George G. Coghill

Channel assignment through evolutionary optimization

The problem of assigning appropriate channels to the individual members of a cellular network is an important challenge facing network designers. Heuristics may be used to solve this problem, although in recent years parallel distributed methods have also been suggested. We investigate how an evolutionary inspired computing technique known as genetic algorithms (GAs) may be used. These global optimization techniques avoid many of the shortcomings exhibited by local search techniques on difficult search spaces. The new approach is tested on several problems of different sizes and complexity. The critical aspects of this technique and additional improvements are also discussed.

Text analysis using local energy

This paper presents a biologically inspired texture-based algorithm using local energy analysis for (1) segmenting text embedded in clutter and (2) classifying text scripts without any explicit knowledge of the type of text present. The local energy model has been shown to work well in texture analysis, where texture segmentation and discrimination are preattentive tasks in human vision. The algorithm for text segmentation involves computing the local energy using a bank of orthogonal pairs of Gabor filters at various orientation and frequency. Similarly, local energy analysis is applied to the task of text script classification using a set of descriptors derived from the local energy information. The segmentation algorithm is quite successful in segmenting text of any arbitrary language script from real images with or without noise. It is also invariant to scale, rotation and position changes in text. The classification scheme was tested on 16 languages with the results obtained consistent with visual classification performed by humans. The scheme is insensitive to orientation of the text script.

Neural Network Implementation Using Bit Streams

A new method for the parallel hardware implementation of artificial neural networks (ANNs) using digital techniques is presented. Signals are represented using uniformly weighted single-bit streams. Techniques for generating bit streams from analog or multibit inputs are also presented. This single-bit representation offers significant advantages over multibit representations since they mitigate the fan-in and fan-out issues which are typical to distributed systems. To process these bit streams using ANNs concepts, functional elements which perform summing, scaling, and squashing have been implemented. These elements are modular and have been designed such that they can be easily interconnected. Two new architectures which act as monotonically increasing differentiable nonlinear squashing functions have also been presented. Using these functional elements, a multilayer perceptron (MLP) can be easily constructed. Two examples successfully demonstrate the use of bit streams in the implementation of ANNs. Since every functional element is individually instantiated, the implementation is genuinely parallel. The results clearly show that this bit-stream technique is viable for the hardware implementation of a variety of distributed systems and for ANNs in particular.

Genetic breeding of control parameters for the Hopfield/Tank neural net

The authors show how genetic algorithms may be used in conjunction with the Hopfield/Tank neural net by breeding an effective set of control parameters in the parameter sub-space to be used by the artificial neural network. They briefly consider the standard Hopfield/Tank neural net followed by a discussion of the genetic algorithm used with this network. Some of the more important operators used with the genetic algorithm are discussed, and the use of this hybrid neural net on two sets of traveling salesman problems of different sizes is illustrated.<<ETX>>

Excessive Noise Injection Training of Neural Networks for Markerless Tracking in Obscured and Segmented Environments

In this letter, we demonstrate that the generalization properties of a neural network (NN) can be extended to encompass objects that obscure or segment the original image in its foreground or background. We achieve this by piloting an extension of the noise injection training technique, which we term excessive noise injection (ENI), on a simple feedforward multilayer perceptron (MLP) network with vanilla backward error propagation to achieve this aim. Six tests are reported that show the ability of an NN to distinguish six similar states of motion of a simplified human figure that has become obscured by moving vertical and horizontal bars and random blocks for different levels of obscuration. Four more extensive tests are then reported to determine the bounds of the technique. The results from the ENI network were compared to results from the same NN trained on clean states only. The results pilot strong evidence that it is possible to track a human subject behind objects using this technique, and thus this technique lends itself to a real-time markerless tracking system from a single video stream.

Factor selection strategies for orthogonal signal correction applied to calibration of near-infrared spectra

Orthogonal signal correction (OSC) is a preprocessing technique used for correction of instrumental drift, bias and scatter in near-infrared spectra. OSC separates the variation into orthogonal factors, where the factors contain the variation within the X matrix data that is not correlated with the y matrix vector data. The aim of this study is to investigate different orthogonal factor selection methods, which will enhance the performance of the OSC routine for quantitative analysis of near-infrared spectra. In order for factor selection methods to be applied to OSC, an implementation of an existing OSC algorithm is used; this method computes OSC factors in a principal component manner. A binarized weighting matrix is then applied to the OSC factors for the purpose of OSC factor subset selection/nonselection. The optimization strategies used for subset selection of OSC factors were a genetic algorithm and stepwise selection. Three data sets were formed: (1) no preprocessing, (2) preprocessing by removal of sequential OSC factors and (3) preprocessing by removal of an OSC factor subset. Combinations of spectral predictors were selected from these data sets by hill climbing, feature selection, genetic algorithm and full-spectrum modeling. Partial least squares regression was undertaken to form calibration models. It was found that selection of OSC factor subsets produced better standard errors of prediction relative to data preprocessed by sequential selection of OSC factors.

A mapping neural network and its application to voiced-unvoiced-silence classification

A mapping neural network which combines unsupervised and supervised training is described and its application to the classification of segments of speech to voiced, unvoiced, and silence (V-UV-S) is demonstrated through computer simulations. The network uses a dynamic variation of competitive learning in the unsupervised layer followed by a supervised associative layer. When used to solve the V-UV-S classification problem, the network outperforms a network based on the frequency sensitive competitive learning.<<ETX>>

Extended genetic algorithm for codesign optimization of DSP systems in FPGAs

The multiobjective genetic algorithm is an effective solution to the complex problem of hardware-software codesign. An extended genetic algorithm (EGA) has been developed that implements a novel selection method with function scaling, adaptive crossover and mutation. This EGA is applied in a codesign optimization stage for dataflow oriented applications and synthesis on field-programmable gate arrays (FPGAs). Its effectiveness is illustrated on the problem of codesign of a self-tuning regulator considering area and performance.

A mapping neural network using unsupervised and supervised training

A two-layer mapping neural network called an extended differentiator network (EDN) is described. The network uses both unsupervised and supervised training in two phases. The differentiator, which is an unsupervised pattern classifier, is followed by the supervised outstar structure of Grossberg. This makes a network somewhat similar to the counterpropagation network of R. Hecht-Nielsen (1987). The unsupervised training of the input patterns by the differentiator provides useful information for the subsequent layer of the network and thus the associations with the target vectors are learned rapidly. As a result, some complex mappings are realizable by the network. The operation of the EDN is demonstrated by some simulation examples.<<ETX>>

Unsupervised clustering and centroid estimation using dynamic competitive learning

In this paper, an unsupervised learning algorithm is developed. Two versions of an artificial neural network, termed a differentiator, are described. It is shown that our algorithm is a dynamic variation of the competitive learning found in most unsupervised learning systems. These systems are frequently used for solving certain pattern recognition tasks such as pattern classification and k-means clustering. Using computer simulation, it is shown that dynamic competitive learning outperforms simple competitive learning methods in solving cluster detection and centroid estimation problems. The simulation results demonstrate that high quality clusters are detected by our method in a short training time. Either a distortion function or the minimum spanning tree method of clustering is used to verify the clustering results. By taking full advantage of all the information presented in the course of training in the differentiator, we demonstrate a powerful adaptive system capable of learning continuously changing patterns.