By Topic

Neural Networks for Processing [1993] III. Proceedings of the 1993 IEEE-SP Workshop

Date 6-9 Sept. 1993

Filter Results

Displaying Results 1 - 25 of 63
  • Neural Networks for Signal Processing III - Proceedings of the 1993 IEEE-SP Workshop

    Save to Project icon | Request Permissions | PDF file iconPDF (37 KB)  
    Freely Available from IEEE
  • Ordered vector quantization for neural network pattern classification

    Page(s): 141 - 150
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (668 KB)  

    The accurate classification of time sequences of vectors is a common goal in signal processing. Vector quantization (VQ) has commonly been used to help encode vectors for subsequent classification. The authors depart from this past approach proposing the use of VQ codebook indices, as opposed to codebook vectors. It is shown that one-dimensional ordering of these indices markedly improves the neural-network-based classification accuracy of acoustic time-frequency patterns. The needs for and extensions of multidimensional codebook indices are described View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Detection of ocean wakes in synthetic aperature radar images with neural networks

    Page(s): 261 - 270
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (496 KB)  

    Two neural networks are combined to detect wakes in synthetic aperture radar (SAR) images of the ocean. The first network detects local wake features in smaller sub-proportions of the image, and the second network integrates the information from the first network to determine the presence or absence of a wake in the entire image. The networks train directly using the gradient descent method on either real SAR images or on synthetic images and are designed to detect wakes in images with low signal-to-noise ratios. When trained on real images, the network detector recognizes the wake in any translation and is robust with respect to rotations. With synthetic images, the network model is able to recognize wakes with all possible translations, rotations and over a wide range of opening angles View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Applying neural network developments to sign language translation

    Page(s): 301 - 310
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (468 KB)  

    Neural networks are used to extract relevant features of sign language from video images of a person communicating in American Sign Language or Signed English. The key features are hand motion, hand location with respect to the body, and handshape. A modular design is under way to apply various techniques, including neural networks, in the development of a translation system that will facilitate communication between deaf and hearing people. Signal processing techniques developed for defense-related programs have been adapted and applied to this project. Algorithm development and transition using neural network architectures has been encouraging. The results of the feasibility study for this project are described View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Competitive learning and winning-weighted competition for optimal vector quantizer design

    Page(s): 50 - 59
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (404 KB)  

    It is essential to build a nonparametric model to estimate a probability density function p(x) in the areas of vector quantization, pattern recognition, control, and many others. A generalization of Kohonen learning, the winning-weighted competitive learning (WWCL), is presented for a better approximation of p(x) and fast learning convergence by introducing the principle of maximum information preservation into the learning. The WWCL is a promising alternative and improvement to the generalized Lloyd algorithm (GLA) which is an iterative descent algorithm with a monotonically decreasing distortion function towards a local minimum. The WWCL is an online algorithm where the codebook is designed while training data is arriving and the reduction of the distortion function is not necessarily monotonic. Experimental results show that the WWCL consistently provides better codebooks than the Kohonen learning and the GLA in distortion or convergence rate View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Text-Dependent speaker verification using recurrent time delay neural networks for feature extraction

    Page(s): 353 - 361
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (304 KB)  

    The possible application of time delay neural network (TDNN) to the text-dependent speaker verification problem is described and evaluated. Each person to be verified has a personalized neural network, which is trained to extract representative feature vector of the speaker by a particular utterance. A novel model called recurrent time delay neural networks is investigated. The training is carried out by backpropagation for sequence (BPS)-a variant of the BP algorithm. The modified structure is shown to outperform both a multilayer perceptron classifier and the original TDNN for feature extraction View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A modular neural network architecture for pattern classification

    Page(s): 232 - 238
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (248 KB)  

    A modular neural network architecture is proposed to classify binary and continuous patterns. This system consists of a supervised feedforward backpropagation network and an unsupervised self-organization map network. The supervised feedforward (basic) network is trained until a saturation error level occurs. Simultaneously, the unsupervised self-organization map (control) network fluids the mapping features for the given input/output patterns. The resultant features are used by Gaussian and linear functions to adjust the hidden and the output weights of the basic network and to classify the given patterns View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A common framework for snakes and Kohonen networks

    Page(s): 251 - 260
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (492 KB)  

    The relationship between snake active contours and Kohonen networks in the scope of object boundary detection is discussed. It is shown that these algorithms share many common features and that they are special cases of a more general structure proposed, which provides a common framework for their study and enables the design of new iterative algorithms for boundary extraction View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A neural net application to signal identification

    Page(s): 549 - 559
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (276 KB)  

    Results from an ongoing research project concerning detection and identification of signals in a relatively quiescent background are discussed. The signals to be identified are of varying duration, and the number of different signal classes is large (> 1,000) and is frequently expanded or reduced. The author's approach to solving this real-world problem is presented, and initial results are shown View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Compressing moving pictures using the APEX neural principal component extractor

    Page(s): 321 - 330
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (444 KB)  

    An application of the optimal Karhunen-Loe`ve transform (KLT) in place of the traditional discrete cosine transform (DCT) for compressing intra-frames in the MPEG protocol is proposed. The I-frames attain the smallest compression ratio since they are coded without reference to any other frames. The difficulty of KLT (additional bit-rate is required to make the image-dependent transform basis known to the decoder) is overcome by using the KLT basis of the previous I- or P-frame which, the authors argue, is very similar to the basis of the current frame. The previous frame is already known to the decoder. Therefore, no additional information needs to be sent out. Paying attention to the nonstationary statistics of image they propose to split the I-frames in N parts and use a dedicated transform basis for each part. Since the KLT basis must be updated continuously they use the adaptive principal component extractor network (APEX) that can incrementally estimate the new basis for the next frame View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Modeling the spectral transition selectivity in the primary auditory cortex

    Page(s): 98 - 107
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (508 KB)  

    From an information processing point of view, a common computational principle in neural systems is to analyze and map the features of the input signal into multidimensional, spatially organized areas. A mathematical model based on biological data mimicking the functions of the primary auditory cortex is presented. The model embraces a three dimensional representation which corresponds respectively to the following acoustic features: frequency components, local spectral shape and local spectral bandwidth. The model representation of nonstationary acoustic features is examined, specifically, the direction and the rate of spectral transitions, which are crucial in speech and sound perception View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Self-organizing feature map with position information and spatial frequency information

    Page(s): 40 - 49
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (436 KB)  

    The representation of scale - position information of local features in images is discussed. A feature map is proposed on which scale information is represented in position explicitly in the same manner of position information. The result of a computer simulation to self-organize such a feature map by using a retina-like filter set in the case of one-dimensional input data is shown. This feature map seems useful for scale invariant image recognition View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Discriminative feature extraction for speech recognition

    Page(s): 392 - 401
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (472 KB)  

    A novel approach to pattern recognition, called discriminative feature extraction (DFE) is introduced as a way to interactively handle the input data with a given classifier. The entire recognizer, consisting of the feature extractor as well as the classifier, is trained with the minimum classification error generalised probabilistic descent learning algorithm. Both the philosophy and implementation examples of this approach are described. DFE realizes a significant departure from conventional approaches, providing a comprehensive base for the entire system design. By way of example, an automatic scaling process is described, and experimental results for designing a cepstrum representation for vowel recognition are presented View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Phoneme recognition based on multi-resolution and non-causal context

    Page(s): 343 - 352
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (480 KB)  

    An alternative view of neural-network-based phoneme recognition using multiresolution ideas and noncausal context is suggested. Some suggestions are made regarding target and error weight functions to improve performance and simplify training. Based on these observations, a simple network with self recurrent links of different delays is proposed and tested on the task of speaker- independent recognition of unvoiced plosives, (p,t,k), with input feature vectors derived from an auditory model View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Target recognition using multiple sensors

    Page(s): 411 - 420
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (536 KB)  

    A novel approach to multisensor target recognition is presented. Currently available multisensor recognition algorithms/systems have low recognition rates when tested in battlefield conditions. The authors' approach make no assumptions on either sensors or targets, and uses some biologically inspired algorithms to build a multisensor target recognition system View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Fuzzy decision neural networks and application to data fusion

    Page(s): 171 - 180
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (476 KB)  

    A decision-based neural network (DBNN) is extended to a fuzzy-decision neural network (FDNN), which is shown to offer classification/generalization performance improvements, especially when the data are not clearly separable. The hierarchical structure adopted make the computation process very efficient. The learning rule and some key properties of FDNN are described. A Bayesian paradigm offers an optimal approach to data fusion. This approach is explored. DBNN, together with a Bayesian approach, is proposed to formulate the data fusion process View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • LS-based training algorithm for neural networks

    Page(s): 22 - 29
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (580 KB)  

    A new training algorithm is presented as a faster alternative to the backpropagation (BP) method. The new approach is based on the solution of a linear system at each step of the learning phase. The squared error at the output of each layer before the nonlinearity is minimized on the entire set of the learning patterns by a block least squares (LS) algorithm. The optimal weights for each layer are then computed by using the singular value decomposition (SVD) technique. The simulation results show considerable improvements from the point of view of both accuracy and speed of convergence View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Nonlinear predictive vector quantisation with recurrent neural nets

    Page(s): 372 - 381
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (528 KB)  

    The nonlinear prediction capability of neural nets is applied to the design of improved predictive speech coders. Performance evaluations and comparisons with linear predictive speech coding are presented. These tests show the applicability of nonlinear prediction to speech coding and the improvement in coding performance View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Evaluation of character recognition systems

    Page(s): 485 - 496
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (536 KB)  

    Eleven different Census Optical Character Recognition Systems systems are evaluated using correlations between the answers of different systems, comparing the decrease in error rate as a function of confidence of recognition, and comparing the writer dependence of recognition. This comparison shows that methods that use different algorithms for feature extraction and recognition perform with very high levels of correlation View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Differentially generated neural network classifiers are efficient

    Page(s): 151 - 160
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (512 KB)  

    Differential learning for statistical pattern classification is based on the classification figure-of-merit (CFM) objective function. It is proved that differential learning is asymptotically efficient, guaranteeing the best generalization allowed by the choice of hypothesis class as the training sample size grows large, while requiring the least classifier complexity necessary for Bayesian (i.e., minimum probability-of-error) discrimination. Differential learning almost always guarantees the best generalization allowed by the choice of hypothesis class for small training sample sizes View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Characterization of network responses to known, unknown, and ambiguous inputs

    Page(s): 226 - 231
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (328 KB)  

    Neural networks typically classify patterns by mapping the input feature space to the corners of the m-dimensional unit hypercube where m is the number of output classes. When classifier networks of graded threshold neurons are presented with patterns that are strong, ambiguous, or unknown, characteristic responses are emitted. A second tier network can be used to characterize the decision of the classifier network. By using a class count-independent mapping as an intermediary, a training set of unknowns can be generated for training the second tier. The two tiered approach is described and the internal function of the second tier is analyzed View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A growing and splitting elastic network for vector quantization

    Page(s): 281 - 290
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (444 KB)  

    A new vector quantization method is proposed which incrementally generates a suitable codebook. During the generation process, new vectors are inserted in areas of the input vector space where the quantization error is especially high. A one-dimensional topological neighborhood makes it possible to interpolate new vectors from existing ones. Vectors not contributing to error minimization are removed. After the desired number of vectors is reached, a stochastic approximation phase fine tunes the codebook. The final quality of the codebooks is exceptional. A comparison with two methods for vector quantization is performed by solving an image compression problem. The results indicate that the new method is clearly superior to both other approaches View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Analysis of coarse parallel architectures for artificial neural processing

    Page(s): 450 - 459
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (692 KB)  

    A methodology for comparing various neural architectures and implementations is illustrated. The methodology consists of writing the artificial neural network (ANN) equations in a summation form and the applying a tool termed algorithmic timing parameter decomposition (ATPD). ATPD decomposes an algorithm or set of equations into a computation time formula comprising basic system primitives. A particular architecture has a corresponding computational time formula. Similarly, the primitive elements are dependent on the actual hardware realization and thus will change with the processor used in the system. Computation times therefore can be estimated for different parallel architectures. Implementation of a multilayer perceptron is analyzed in several digital signal processor (DSP)-based parallel architectures View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • Quantized, piecewise linear filter network

    Page(s): 470 - 474
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (180 KB)  

    A quantization based piecewise linear filter network is defined. A method for the training of this network based on local approximation in the input space is devised. The training is carried out by repeatedly alternating between vector quantization of the training set into quantization classes and equalization of the quantization classes linear filter mean square training errors. The equalization of the mean square training errors is carried out by adapting the boundaries between neighbor quantization classes such that the differences in mean square training errors are reduced View full abstract»

    Full text access may be available. Click article title to sign in or learn about subscription options.
  • A Lyapunov function for additive neural networks and nonlinear integral equations of Hammerstein type

    Page(s): 11 - 13
    Save to Project icon | Request Permissions | Click to expandQuick Abstract | PDF file iconPDF (140 KB)  

    Using the properties of the nonlinear integral equations of the Hammerstein type, a new Lyapunov function for additive neural networks is constructed. The function does not require monotonicity of the transfer function as does the previously discovered Lyapunov function for the additive networks. Instead positivity of the symmetric part of the weight matrix is required. The results on the Hammerstein equation also allow one to provide simple criteria for estimation of the number of fixed points and their bifurcation. The criteria combine the spectral properties of the weight matrix and the growth properties of the transfer function View full abstract»

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