Home Home
Barbara Varone, Politecnico di Milano (Italy)
Jarno M.A. Tanskanen, Helsinki University of Technology (Finland)
Seppo J. Ovaska, Helsinki University of Technology (Finland)
In this paper, we investigate the characteristics of some one-step-ahead nonlinear predictors based on a two-layer feed-forward neural network (2LFNN). The behavior of neural networks (NN) is investigated in the frequency domain using two frequency response estimation techniques, and in the time domain, by analyzing the unit step and triangular pulse responses. Some of the estimated frequency responses of these NNs resemble those of corresponding linear polynomial predictors, revealing the nearly polynomial nature of the applied training signals. Similarity of the two frequency response estimates is an indication of good generalization properties.
ic973309.pdf ic973309.pdf
TOPCyril Goutte, D.T.U. (Denmark)
The purpose of this contribution is to investigate some techniques for finding the relevant lag-space, i.e. input information, for time series modelling. This is an important aspect of time series modelling, as it conditions the design of the model through the regressor vector a.k.a. input layer in a neural network. We give a rough description of the problem, insist on the concept of generalisation, and propose a generalisation-based method. We compare it to a non-parametric test, and carry out experiments, both on the well-known Hénon map, and on a real data set.
ic973313.pdf ic973313.pdf TOP
Janusz Mazurek, Neurolab (Germany)
Adam Krzyzak, Neurolab (Germany)
Andrzej Cichocki, Neurolab (Germany)
Recursive radial basis function (RRBF) neural networks are introduced and discussed. We study in detail the nets with diagonal receptive field matrices. Parameters of the networks are learned by a simple procedure. Convergence and the rates of convergence of RRBF nets in the mean integrated absolute error (MIAE) sense are studied under mild conditions imposed on some of the network parameters. Obtained results give also upper bounds on the performance of RRBF nets learned by minimizing empirical L1 error.
ic973317.pdf ic973317.pdf TOP
A. Neil Birkett, Carleton University (Canada)
Rafik A. Goubran, Carleton University (Canada)
This paper focuses on multilayer perceptron neural networks where the activation functions are adaptive and where each neuron synapse is modelled by a finite impulse response (FIR) filter. A simplified architecture consisting of a variable activation (VA) function which is sandwiched between two FIR synapses is studied. The VA function consists of a mixed linear-tanh sigmoid with a parameter which controls the linear region.The VA parameters and FIR synaptic weights are updated using a modified form of the instantaneous-cost (IC) temporal backpropagation algorithm. Simulations for identifying cascaded nonlinear transfer functions with internal memory and arbitrary activation functions illustrate the improved modelling performance over models with non-adaptive activation functions.
ic973321.pdf TOP
Craig L. Fancourt, University of Florida (U.S.A.)
Jose C. Principe, University of Florida (U.S.A.)
Two self-organizing principles for the competitive identification and segmentation of piecewise stationary time series are described. In the first, a neighborhood map of one step predictors competes for the data during training. The winner is granted the largest parameter update, while other predictors are allowed smaller updates, decreasing with distance from the winner on the neighborhood map. In addition to performing piecewise segmentation and identification, the technique maps similar segments of the time series as neighbors on the neighborhood map. In the second, we propose a new cost function for competitive prediction that imbeds memory in the error metric and couples the memory with the degree of competition. Performing gradient descent on the cost function yields a self-annealing system that can also perform piecewise segmentation and identification of a time series.
ic973325.pdf ic973325.pdf TOP
Mark Girolami, University of Paisley, Scotland (U.K.)
Colin Fyfe, University of Paisley, Scotland (U.K.)
We propose a nonlinear self-organising network which solely employs computationally simple hebbian and anti-hebbian learning in approximating a linear independent component analysis (ICA). Current neural architectures and algorithms which perform parallel ICA are either restricted to positively kurtotic data distributions or data which exhibits one sign of kurtosis. We show that the proposed network is capable of separating mixtures of speech, noise and signals with both platykurtic (positive kurtosis) and leptokurtic (negative kurtosis) distributions in a blind manner. A simulation is reported which successfully separates a mixture of twenty sources of music, speech, noise and fundamental frequencies.
ic973329.pdf ic973329.pdf TOP
Guiqing He, NTU (Singapore)
Alex C. Kot, NTU (Singapore)
In this paper a new adaptive neural network multiuser detector is proposed and investigated for synchronous code-division multiple-access (CDMA) systems. The proposed multiuser detector includes two parts: a decorrelating detector as its auxiliary detector and an adaptive multiple layer perceptron (MLP) detector as its main detector. At the setup stage, the auxiliary detector detects the user's transmitted data and at the same time feeds these output data to the main MLP detector as its training data, and the main detector is trained by using the well-known backpropagation (BP) algorithm. After the training process, the auxiliary detector stops work and the main detector starts detecting the user's transmitted data self-adaptively. The proposed detector is blind and can provide near minimum bit-error-rate performance.
ic973333.pdf TOP
Sylvain Nadaud, L.A.M.I. (France)
J.F. Trouilhet, L.A.M.I. (France)
This article presents two modelling methods using wavelet networks. Both methods are intended to be used for an acoustic pulse signal classifier. We present a few results obtained with signals coming from a recording of the percussive response of metal parts. The object of this application is the non-destructive testing of these parts, as defects perturb the acoustic signature. The first modelling method uses wavelet networks to perform a non-linear regression on the signal to be classified. The second consists of non-linear auto-recursive modelling of the signal by means of the networks. The use of wavelet networks enables us to combine the generalizing capacities of neural networks with the efficiency of wavelet analysis of pulse signals.
ic973337.pdf TOPAmir Hussain, University of Paisley (U.K.)
In this paper a new two-layer linear-in-the-parameters feedforward network termed the Functionally Expanded Neural Network (FENN) is presented, together with its design strategy and learning algorithm. It can be considered to a hybrid neural network incorporating a variety of non-linear basis functions within its single hidden layer which emulate other universal approximators employed in the conventinal Multi-Layered Perceptron (MLP), Radial Basis Function (RBF) and Volterra Neural Networks (VNN). The FENN's output error surface is shown to be uni-modal allowing high speed single run learning. A simple strategy based on an iterative pruning re-training scheme coupled with statistical model validation tests has been proposed for pruning the FENN. Both simulated chaotic (Mackey Glass time series) and real-world noisy, highly non-stationary (sunspot) time series have been used to illustrate the superior modeling and prediction performance of the FENN compared with other recently reported , more complex feedforward and recurrent neural network based predictor models.
ic973341.pdf ic973341.pdf TOP
Konstantinos Plataniotis, University of Toronto (Canada)
Dimitrios Androutsos, University of Toronto (Canada)
Anastasios Venetsanopoulos, University of Toronto (Canada)
A new approach to the problem of time series classification is discussed in this paper. A new adaptive classification scheme is introduced and compared w ith existing approaches, such as the Bayesian approach and the Incremental Credit Assignme nt approach. Simulation results are included to demonstrate the effectiveness of the new m ethodology.
ic973345.pdf ic973345.pdf TOP
Simone Fiori, University of Ancona (Italy)
Paolo Campolucci, University of Ancona (Italy)
Aurelio Uncini, University of Ancona (Italy)
Francesco Piazza, University of Ancona (Italy)
We derive a new class of neural unsupervised learning rules which arises from the analysis of the dynamics of an abstract mechanical system. The corresponding algorithms can be used to solve several problems in Digital Signal Processing area, where orthonormal matrices are involved. We present an application which deals with blind separation of sources, i.e. a new method to perform efficient Independent Component Analysis (ICA) of random signals.
ic973349.pdf ic973349.pdf TOP
Ruck Thawonmas, RIKEN (Japan)
Andrzej Cichocki, RIKEN (Japan)
We present a neural-network approach which allows sequential extraction of source signals from a linear mixture of multiple sources in the order determined by absolute values of normalized kurtosis. To achieve this, we develop a non-linear Hebbian learning rule for extraction of a single signal. We discuss several techniques which enable extraction of signals not randomly but in the desired order. To prevent the same signals from being extracted several times, a robust deflation technique is used which eliminates from the mixture the already extracted signals. Extensive computer simulations confirm the validity and high performance of our method.
ic973353.pdf ic973353.pdf TOP
Sung Yoon, Villanova University (U.S.A.)
Sathyanarayan S. Rao, Villanova University (U.S.A.)
We present a high performance neural network based multiuser detector in code division multiple access (CDMA) communications. This detector retains salient features of the annealed neural network based detector (ANNMD) and reduces its computational complexity. The BER performance of the proposed detector is close to the theoretical lower bound of the BER performance. We present a theoretical derivation of the hybrid type multiuser detector with improved hardware complexity. Extensive numerical evaluation of the proposed techniques as well as various suboptimal multiuser detectors is conducted using Monte-Carlo simulation.
ic973357.pdf TOP
Juha Karhunen, Helsinki University of Technology (Finland)
Petteri Pajunen, Helsinki University of Technology (Finland)
In this paper adaptive least-squares type algorithms are introduced for blind source separation. They are based on minimizing a criterion used in context with nonlinear PCA (Principal Component Analysis) networks. The new algorithms converge clearly faster and provide more accurate results than typical current adaptive blind separation algorithms based on instantaneous gradients. They are also applicable to the difficult case nonstationary mixtures. The proposed algorithms have a close relationship to a nonlinear extension of Oja's PCA learning rule. A batch algorithm based on the same criterion is also presented.
ic973361.pdf ic973361.pdf TOP