Learning Vector Quantization Network Research Articles

Multiple competitive learning network fusion for object classification

This paper introduces a multiple competitive learning neural network fusion method for pattern recognition. By defining a confidence level measure for the learning vector quantization network classifier, we develop both a serial and a parallel network fusion algorithm to combine the discriminatory ability of different individually trained networks. We use two distinct feature vectors, gray-scale morphological granulometry and Fourier boundary descriptor, to demonstrate the efficacy of the classifier. The algorithms are applied on the classification of more than 8000 underwater plankton images. The classification accuracy for training data and for testing data are over 92% and 85%, respectively.

A three-dimensional neural network model for unconstrained handwritten numeral recognition: a new approach

The paper describes a three-dimensional (3-D) neural network recognition system for conflict resolution in recognition of unconstrained handwritten numerals. This neural network classifier is a combination of modified self-organizing map (MSOM) and learning vector quantization (LVQ). The 3-D neural network recognition system has many layers of such neural network classifiers and the number of layers forms the third dimension. The Experiments are conducted employing SOM, MSOM, SOM and LVQ, and MSOM and LVQ networks. These experiments on a database of unconstrained handwritten samples show that the combination of MSOM and LVQ performs better than other networks in terms of classification, recognition and training time. The 3-D neural network eliminates the substitution error.

Automatic Plankton Image Recognition

Plankton form the base of the food chain in the ocean and are fundamental to marine ecosystem dynamics. The rapid mapping of plankton abundance together with taxonomic and size composition is very important for ocean environmental research, but difficult or impossible to accomplish using traditional techniques. In this paper, we present a new pattern recognition system to classify large numbers of plankton images detected in real time by the Video Plankton Recorder (VPR), a towed underwater video microscope system. The difficulty of such classification is compounded because: 1) underwater images are typically very noisy, 2) many plankton objects are in partial occlusion, 3) the objects are deformable and 4) images are projection variant, i.e., the images are video records of three-dimensional objects in arbitrary positions and orientations. Our approach combines traditional invariant moment features and Fourier boundary descriptors with gray-scale morphological granulometries to form a feature vector capturing both shape and texture information of plankton images. With an improved learning vector quantization network classifier, we achieve 95% classification accuracy on six plankton taxa taken from nearly 2,000 images. This result is comparable with what a trained biologist can achieve by using conventional manual techniques, making possible for the first time a fully automated, at sea-approach to real-time mapping of plankton populations.

Aspect-independent classification of ‘‘dolphin’’ ensonified mines using Choi–Williams representations

Contemporary anti-invasion mines have aspect-dependent shapes, making discrimination of mines from nonthreat objects, and classification of mines, a difficult task for human mine countermeasures personnel. Shallow-water (SW) noise and bottom reverberation substantially degrade the acoustic structure of mine echoes. Traditional MCM target strength and FFT classifiers are not effective under these conditions. The testing of a novel neural network classifier to solve the task of classifying mines in the SW acoustic environment was begun. Three mine types were ensonified at 1 deg rotations using synthetic dolphin clicks. A learning vector quantization network was trained to classify the mines using the Choi–Williams joint time-frequency distributions of echoes. A training set of 36 echoes per mine (5, 15, 25,..., 355 deg) was created, with the remaining echoes (0, 10, 20,..., 350 deg) reserved for generalization testing. The network correctly classified the mines at novel orientations with 85% accuracy. Performance using CWD will be contrasted with biomimetic sonar target classifiers under varying signal-to-noise ratios. [Research funded by ONR Underseas Active Signal Processing Program.]

Localization of Myocardial Infarction Based on Learning Vector Quantization Networks Applied to ST Elevations of the 12‐Lead ECG

BackgroundDuring recent years artificial neural networks have been proposed as a diagnostic tool in different fields of cardiology. Most of the studies have utilized the multilayer perceptron with backpropagation learning rule for the design of the network. As a new approach, Learning Vector Quantization (LVQ) which belongs to the class of competitive learning networks, was developed particularly for classification problems. So far there are no data available on the application of LVQ for classification tasks in cardiology. The present study aims at investigating the performance of LVQ for localization of myocardial infarction (Ml) based on ST elevations in the standard 12‐lead ECG.MethodsAltogether, 769 male patients (age 53 ± 7 years) with an acute Ml were included into the study. Three hundred fifty‐three patients (46%) presented with anterior and 416 patients (54%) with inferior Ml based on typical changes in the standard 12‐lead ECG. Standardized ST elevations in all 12 leads were used as input structure for the network. The performance of the network was studied using two different learning and test sets. The influence of the number of reference vectors and training steps on the classification accuracy for infarct location was investigated.ResultsThe highest classification accuracy of 88.6% for infarct location was achieved using the learning set with 66% of all patients. This setup was based on five reference vectors and 200 training steps. The best accuracies for anterior Ml were higher as compared to inferior infarctions in both the test and training set. Using more than 50 reference vectors resulted in a decrease of classification accuracy due to overtraining of the network.ConclusionAppropriately initialized and trained artificial neural networks based on LVQ give a high accuracy for localization of Ml using only ST elevations of the standard 12‐lead ECG.

Featureless classification of tactile contacts in a gripper using neural networks

A direct, featureless process to classify contact impressions of objects gripped by a robot hand is presented. The inforlnation about the type of contact allows the selection of the most appropriate manipulating strategy to handle the grasped object. A learning vector quantization (LVQ) network is applied introducing a contact-pattern preprocessing technique to improve the robustness of tbe classification with respect to the pattern variations in position, orientation and size.

Comparison of the BP training algorithm and LVQ neural networks for e, μ, π identification

Two different kinds of neural networks, feed-forward multi-layer mode with Back-Propagation training algorithm (BP) and Kohonen's Learning Vector Quantization networks (LVQ), are adopted for the identification of e, μ, π particles in Beijing Spectrometer (BES) experiment. The data samples for training and test consist of μ from cosmic ray, e and π from experimental data by strict selection. Although their momentum spectra are non-uniform, the identification efficiencies given by BP are quite uniform versus momentum, and LVQ is little worse. At least in this application BP is shown to be more powerful in pattern recognition than LVQ.

Detection of Seizure Activity in EEG by an Artificial Neural Network: A Preliminary Study

Neural networks, inspired by the organizational principles of the human brain, have recently been used in various fields of application such as pattern recognition, identification, classification, speech, vision, signal processing, and control systems. In this study, a two-layered neural network has been trained for the recognition of temporal patterns of the electroencephalogram (EEG). This network is called a Learning Vector Quantization (LVQ) neural network since it learns the characteristics of the signal presented to it as a vector. The first layer is a competitive layer which learns to classify the input vectors. The second, linear, layer transforms the output of the competitive layer to target classes defined by the user. We have tested and evaluated the LVQ network. The network successfully detects epileptiform discharges (EDs) when trained using EEG records scored by a neurologist. Epochs of EEG containing EDs from one subject have been used for training the network, and EEGs of other subjects have been used for testing the network. The results demonstrate that the LVQ detector can generalize the learning to previously “unseen” records of subjects. This study shows that the LVQ network offers a practical solution for ED detection which is easily adjusted to an individual neurologist's style and is as sensitive and specific as an expert visual analysis.

Fuzzy Model Using a Modified LVQ Network for Sinter Strand Process

Rule-based fuzzy systems can be structurally mapped into Kohonen's Learning Vector Quantization (LVQ) network. However the structure of the Kohonen layer is fixed in standard LVQ and the recall is in the manner of winner-takes-all. This is not suitable for the self-organizing of rule-based fuzzy systems. Modifications are made to the LVQ's learning algorithm and the activity function of the Kohonen node. With the modified LVQ network the fuzzy model of an iron-ore sinter strand process is constructed. The resulting fuzzy model is dynamically self-organized from input-output observation data with very little a priori knowledge about the sinter strand process and is adaptive to the change in the environment in which the sinter strand process operates. The results of the simulation show that the model posseses very good performance in terms of prediction accuracy, adaptive ability and robustness.

Neural network analysis of the EMG interference pattern

This paper investigates the performance of artificial neural networks for analysing and classifying EMG signals from healthy subjects and patients with myopathic and neuropathic disorders. EMG interference patterns (IP) were recorded under maximum voluntary contraction from the right biceps of a total of 50 subjects. Parameters were obtained from the signals using recognized quantification techniques including turns analysis, small segments analysis and frequency analysis. Supervised networks examined were an improved backpropagation network (IBPN), a radial basis network (RBN), and a learning vector quantization network (LVQ). Supervised networks using different combinations of parameters from turns analysis and small segments analysis gave diagnostic yields of 60–80%. Combinations using frequency analysis parameters produced similar results. The performance of unsupervised Self-Organising Feature Maps (SOFM) was generally lower than that of the supervised networks. Including personal data (sex and age) did not improve the overall performance.

Fuzzy algorithms for learning vector quantization

This paper presents the development of fuzzy algorithms for learning vector quantization (FALVQ). These algorithms are derived by minimizing the weighted sum of the squared Euclidean distances between an input vector, which represents a feature vector, and the weight vectors of a competitive learning vector quantization (LVQ) network, which represent the prototypes. This formulation leads to competitive algorithms, which allow each input vector to attract all prototypes. The strength of attraction between each input and the prototypes is determined by a set of membership functions, which can be selected on the basis of specific criteria. A gradient-descent-based learning rule is derived for a general class of admissible membership functions which satisfy certain properties. The FALVQ 1, FALVQ 2, and FALVQ 3 families of algorithms are developed by selecting admissible membership functions with different properties. The proposed algorithms are tested and evaluated using the IRIS data set. The efficiency of the proposed algorithms is also illustrated by their use in codebook design required for image compression based on vector quantization.

Three Artificial Neural Network Paradigms in High Dimensional Multisource Spatial Data Classification

Abstract In this paper, fundamentals of three artificial neural network (ANN) paradigms of backpropagation artificial neutral network (BPANN), modular artificial neural network (MANN) and learning vector quantization (LVQ) network are discussed. The capabilities of the three ANN paradigms in high dimensional multisource spatial data classification were evaluated and compared. The research indicates that all three ANN paradigms are capable in the high dimensional multisource spatial data classification. Paradigms of MANN and LVQ networks, with modular design and competition mechanism, achieved more reliable and better classification results than BPANN.

Detection of Pc-2–5 groups of geomagnetic micropulsations with neural networks

The following paper is focused on qualitative analysis, i.e. classification and recognition, of sensor signals with neural networks. The detection of Pc-2–5 groups of micropulsations in the signal of the geomagnetic field has been chosen as an example through which some potentials of neural networks for solving such problems are revealed. The entire preprocessing stage and the generation of learning and test samples are described. Classification has been carried out with four different neural network architectures: multilayer perceptrons, self-organizing maps into categorization, learning vector quantization networks and probabilistic neural networks. All but the last ones have given quite satisfactory results. Furthermore, analyses results have provided some global conclusions concerning the use of neural networks for solving classification problems.

A labeled object identification system using multilevel neural networks

This paper describes the design of a neural network based labeled object identification system, to be used for product classification at the final inspection stage of an IBM personal computer manufacturing line. The objective was to design and identification system using existing equipment that would provide robust and accurate classification, as well as a simple means for adding new product models to the system. In the first stage of the identification system, an image of the product is obtained, and the region containing the label is segmented from the rest of the image. Preprocessing operations are performed to extract the region of interest from the segmented image. Normalized and preprocessed images of the labels are compressed using a fully-connected back-propagation autoencoder network. Features extracted in this manner are used as inputs to a Learning Vector Quantization (LVQ) network, trained to classify the labels. The system so designed is shown to satisfy the primary requirements of a typical industrial classification system.

A neural network architecture for noise prediction

This paper presents application of neural networks (NNs) to the problem of prediction of noise caused by urban traffic. The most representative physical variable quantifying noise emissions is the equivalent sound pressure level. Up to now it has been identified on the basis of semi-empirical models, typically regression analysis, which generally do not provide very accurate approximations of the trend followed by sound pressure level. The authors have attempted to overcome this difficulty by adopting a neural approach based on a Backpropagation Network (BPN). Results obtained by the comparison of the BPN approach with those provided by selected relationships found in relevant literature, show how good is the approach proposed. The neural solution to the problem has shown the necessity, in certain phases, of a set of acoustic measurements which is as free as possible of error. The complexity of error identification by means of classical approaches has led the authors to explore the possibility of a neural solution to this problem as well. The authors therefore propose use of a neural architecture made up of two cascading levels. At the first level a supervised classifying network, the learning vector quantization (LVQ) network, filters the data discarding all the wrong measurements, while at the second level the BPN predicts the sound pressure level.

Prototype-based minimum classification error/generalized probabilistic descent training for various speech units

In previous work we reported high classification rates for learning vector quantization (LVQ) networks trained to classify phoneme tokens shifted in time. It has since been shown that the framework of minimum classification error (MCE) and generalized probabilistic descent (GPD) can treat LVQ as a special case of a general method for gradient descent on a rigorously defined classification loss measure that closely reflects the misclassification rate. This framework allows us to extend LVQ into a prototype-based minimum error classifier (PBMEC) appropriate for the classification of various speech units which the original LVQ was unable to treat. Speech categories are represented using a prototype-based multi-state architecture incorporating a dynamic time warping procedure. We present results for the difficult E-set task, as well as for isolated word recognition for a vocabulary of 5240 words, that reveal clear gains in performance as a result of using PBMEC. In addition, we discuss the issue of smoothing the loss function from the perspective of increasing classifier robustness.

Neural Classifiers for Automated Visual Inspection

This paper discusses the application of a back-propagation multi-layer perceptron and a learning vector quantization network to the classification of defects in valve stem seals for car engines. Both networks were trained with vectors containing descriptive attributes of known flaws. These attribute vectors (‘signatures’) were extracted from images of the seals captured by an industrial vision system. The paper describes the hardware and techniques used and the results obtained.

Control chart pattern recognition using learning vector quantization networks

Pattern recognition systems using neural networks for discriminating between different types of control chart patterns are discussed. A class of pattern recognizers based on the Learning Vector Quantization (LVQ) network is described. A procedure to increase the classification accuracy and decrease the learning time for LVQ networks is presented. The results of control chart pattern recognition experiments using both existing LVQ networks and an LVQ network implementing the proposed procedure are given.

Artificial neural networks in diagnosis of thyroid function from in vitro laboratory tests

We studied the potential benefit of using artificial neural networks (ANNs) for the diagnosis of thyroid function. We examined two types of ANN architecture and assessed their robustness in the face of diagnostic noise. The thyroid function data we used had previously been studied by multivariate statistical methods and a variety of pattern-recognition techniques. The total data set comprised 392 cases that had been classified according to both thyroid function and 19 clinical categories. All cases had a complete set of results of six laboratory tests (total thyroxine, free thyroxine, triiodothyronine, triiodothyronine uptake test, thyrotropin, and thyroxine-binding globulin). This data set was divided into subsets used for training the networks and for testing their performance; the test subsets contained various proportions of cases with diagnostic noise to mimic real-life diagnostic situations. The networks studied were a multilayer perceptron trained by back-propagation, and a learning vector quantization network. The training data subsets were selected according to two strategies: either training data based on cases with extreme values for the laboratory tests with randomly selected nonextreme cases added, or training cases from very pure functional groups. Both network architectures were efficient irrespective of the type of training data. The correct allocation of cases in test data subsets was 96.4-99.7% when extreme values were used for training and 92.7-98.8% when only pure cases were used.

A comparison of neural network performance for seismic phase identification

Traditional techniques of analysis and interpretation of seismic events involve a series of complex steps involving sophisticated signal processing as well as many manual tasks. Automating each of these steps is an important goal of this ongoing research. The paper discusses the use of neural networks in performing phase identification, namely the discrimination of distinct seismic waves within a seismogram. The scope is further restricted to the identification of only two of the regional principal phases, Pg and Lg, among the signals collected in the western United States. Using a database of 75 earthquakes and 75 underground nuclear explosions, the performance of several types of neural networks was compared. The performance of probabilistic neural network (PNN), radial basis function (RBF) network and learning vector quantization (LVQ) network is compared with a back-propagation network that combines the conjugate-gradient method with a weight-elimination strategy. The results indicate that the latter outperformed all other methods tested.