Chair: C. Lee Giles, NEC Corporation (USA)
Xieting Ling, Fudan University (CHINA)
Wei Tian, Fudan University (CHINA)
Bin Liu, Fudan University (CHINA)
Ruey-wen Liu, University of Notre Dame (USA)
This paper presents a linear feedback neural network and an adaptive algorithm to achieve the blind signal separation under near-field situation. The convergence property of algorithm and stability of equilibrium state are discussed. Some simulations are provided.
John Aasted Sorensen, Technical University of Denmark (DENMARK)
A time signal prediction algorithm based on Relative Neighborhood Graph (RNG) localized FIR filters is defined. The RNG connects two nodes, of input space dimension $D$, if their lune does not contain any other node. The FIR filters associated with the nodes, are used for local approximation of the training vectors belonging to the lunes formed by the nodes. The predictor training is carried out by iteration through 3 stages: Initialization of the RNG of the training signal by vector quantization, LS estimation of the FIR filters localized in the input space by RNG nodes and adaptation of the RNG nodes by equalizing the LS approximation error among the lunes formed by the nodes of the RNG. The training properties of the predictor is exemplified on a burst signal and characterized by the normalized mean square error (NMSE) and the mean valence of the RNG nodes through the adaptation.
F. Bartolini, Universita di Firenze
V. Cappellini, Universita di Firenze
S. Nerozzi, Universita di Firenze
A. Mecocci, Universita di Pavia (ITALY)
The progress of digital electroencephalography gave rise to the problem of EEG data recording. In this paper a DPCM scheme for EEG signal compression is discussed. In particular the performance of a class of predictors based on recurrent neural networks is presented. The training strategy is accurately described and the results of a comparison with some other classical linear and static neural predictors are given. The proposed recurrent neural predictor demonstrates to be competitive with the others in offering good performance at a very low computational cost.
M.J. Bradley, University of Durham (UK)
P. Mars, University of Durham (UK)
This paper examines the mechanism by which Recurrent Neural Networks (RNNs) acheive equalization whilst operating on simple digital communication channels. The mode of operation is seen to be essentially similar to the conventional Decision Feedback Equalizer (DFE) and the RNN node nonlinearity is identified as a limiting factor. Two versions of an alternative structure are formulated for channels with longer impulse responses based on soft-decision feedback. Simulations demonstrate the improved performance compared with the DFE.
Pauli Kuosmanen, Tampere University of Technology (FINLAND)
Karen Egiazarian, Tampere University of Technology (FINLAND)
Jaakko Astola, Tampere University of Technology (FINLAND)
In the present work weighted Chow parameters are developed with the aim of their application in the statistical analysis of a class of nonlinear filters, namely stack filters, which are specified by positive Boolean functions (PBF) representing the binary output at each threshold level of the continuous-valued signal. Selection probabilities of stack filters were defined based on the fact that the output of a continuous stack filter is one of the samples within the input window. The notion of weighted Chow parameters is introduced in this paper for analysis and computation of the sample selection probability vector of a continuous stack filter.
Bryan W. Stiles, University of Texas-Austin (USA)
Joydeep Ghosh, University of Texas-Austin (USA)
Based on the habituation mechanism found in biological neural systems, novel dynamic neural networks are proposed for recognizing temporal patterns. The specific task considered in this paper is the classification of whale songs from passive SONAR data, but the networks are also readily applicable to other temporal pattern recognition problems. The fact that the networks designed operate dynamically is important, because it makes the goal of real time data analysis possible.
X.Q. Gao, Southeast University (PEOPLES REPUBLIC OF CHINA)
X. D. Wang, Southeast University (PEOPLES REPUBLIC OF CHINA)
L.H. Li, Southeast University (PEOPLES REPUBLIC OF CHINA)
H. Zhang, Southeast University (PEOPLES REPUBLIC OF CHINA)
Z.Y. He, Southeast University (PEOPLES REPUBLIC OF CHINA)
This paper presents a decision feedback neural network (NN) coherent receiver schemefor continuous phase modulation based on frequency domain. Through decision feedback pre-processing, the effect of the previous transmitted symbols can be removed for the present symbol decision. By employing Karhunen-Loeve transform (KLT) or discrete cosine transform (DCT), the input data number of neural networks can be reduced significantly. To obtain more sufficient convergence of neural networks a modified "delta-bar-delta" BP learning algorithm is proposed. Despite low complexity in NN training and implementation, computer simulation results show that our NN receivers can achieve near optimal demodulation performance.
Anthony J. Bell, The Salk Institute (USA)
Terrence J. Sejnowski, The Salk Institute (USA)
A new unsupervised learning algorithm is derived and applied to the problems of blind separation and blind deconvolution. It maximises, by stochastic gradient ascent, the joint entropy of a non-linearly deterministically transformed representation of the signal or mixtures. With one extra (weak) assumption, this process can be shown to yield a factorial representation, splitting inputs into their statistically independent components, or removing inter-symbol dependencies in a time series. The non-linearity supplies the higher-order statistical information required for performing these tasks,circumventing the need for cumulant expansions. The algorithm has advantages over Bussgang methods for blind deconvolution and the Herault-Jutten method for blind separation, in that the use of non-linearity is rigorously linked to an information-theoretic objective function. In simulations, the algorithm perfectly separates 10 digitally mixed signals (voices, music), and also cancels echoes and reverberations in a speech signal.
Paul Yee, McMaster University (CANADA)
Simon Haykin, McMaster University (CANADA)
A dynamic network of regularized Gaussian radial basis functions (GaRBF) is described for the one-step prediction of nonlinear, nonstationary autoregressive (NLAR) processes governed by a smooth process map and a zero-mean, independent additive disturbance process of bounded variance. For N basis functions, both full-order and reduced- order updating algorithms are introduced, having computational complexities of O(N^3) and O(N^2), respectively, per time step. Simulations on a 10,000 point, 8-bit quantized 64kbps rate speech signal show that the proposed dynamic algorithm has a prediction performance comparable and, in some cases, superior to that of AT&T's LMS-based speech predictor designed for the ITU-T G.721 standard on the 32kbps ADPCM of speech. The results indicate that the proposed dynamic regularized GaRBF predictor provides a useful tradeoff between its minimal need for prior knowledge of the speech data characteristics and its consequently heavier computational burden.
Kanaan Jemili, University of Dayton (USA)
John J. Westerkamp, University of Dayton (USA)
A new signal estimation technique is introduced for highly non-stationary signals. The system uses the wavelet transform to extract time-frequency components of the signal plus noise, followed by a radial basis function neural network that adaptively estimates the underlying signal. The method is applied to the visual evoked potential (EP) signal, which is a transient signal corrupted by the ongoing elctroencephalogram (EEG) noise, with a signal-to-noise ratio often less than -6 dB. The proposed system gives good time-varying estimates of the EP, while suppressing the on-going EEG. The performance is given in terms of mean-square error, bias factor and noise reduction factor.