Scheduled System Maintenance:
Some services will be unavailable Sunday, March 29th through Monday, March 30th. We apologize for the inconvenience.
By Topic

Neural Networks for Signal Processing [1996] VI. Proceedings of the 1996 IEEE Signal Processing Society Workshop

Date 4-6 Sept. 1996

Filter Results

Displaying Results 1 - 25 of 61
  • Neural Networks for Signal Processing VI. Proceedings of the 1996 IEEE Signal Processing Society Workshop

    Publication Year: 1996
    Save to Project icon | Request Permissions | PDF file iconPDF (270 KB)  
    Freely Available from IEEE
  • Author index

    Publication Year: 1996
    Save to Project icon | Request Permissions | PDF file iconPDF (60 KB)  
    Freely Available from IEEE
  • An incremental learning method with relearning of recalled interfered patterns

    Publication Year: 1996 , Page(s): 243 - 252
    Cited by:  Papers (2)  |  Patents (1)
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (496 KB)  

    This paper presents a new incremental learning method for neural networks. If a neural network is trained to memorize novel patterns only by their presentation, the network will forget some patterns that have been already learnt. This problem is caused by the fact that the learning of novel patterns usually interfere in the internal representation corresponding to the old training patterns. In the new method, the network recalls the patterns that the novel patterns possibly interfere in, and then learns both novel and recalled patterns. In the computer simulation, we demonstrate this system in the learning of alphabetic characters View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Sparse distributed memory implementations on tree shape parallel neurocomputer

    Publication Year: 1996 , Page(s): 539 - 548
    Cited by:  Papers (1)
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (544 KB)  

    This paper presents two different realizations of a sparse distributed memory (SDM) model. For parallelization purposes, addressing, storage and retrieval operations are explained in detail and some existing implementations in various computing platforms are considered before introducing the tree shaped parallel computer, TUTNC (Tampere University of Technology Neural Computer). The architecture and main features of TUTNC are presented in order to map SDM to the system in columnwise and rowwise manner. Mappings are compared in terms of measured execution time with different parameter sets. Speedup and performance estimations are also given for a larger system. The results show, that SDM can be well parallelized in TUTNC View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Structured neural networks for multi-font Chinese character recognition using a newly developed digital neural network chip with adaptive segmentation of quantizer neuron architecture (ASQA)

    Publication Year: 1996 , Page(s): 330 - 339
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (348 KB)  

    This paper describes structured networks that use a digital network chip with having adaptive segmentation of quantizer neuron architecture (ASQA) and presents results of applying the ASQA chip to the large scale problem of multi-font Chinese character recognition. The ASQA chip can simulate neural networks using ASQA model which can provide a proliferation of neurons based on input data for learning and can generate appropriate network structure with extremely fast processing speed. Moreover, this chip can simulate not only a single network but also sets of several structured networks; consequently, the chip can handle large scale problems. By applying the chip to multi-font Chinese character recognition, average accuracy of the open test increased to 97% and a recognition speed of 6 msec/character was achieved View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A bit-serial VLSI processor for kernel-based classifiers

    Publication Year: 1996 , Page(s): 549 - 558
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (424 KB)  

    A bit-serial digital VLSI processor that emulates the operation of kernel-based classifiers is described. First the operating principle and the relevant characteristics of that kind of neural classifiers are introduced. Operations are identified and suitable approximations are proposed in order to make the required computations compatible with a digital VLSI implementation. Then, the bit-serial processor architecture is described. Basically it consists of three modules: distance calculation, kernel function approximation and accumulation-comparison. Estimated results on area, processing speed and power consumption of a synthesized ASIC are given in comparison with a published analog processor. The obtained data show that the reported approach is very promising for high throughput applications View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A digital neuro chip with adaptive segmentation quantizer neuron architecture (ASQA)

    Publication Year: 1996 , Page(s): 559 - 568
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (448 KB)  

    We discuss a chip which simulates a neural network and automatically generates optimum network structure according to input data. It handles 128 sub neural networks which compose a large scale neural network. By our original architecture, necessary memory size to get the same recognition performance as a conventional chip is reduced to 9%. It classifies up to 16.384 categories and solves large size problems such as Kanji recognition on a single chip. It consists of 250 K transistors on a 6.92 mm×7.08 mm chip by 0.5 μm double metal CMOS technology View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Neural network approaches for the extraction of the eigenstructure

    Publication Year: 1996 , Page(s): 23 - 32
    Cited by:  Papers (2)
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (372 KB)  

    A feedback neural network for eigen-decomposition of a positive semidefinite matrix is presented. In the paper, we have shown the stability and real-time eigen-decomposition computation ability of the proposed neural network. The network can go into the stable state in the magnitude of the circuit time constant. The output voltage of the net, under the smallest energy state, is just the eigenvector corresponding to the minimum eigenvalue of the NN's connection strength matrix R, which is the direct use of the data covariance matrix without any pre-processing. By taking reasonable process to R, we can further extract the other eigen-parameters of R. A number of computer simulation results have been made to verify the effectiveness of the network. Both the theoretical analysis and the experimental results show that the proposed net can perform the extraction of the eigenstructure in real time View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A stochastic natural gradient descent algorithm for blind signal separation

    Publication Year: 1996 , Page(s): 433 - 442
    Cited by:  Papers (2)
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (340 KB)  

    A new blind separation algorithm is derived based on minimizing the mutual information of the output of the de-mixing system using natural gradient descent method. The algorithm can be easily implemented on a neural network with data dependent activation functions. A new performance function which depends only on the output and the de-mixing matrix is introduced. The new performance function is evaluated without any knowledge of the mixing matrix except for its order. It is very useful for comparing the performance of different blind separation algorithms. The performance of the new algorithm is compared to that of some existing blind separation algorithms by using the performance function. The new algorithm generally outperforms the existing algorithms because it minimizes the mutual information directly. This is verified by the simulation results View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Supervised learning for multilayered neural network with non-monotonic activation functions

    Publication Year: 1996 , Page(s): 13 - 22
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (356 KB)  

    We describe the performance of multilayer neural network with hidden units of nonmonotonic activation functions. Our previous work has shown that the network was effective in improving two difficulties: a convergence to local minima and a slow learning speed for the exclusive-OR and the binary addition problems. The purpose of this paper is to evaluate the performance of the proposed network for more complicated tasks, that is, N-bits parity tasks and two spirals task. Furthermore, we evaluate the generalization performance of the network for an acoustic diagnosis. The results show that the networks are effective for the tasks and have the same generalization performance as the networks with the sigmoidal activation functions View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Defect diagnosis of solder joints using fuzzy logic

    Publication Year: 1996 , Page(s): 502 - 509
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (412 KB)  

    This paper describes how the methods of fuzzy logic can be used in the classification of solder joint inspection results. The inspection results of over 900 circuit boards have been analysed for the purpose of distinguishing the essential characters of defective and decent solder joints. Also the effect of the prevailing average solder amount of each component type on the decision making process has been considered. The developed prototype of a fuzzy classifier of defect reports accomplishes the classification by using the information received from the solder joint inspection system. Defect reports are classified to be either justified or unnecessary. The classifier is able to detect about 80% of unnecessary defect reports and thus reduces the amount of work in the required visual re-inspection View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • System identification using neural networks

    Publication Year: 1996 , Page(s): 82 - 88
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (348 KB)  

    We examine the complexity of neural networks required to approximate an unknown system to a given degree of accuracy. We establish lower bounds on the number of neurons, as well as constructing networks to “almost” achieve this lower bound in the worst case analysis. Our constructions are simple, deterministic, and involve no optimization based training, such as backpropagation View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Texture feature extraction by wavelet projection pursuit for cloud detection in AVIRIS and AVHRR imagery

    Publication Year: 1996 , Page(s): 351 - 360
    Cited by:  Papers (1)
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (472 KB)  

    A constrained projection pursuit model is used to extract textural features from two multi-spectral, remote sensing databases of AVIRIS and AVHRR imagery. The goal is to evaluate methods which may eventually be useful for cloud detection in multi-angle imaging spectrometer (MISR) imagery. The method is an unsupervised adaptive wavelet technique that topologically constrains the projection vector and complements other techniques for extracting texture information View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Unsupervised segmentation of multispectral images using hierarchical MRF model

    Publication Year: 1996 , Page(s): 381 - 390
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (480 KB)  

    This paper proposes an Markov random field (MRF) model-based method for unsupervised segmentation of multispectral images, in which the intra-class correlation of multispectral data as well as the class correlation are taken into account. In this method a set of multispectral images is modeled by a hierarchical MRF model. The proposed segmentation method is an iterative method composed of parameter estimation and segmentation which is based on the framework of the expectation-maximization (EM) method. Making use of an approximation for the Baum function in the expectation step, parameter estimation is reduced to the conventional maximum likelihood (ML) estimation given the current estimate of the hidden class label. The estimation of the class label, which corresponds to image segmentation, is carried out by a deterministic relaxation method proposed by us View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • On the use of a pruning prior for neural networks

    Publication Year: 1996 , Page(s): 52 - 61
    Cited by:  Papers (2)
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (448 KB)  

    We address the problem of using a regularization prior that prunes unnecessary weights in a neural network architecture. This prior provides a convenient alternative to traditional weight-decay. Two examples are studied to support this method and illustrate its use. First we use the sunspots benchmark problem as an example of time series processing. Then we address the problem of system identification on a small artificial system View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Scaling properties of neural networks for the prediction of time series

    Publication Year: 1996 , Page(s): 190 - 199
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (448 KB)  

    Sealing properties of neural networks, i.e. relations between the number of hidden units and the training or generalization error, recently have been investigated theoretically with encouraging results. In this paper we investigate experimentally, whether the theoretic results may be expected in practical applications. We investigate different neural network structures with varying number of hidden units for solving two time series prediction tasks. The results show a considerable difference of the scaling behavior of multilayer perceptrons and radial basis function networks View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Speech enhancement based on extended Kalman filter and neural predictive hidden Markov model

    Publication Year: 1996 , Page(s): 302 - 310
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (360 KB)  

    To represent the nonlinear and nonstationarity nature of speech, we assume that speech is the output of an NPHMM combining a neural network and hidden Markov model (HMM). The NPHMM is a nonlinear autoregressive process whose time-varying parameters are controlled by a hidden Markov chain. Given some speech data for training, the parameter of NPHMM is estimated by a learning algorithm based on the combination of Baum-Welch's algorithm and a neural network learning algorithm using the backpropagation algorithm. A recursive method using a extended Kalman filter with the parameters of a trained NPHMM is developed for enhancing speech signals degraded by statistically independent additive noise assumed to be white Gaussian. Our recursive speech enhancement method achieves an improvement over the method hidden filter model of about 1-1.5 dB in SNR View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Temporal decorrelation using teacher forcing anti-Hebbian learning and its application in adaptive blind source separation

    Publication Year: 1996 , Page(s): 413 - 422
    Cited by:  Papers (3)
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (408 KB)  

    This paper proposes a network architecture to compute on-line the temporal crosscorrelation function between two signals, either stationary or locally stationary. We show that the weights of a multi-FIR (finite impulse response) filter trained with a teacher forcing anti-Hebbian rule encode the crosscorrelation function between the input and the desired response. We extend this network to the Gamma filter which is an IIR (infinite impulse response) filter and also to nonlinear filters. This temporal correlation idea is applied to the blind source separation problem. From these networks we build a recurrent system trained on-line with anti-Hebbian learning which performs temporal decorrelation on the mixed signals. The system performance is tested in speech signals mixed in time with good results. A comparison of the performance among the different topologies is also presented in the paper View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • The fuzzy leaky bucket for policing mechanism in ATM networks

    Publication Year: 1996 , Page(s): 510 - 517
    Cited by:  Papers (1)
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (372 KB)  

    This paper presents a model of fuzzy leaky bucket (FLB) for policing mechanism (PM) in ATM networks. Because of the fundamental contract, call acceptance control (CAC) may not be reasonable and some traffic is very difficult to model, thus dynamic adjustment is needed during practical applications. We have presented the function and theory of the designed FLB. The conclusions show that the FLB is a better dynamic method of PM than the conventional one. The cell loss probability and mean time delay are lowered and the utilization of bandwidth is improved in the FLB View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Periodic multiresolution analysis using quasi-orthogonals splines

    Publication Year: 1996 , Page(s): 89 - 98
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (348 KB)  

    We present a periodic spline quasi-wavelet generated from a quasi-orthogonal spline basis. By construction the quasi-wavelet and related multiresolution spaces are approximately orthogonal. The quasi-wavelet allows for multiresolution decomposition using a single scaling function. This is in contrast with classic periodic wavelets which require a different scaling function between each resolution. This results in a fast algorithm to create the periodic multiresolution pyramid of least squares fits to periodic data View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Classification of seismic waveforms by integrating ensembles of neural networks

    Publication Year: 1996 , Page(s): 453 - 462
    Cited by:  Patents (1)
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (592 KB)  

    The problem considered is the discrimination between natural and artificial seismic events, based on their waveform recording. We build a classification environment consists of several ensembles of neural networks trained on boot-strap sample sets, using various data representations and architectures. The integration of the different ensembles is made in a nonconstant signal adaptive manner, using a posterior confidence measure based on the agreement (variance) within the ensembles. The proposed integrated classification machine achieved 92.1% correct classification on the seismic test data. Cross validation tests and comparisons indicate that such integration of a collection of ANN's ensembles is a robust way for handling high dimensional problems with a complex nonstationary signal space as in the current seismic classification problem View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A neural network trained microphone array system for noise reduction

    Publication Year: 1996 , Page(s): 311 - 319
    Cited by:  Papers (1)
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (412 KB)  

    This paper presents a neural network based microphone array system, which is capable to continuously perform speech enhancement and adaptation to nonuniform quantization, such as A-law and μ-law. Such a quantizer is designed to increase the signal to quantization noise ratio (SQNR) for small amplitudes in telecommunications systems. The proposed method primarily developed for hand-free mobile telephones, suppresses the ambient car noise with approximately 10 dB. The system is based upon a multilayer nonlinear backpropagation trained network by using a built-in calibration technique View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • An adaptive Kalman filter model for stable perception of visual space unaffected by eye movement

    Publication Year: 1996 , Page(s): 492 - 501
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (556 KB)  

    A number of visual psychological-physical (PP) models, i.e., the saccadic suppression model, exist for explaining the phenomenon of visually stable perception in visual space regardless of eye movement (EM). However, most models ignore the visual system noise that is added between the retina and the high visual information processing system. This paper introduces a new visual model based on an adaptive Kalman filter (AKF) model, derived from an optimum control theory, which can optimally estimate and predict near-original visual information from a visual information signal disturbed by noise. The AKF model has three main mechanisms that eliminate some of the deficiencies appearing in conventional PP models: filter gain adapts in response to texture similarity, optimal estimation/prediction, and a smoothing effect from step-wise motion of EM. Based on visual experiments, three autocorrelation functions concerned with EM acceleration when viewing eleven texture images are clustered. Numerical simulations show that AKF can cover several of the visual phenomena which traditionally thought to be caused by different mechanisms View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • An analysis on genetic algorithms using Markov process with rewards

    Publication Year: 1996 , Page(s): 129 - 138
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (388 KB)  

    We propose a new method to analyze the behavior of genetic algorithms (GAs) using Markov processes with rewards, which are extensions of Markov processes by introducing a concept of rewards. We analyze some simple models of GAs by our method and derive expected maximum and mean fitness values of these models. These values are explicitly expressed as functions of generations and can be calculated without simulations, even for the generations at infinity. We discuss the optimum value of mutation rate and compare the maximum and mean fitness based on these results View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Genetic programming techniques that evolve recurrent neural network architectures for signal processing

    Publication Year: 1996 , Page(s): 139 - 148
    Cited by:  Papers (2)
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (416 KB)  

    We propose a novel design paradigm for recurrent neural networks. This employs a two-stage genetic programming/simulated annealing hybrid algorithm to produce a neural network which satisfies a set of design constraints. The genetic programming part of the algorithm is used to evolve the general topology of the network, along with specifications for the neuronal transfer functions, while the simulated annealing component of the algorithm adapts the network's connection weights in response to a set of training data. Our approach offers important advantages over existing methods for automated network design. Firstly, it allows us to develop recurrent network connections; secondly, we are able to employ neurones with arbitrary transfer functions, and thirdly, the approach yields an efficient and easy to implement on-line training algorithm. The procedures involved in using the GP/SA hybrid algorithm are illustrated by using it to design a neural network for adaptive filtering in a signal processing application View full abstract»

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