Y.F. Huang

AMUSE: a new blind identification algorithm

The mathematical formulation of the blind identification problem is presented. Various theoretical properties are discussed. The AMUSE algorithm is derived on the basis of the necessary condition of source identifiability and shown to have good performance and wide application.<<ETX>>

An extended fourth order blind identification algorithm in spatially correlated noise

An eigenstructure-based method called extended fourth-order blind identification (EFOBI) is presented for the problem of blind decomposition of multiple source signals in spatially correlated noise. Estimates of the source signals and the unknown model parameters are computed by the algorithm. The method is applied to a speech enhancement problem and to the direction-of-arrival problem in array signal processing under sensor gain uncertainties and shows improved performance over traditional methods such as MUSIC.<<ETX>>

Bounds on number of hidden neurons of multilayer perceptrons in classification and recognition

The use of multilayer perceptrons (MLP) in the realization of arbitrary functions which map from a finite subset of E/sup n/ into E/sup m/ is investigated. A least upper bound of hidden neurons needed to solve this problem is derived. It is shown that as long as the number of hidden neurons exceeds this bound, an MLP can realize arbitrary switching functions without requiring learning algorithms. In studying classification problems, an upper bound which is tighter than the ones obtained with the common assumption of the general position condition on the input set is derived. In addition, a lower bound is derived in addressing recognition problems.<<ETX>>

A necessary and sufficient condition for the blind identification of memoryless systems

The blind identification of multiple-input and multiple-output memoryless systems is considered. It is found that the system is identifiable with respect to uncorrelated sources if and only if the source signals have a distinct spectrum. A blind identification algorithm that is computationally simpler than previous ones is developed.<<ETX>>

A novel approach to the convergence of neural networks for signal processing

A novel deterministic approach to the convergence analysis of (stochastic) learning algorithms is presented. The link between the two is the new concept of time-average invariance, which is a property of deterministic signals but resembles that of stochastic signals which are ergodic and stationary. >

A neural network for blind signal separation

An unsupervised neural network is constructed for the problem of blind signal separation. It is designed based on the condition that the outputs of the neural network are independent. A study of the stability of the neural network in the sense of expectation is presented. A stability condition on the system matrix A is obtained. Simulation studies show that this neural network is robust. It can separate two signals with strength ratio 100:1.<<ETX>>

A wideband blind identification approach to speech acquisition using a microphone array

The problem of speech separation from an array of microphones is considered. The problem is modeled as a wideband process with unknown model parameters, namely, sensor gains and time delays. The technique of blind identification is extended to the wideband case. It is shown that the wideband blind identification approach is capable of handling multipath situations. Simulations are presented which show that the wideband signals can indeed be separated by the proposed approach.<<ETX>>

Greedy tree growing algorithms for designing variable rate vector quantizers

The performance of vector quantization for image compression can be improved by using a variable-rate code which is able to designate more bits to regions of an image that are active or difficult to code, and fewer bits to less active regions. Two schemes are presented for directly designing variable-rate tree-structured vector quantizers by growing the tree one node at a time. One involves selecting the node with the largest average distortion within one node to split. The other involves splitting the node with the largest eigenvalue of the input covariance matrix. A comparison with the scheme of E.A. Riskin and R.M. Gray (1991) shows that the proposed schemes perform better in terms of visual quality with reduced complexity.<<ETX>>

A neural network structure for vector quantizers

A variable perturbation method for codebook design in vector quantization (VQ) is proposed. The resulting codebook can be used as the initial codebook in the implementation of the LBG (Line, Buzo, Gray, 1990) VQ algorithm. The proposed method is based on the concept of entropy implemented as a learning algorithm for a feedforward neural network. Such a neural network with the proposed learning algorithm can construct a codebook for input vectors with an unknown distribution without memorizing long training data.<<ETX>>

Principal component vector quantization for abrupt scene changes

The authors present a vector quantization technique, referred to as principal component vector quantization (PCVQ), and investigate the problem of abrupt scene changes of video signals. This technique, featuring a simple design procedure, is implemented by an artificial neural network with a learning algorithm for online codebook design based on the local statistics for each difference scene. This network features online learning and a constant encoding time that is independent of the codebook size. The PCVQ was examined via simulation on an abrupt scene change problem which had nonstationary training sequences.<<ETX>>