By Topic

Neural Networks for Signal Processing Proceedings of the 1991 IEEE Workshop

Filter Results

Displaying Results 1 - 25 of 64
• Tutorial: digital neurocomputing for signal/image processing

Publication Year: 1991, Page(s):616 - 644
Cited by:  Papers (4)
| | PDF (1110 KB)

The requirements on both the computations and storage for neural networks are extremely demanding. Neural information processing would be practical only when efficient and high-speed computing hardware can be made available. The author reviews several approaches to architecture and implementation of neural networks for signal and image processing. The author discusses direct design of dedicated ne... View full abstract»

• Segment-based speaker adaptation by neural network

Publication Year: 1991, Page(s):442 - 451
Cited by:  Papers (3)
| | PDF (370 KB)

The authors propose a segment-to-segment speaker adaptation technique using a feed-forward neural network with a time shifted sub-connection architecture. Differences in voice individuality exist in both the spectral and temporal domains. It is generally known that frame based speaker adaptation techniques can not compensate for speaker individuality in the temporal domain. Segment based speaker a... View full abstract»

• Neural Networks for Signal Processing. Proceedings of the 1991 IEEE Workshop (Cat. No.91TH0385-5)

Publication Year: 1991
| PDF (27 KB)
• New discriminative training algorithms based on the generalized probabilistic descent method

Publication Year: 1991, Page(s):299 - 308
Cited by:  Papers (91)  |  Patents (2)
| | PDF (416 KB)

The authors developed a generalized probabilistic descent (GPD) method by extending the classical theory on adaptive training by Amari (1967). Their generalization makes it possible to treat dynamic patterns (of a variable duration or dimension) such as speech as well as static patterns (of a fixed duration or dimension), for pattern classification problems. The key ideas of GPD formulations inclu... View full abstract»

• Probability estimation by feed-forward networks in continuous speech recognition

Publication Year: 1991, Page(s):309 - 318
Cited by:  Papers (6)
| | PDF (436 KB)

The authors review the use of feedforward neural networks as estimators of probability densities in hidden Markov modelling. In this paper, they are mostly concerned with radial basis functions (RBF) networks. They not the isomorphism of RBF networks to tied mixture density estimators; additionally they note that RBF networks are trained to estimate posteriors rather than the likelihoods estimated... View full abstract»

• Nonlinear resampling transformation for automatic speech recognition

Publication Year: 1991, Page(s):319 - 326
Cited by:  Papers (3)
| | PDF (260 KB)

A new technique for speech signal processing called nonlinear resampling transformation (NRT) is proposed. The representation of a speech pattern derived from this technique has two important features: first, it reduces redundancy; second, it effectively removes the nonlinear variations of speech signals in time. The authors have applied NRT to the TI isolated-word database achieving a 99.66% reco... View full abstract»

• Speech recognition by combining pairwise discriminant time-delay neural networks and predictive LR-parser

Publication Year: 1991, Page(s):327 - 336
Cited by:  Papers (1)
| | PDF (416 KB)

A phoneme recognition method using pairwise discriminant time-delay neural networks (PD-TDNNs) and a continuous speech recognition method using the PD-TDNNs are proposed. It is shown that classification-type neural networks have poor robustness against the difference in speaking rates between training data and testing data. To improve the robustness, the authors developed a phoneme recognition met... View full abstract»

• Improved structures based on neural networks for image compression

Publication Year: 1991, Page(s):493 - 502
Cited by:  Papers (5)
| | PDF (436 KB)

The problem of efficient image compression through neural networks (NNs) is addressed. Some theoretical results on the application of 2-layer linear NNs to this problem are given. Two more elaborate structures, based on a set of NNs, are further presented; they are shown to be very efficient while remaining computationally rather simple View full abstract»

• Edge detection for optical image metrology using unsupervised neural network learning

Publication Year: 1991, Page(s):188 - 197
Cited by:  Papers (1)  |  Patents (2)
| | PDF (492 KB)

Several unsupervised neural network learning methods are explored and applied to edge detection of microlithography optical images. Lack of a priori knowledge about correct state assignments for learning procedure in optical microlithography environment makes the metrology problem a suitable area for applying unsupervised learning strategies. The methods studied include a self-organizing competiti... View full abstract»

• Speech recognition using time-warping neural networks

Publication Year: 1991, Page(s):337 - 346
Cited by:  Patents (53)
| | PDF (396 KB)

The author proposes a time-warping neural network (TWNN) for phoneme-based speech recognition. The TWNN is designed to accept phonemes with arbitrary duration, whereas conventional phoneme recognition networks have a fixed-length input window. The purpose of this network is to cope with not only variability of phoneme duration but also time warping in a phoneme. The proposed network is composed of... View full abstract»

• Nonlinear prediction of speech signals using memory neuron networks

Publication Year: 1991, Page(s):395 - 404
Cited by:  Papers (7)  |  Patents (3)
| | PDF (364 KB)

The authors present a feed-forward neural network architecture that can be used for nonlinear autoregressive prediction of multivariate time-series. It uses specialized neurons (called memory neurons) to store past activations of the network in an efficient fashion. The network learns to be a nonlinear predictor of the appropriate order to model temporal waveforms of speech signals. Arrays of such... View full abstract»

• Adaptive neural filters

Publication Year: 1991, Page(s):503 - 512
Cited by:  Papers (11)  |  Patents (1)
| | PDF (436 KB)

The authors introduce a new class of nonlinear filters called neural filters based on the threshold decomposition and neural networks. Neural filters can approximate both linear FIR filters and weighted order statistic (WOS) filters which include median, rank order, and weighted median filters. An adaptive algorithm is derived for determining optimal neural filters under the mean squared error (MS... View full abstract»

• Improving generalization performance in character recognition

Publication Year: 1991, Page(s):198 - 207
Cited by:  Papers (1)
| | PDF (340 KB)

One test of a new training algorithm is how well the algorithm generalizes from the training data to the test data. A new neural net training algorithm termed double backpropagation improves generalization in character recognition by minimizing the change in the output due to small changes in the input. This is accomplished by minimizing the normal energy term found in backpropagation and an addit... View full abstract»

• Three-dimensional structured networks for matrix equation solving

Publication Year: 1991, Page(s):80 - 89
| | PDF (380 KB)

Structured networks are feedforward neural networks with linear neurons than use special training algorithms. Two three-dimensional (3-D) structured networks are developed for solving linear equations and the Lyapunov equation. The basic idea of the structured network approaches is: first, represent a given equation-solving problem by a 3-D structured network so that if the network matches a desir... View full abstract»

• A hybrid continuous speech recognition system using segmental neural nets with hidden Markov models

Publication Year: 1991, Page(s):347 - 356
Cited by:  Papers (2)  |  Patents (1)
| | PDF (448 KB)

The authors present the concept of a segmental neural net' (SNN) for phonetic modeling in continuous speech recognition (CSR) and demonstrate how than can be used with a multiple hypothesis (or N-Best) paradigm to combine different CSR systems. In particular, they have developed a system that combines the SNN with a hidden Markov model (HMM) system. They believe that this is the first system inco... View full abstract»

• Fingerprint recognition using neural network

Publication Year: 1991, Page(s):226 - 235
Cited by:  Papers (12)  |  Patents (3)
| | PDF (420 KB)

The authors describe a neural network based approach for automated fingerprint recognition. Minutiae are extracted from the fingerprint image via a multilayer perceptron (MLP) classifier with one hidden layer. The backpropagation learning technique is used for its training. Selected features are represented in a special way such that they are simultaneously invariant under shift, rotation and scal... View full abstract»

• Experiments with temporal resolution for continuous speech recognition with multi-layer perceptrons

Publication Year: 1991, Page(s):405 - 410
| | PDF (316 KB)

Previous work by the authors focused on the integration of multilayer perceptrons (MLP) into hidden Markov models (HMM) and on the use of perceptual linear prediction (PLP) parameters for the feature inputs to such nets. The system uses the Viterbi algorithm for temporal alignment. This algorithm is a simple and optimal procedure, but it necessitates a frame-based analysis in which all features ha... View full abstract»

• A surface reconstruction neural network for absolute orientation problems

Publication Year: 1991, Page(s):513 - 522
Cited by:  Papers (4)
| | PDF (368 KB)

The authors propose a neural network for representation and reconstruction of 2-D curves or 3-D surfaces of complex objects with application to absolute orientation problems of rigid bodies. The surface reconstruction network is trained by a set of roots (the points on the curve or the surface of the object) via forming a very steep cliff between the exterior and interior of the surface, with the ... View full abstract»

• Workstation-based phonetic typewriter

Publication Year: 1991, Page(s):279 - 288
| | PDF (460 KB)

The author presents a general description of his phonetic typewriter' system that transcribes unlimited speech into orthographically correct text. The purpose of this paper is to motivate certain choices made in the partitioning of the problem into tasks and describe their implementation. The combination of algorithms he has selected has proven effective for well-articulated dictation in a phonem... View full abstract»

• A relaxation neural network model for optimal multi-level image representation by local-parallel computations

Publication Year: 1991, Page(s):473 - 482
| | PDF (828 KB)

A relaxation neural network model is proposed to solve the multi-level image representation problem by energy minimization in local and parallel computations. This network iteratively minimizes the computational energy defined by the local error in neighboring picture elements. This optimization method can generate high quality binary and multi-level images depending on local features, and can be ... View full abstract»

• Neural networks for sidescan sonar automatic target detection

Publication Year: 1991, Page(s):208 - 216
Cited by:  Papers (2)
| | PDF (376 KB)

The goal of this research is to develop a multi-layer feedforward neural network architecture which can distinguish targets (in this case, mines) from background clutter in sidescan sonar images. The network is to be implemented on a hardware neurocomputer currently in development at CSDL, with the goal of eventual real-time performance in the field. A variety of neural network architectures are d... View full abstract»

• A simple word-recognition network with the ability to choose its own decision criteria

Publication Year: 1991, Page(s):452 - 459
Cited by:  Papers (1)
| | PDF (268 KB)

Various reliable algorithms for the word classification problem have been developed. All these models are necessarily based on the classification of certain `features' that have to be extracted from the presented word. The general problem in speech recognition is: what kind of features are both word dependent as well as speaker independent? The majority of the existing systems requires a feature s... View full abstract»

• Nonlinear adaptive filtering of systems with hysteresis by quantized mean field annealing

Publication Year: 1991, Page(s):151 - 160
| | PDF (336 KB)

A technique for nonlinear adaptive filtering of systems with hysteresis has been developed which combines quantized mean field annealing (QMFA) and conventional RLS/FTF adaptive filtering. Hysteresis is modeled as a nonlinear system with memory. Unlike other methods which rely on Volterra and Wiener models, this technique can efficiently handle large order nonlinearities with or without hysteresis... View full abstract»

• A parallel learning filter system that learns the KL-expansion from examples

Publication Year: 1991, Page(s):121 - 130
| | PDF (384 KB)

A new method for learning in a single-layer linear neural network is investigated. It is based on an optimality criterion that maximizes the information in the outputs and simultaneously concentrates the outputs. The system consists of a number of so-called basic units and it is shown that the stable states of these basic units correspond to the (pure) eigenvectors of the input correlation matrix.... View full abstract»

• Improving learning rate of neural tree networks using thermal perceptrons

Publication Year: 1991, Page(s):90 - 100
Cited by:  Papers (1)
| | PDF (428 KB)

A new neural network called the neural tree network (NTN) is a combination of decision trees and multi-layer perceptrons (MLP). The NTN grows the network as opposed to MLPs. The learning algorithm for growing NTNs is more efficient that standard decision tree algorithms. Simulation results have shown that the NTN is superior in performance to both decision trees and MLPs. A new NTN learning algori... View full abstract»