved as fossils discovered in South Australia, the striation widths constitute clues to ancient solar activity.

Utilizing this noisy data, we have improved our ability to predict the modern sunspot series. In this paper, we detail how the prediction results were achieved through training on the fossil data and committee predictions with the sunspots. Through this exercise, we develop general methods for combining predictors and also time series that may be related but separated in time.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS6E Paper Number: 637 Oral

Overcoming recurrent neural networks' compactness limitation for neurofiltering

James Ting-Ho Lo and Lei Yu

Keywords: recurrent neural networks compactness neurofiltering

Abstract:

Two range extenders and one range reducer for neural filtering are herein disclosed. The two range extenders are essentially an EKF and an accumulator respectively, which are used to extend the range of a recurrent neural network to cover the range of a signal process to be estimated. The range reducer herein disclosed is a differencer, which is used to reduce the range of the measurement process available for filtering.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TS6F Paper Number: 642 Oral

RBF neural network, basis functions and genetic algorithms

Eric P. Maillard and Didier Gueriot

Keywords: RBF neural network basis functions genetic algorithms

Abstract:

Theoretical researches focus on the capabilities of the network to reach an optimal solution. Unfortunately, few results concerning the design and training of the network are available. When dealing with a specific application, the performances of the network dramatically depend on the number of neurons and on the distribution of the hidden neurons in the input space. Generally, the network resulting from learning applied to a predetermined architecture, is either insufficient or over-complicated. In this study, we focus on genetic learning for the RBF network applied to prediction of chaotic time series.

The centers and widths of the hidden layer neurons basis function -- defined as the barycenter and distance between two input patterns -- are coded into a chromosome. It is shown that the basis functions which are also coded as a paramater of the neurons provide an additional degree of freedom resulting in a smaller optimal network. A direct inversion of matrix provides the weights between the hidden layer and the output layer and avoids the risk of getting stuck into a local minimum.

The performances of a network with Gaussian basis functions is compared with those of a network with genetic determination of the basis functions on the Mackey-Glass delay differential equation.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 463 Poster

Separation of two independent sources by the information-theoretic approach with cubic nonlinearity

Chi Chiu Cheung and Lei Xu

Keywords: source separation cubic nonlinearity Independent component analysis

Abstract:

We investigate the use of the simplest nonlinearity - cubic nonlinearity by the Information-theoretic approach on two signals in the Independent Component Analysis (ICA) problem. The mathematical analysis in this paper provides a global description of the cost function in the parameter space. It has also been proved that the general gradient algorithm can perform source separation on mixtures of two sources whose distributions are sub-gaussian in average. Experiments that demonstrate the results are presented. This paper provides an interesting insight in the role of nonlinearity in adaptive ICA algorithm.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 474 Poster

Financial prediction using higher order trigonometric polynomial neural network group model

Jing Chun Zhang and Ming Zhang

Keywords: Financial prediction Trigonometric polynomial NN.

Abstract:

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 415 Poster

Neural networks for periodicity analysis of unevenly spaced data

M. Rasile, L. Milano, R. Tagliaferri and G. Longo

Keywords: periodicity analysis time series

Abstract:

Periodicity analysis of unevenly collected data is a relevant issue in several scientific fields. In astrophysics, for example, we have to find the fundamental period of light or radial velocity curves which are unevenly sampled observations of stars. Classical spectral analysis methods are unsatisfactory to face the problem. In this paper we present a neural network based estimator system which performs well the frequency extraction in unevenly sampled signals. It uses a unsupervised Hebbian nonlinear neural algorithm to extract, from the interpolated signal, the principal components which, in turn, are used by the MUSIC frequency estimator algorithm to extract the frequencies. The neural network is tolerant to noise amplification due to interpolation and, above all, to blank time window in the data. We benchmark the system on synthetic, realistic and real signals with the Periodogram.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 310 Poster

Signal representation using adaptive wavelet-net

Chun-Ta Chen, Mahmood R. Azimi-Sadjadi and Chunhua Yuan

Keywords: Signal representation adaptive wavelet-net wavelet transform

Abstract:

In this paper, an adaptive wavelet network (Wave-Net) is introduced for optimal signal representation using continuous wavelet transform. This adaptive Wave-Net has the ability to adapt the number of hidden-layer nodes or wavelet basis to the characteristics of the signal. A node creation scheme is proposed to accomplish this structural adaptability. The wavelet coefficients for the specific signal representation problems are obtained by updating the weights of the adaptive Wave-Net using the conjugate gradient method. The effectiveness of the proposed adaptive Wave-Net is demonstrated on a signal representation example and the results are compared with those of the standard Wave-Net structure.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 294 Poster

A perceptron-like online algorithm for tracking the median

Tom Bylander and Bruce Rosen

Keywords: perceptron median series

Abstract:

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 266 Poster

Neural network autoregressive modeling of vibrations for condition monitoring of rotating shafts

A. C. McCormick and A. K. Nandi

Keywords: Neural network autoregressive modeling vibrations for condition monitoring of rotati

Abstract:

Artificial neural networks provide a means of capturing stationary statistical information about machine vibrations in the form of non-linear autoregressive models. These models can be used as one step ahead predictors allowing comparison of signals for the purposes of fault detection and diagnosis. From the prediction error, features can be extracted and used to determine the machine's condition. In this paper, the higher-order statistics of the error time series are extracted and used to compare vibration time series. Vibration data from a rotating shaft placed under different fault conditions were used for training and testing models. A statistical approach which assesses the probability that a fault has occurred is used and results indicate that this approach could be used to diagnose known conditions and even detect unknown faults.

Keywords: Machine Condition Monitoring, Non-linear Autoregressive Modeling, Higher-Order Statistics

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 666 Poster

Robust adaptive neurofilters with or without on-line weight adjustment

James Ting-Ho Lo

Keywords: Robust adaptive neurofilters on-line weight adjustment risk-sensitive criteria

Abstract:

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 620 Poster

Prediction of chaos and bifurcation: an asymmetric radial basis function approach

Hiroyuki Shibayama and Toshimichi Saito

Keywords: Prediction chaos bifurcation asymmetric radial basis function

Abstract:

This paper proposes an asymmetric basis function(ab. ABF) network and considers its application for prediction of chaotic time series and bifurcation phenomena. Using chaotic time series from an autonomous circuit, we have performed numerical simulation for the prediction problems and have confirmed that the ABF network has much better performance than conventional RBF networks.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 52 Poster

Neural prediction of missile dynamics during hardware-in-the-loop captive-carry experiments

Charles W. Gill and Phillip E. Pace

Keywords: Neural prediction missile dynamics hardware-in-the-loop captive-carry experiment

Abstract:

P-3 captive carry correlation algorithms combine hardware-in-the-loop missile simulator closed-loop anechoic chamber results and open-loop field test results to calculate a single measure of electronic attack effectiveness (miss distance). Prediction of the missile dynamics from the open-loop captive-carry seeker response represents an important component of these algorithms. This paper reports the use of neural networks to predict the missile dynamics by training Levenburg-Marquardt type predictors with the closed-loop seeker azimuth, elevation and range as inputs, the change in position of the missile in spherical coordinates (relative to the missile body), and the missile yaw and pitch as outputs.

The trained networks are then used to process the open-loop seeker information to predict the dynamics. The performance of the precictors is evaluated numerically using recent test results. The prediction error is also quantified.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 57 Poster

Improved option pricing using bootstrap methods

P.R. Lajbcygier and J.T. Connor

Keywords: option pricing bootstrap methods hybrid neural network

Abstract:

A "hybrid" neural network is used to predict the difference between the conventionally accepted modified Black option pricing model and observed intraday option pricies for stock index option futures. Confidence intervals derived with bootstrap methods are used in a trading strategy which allows only trades outsude the range of spurious model fits to be executed. Furthermore, "hybrid" neural network option pricing models can improve predictions but have bias which can be reduced with bootstrap methods. A modified bootstrap predictor is indexed by a parameter which allows the predictor to range from a pure bootstrap predictor, to a hybrid predictor, and finally the bagging predictor. Our results show that a modified bootstrap predictor outperforms the hybrid and bagging predictors. Greatly improved performance was observed in particular regions of the input space, namely out of the money options.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 70 Poster

A learning algorithm for modified recurrent neural networks

C.H. Chen and Liwen Yu

Keywords: learning algorithm modified recurrent neural networks nonstationary time series

Abstract:

An improved recurrent neural network structure is proposed. The exact form of gradient-following learning algorithm for the continuously running neural networks is derived for temporal supervised learning tasks. The algorithm allows networks to learn complex tasks that require the retention of information over time periods. The algorithm also compensates for the information that is missed by the traditional recurrent neural networks. Empirical results show that the networks trained using this algorithm have improved prediction performance over the backpropagation trained network and the Elman recurrent neural network._____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 134 Poster

Prediction of onset of respiratory disorder in neonates

Emma Braithwaite, Jimmy Dripps and Alan F. Murray

Keywords: Prediction respiratory disorder neonates

Abstract:

Premature extremely sick babies are currently monitored by skilled medical staff using numerous dedicated non-invasive sensors and associated monitoring equipment. This paper describes a method of ``fusing'' a number of the physiological signals and, by examining them simultaneously and continously in time, produces an early warning for the onset of respiratory disorder (RD). The method uses a multi-layer perceptron neural network to produce probabilities that the patient is going to suffer from RD at some point within the next thirty minutes. Initial results from this classification system are shown and suggestions for further work are given.

_____

ICNN97 Speech Processing, Time Series & Filtering Session: TSP1 Paper Number: 427 Poster

Minor Component analysis with implementation to blind 2-channel equalization

Jie Luo and Xieting Ling

Keywords: Minor component analysis mobile communication equalizing

Abstract:


Web Site Author: Mary Lou Padgett (m.padgett@ieee.org)
URL: http://www.mindspring.com/~pci-inc/ICNN97/paperts.htm
(Last Modified: 30-Apr-1997)