Supervised and Unsupervised Pattern Recognition

Feature Extraction and Computational Intelligence

Evangelia Miche Tzanakou

• Publisher: CRC Press
• Year: 2017
• eBook ISBN: 9781351835558
• Edition: 1
• Topic: Technology & Engineering
• Subtopic: Electrical Engineering & Telecommunications
• Index: Technology & Engineering

There are many books on neural networks, some of which cover computational intelligence, but none that incorporate both feature extraction and computational intelligence, as Supervised and Unsupervised Pattern Recognition does. This volume describes the application of a novel, unsupervised pattern recognition scheme to the classification of various types of waveforms and images.This substantial collection of recent research begins with an introduction to Neural Networks, classifiers, and feature extraction methods. It then addresses unsupervised and fuzzy neural networks and their applications to handwritten character recognition and recognition of normal and abnormal visual evoked potentials. The third section deals with advanced neural network architectures-including modular design-and their applications to medicine and three-dimensional NN architecture simulating brain functions. The final section discusses general applications and simulations, such as the establishment of a brain-computer link, speaker identification, and face recognition.In the quickly changing field of computational intelligence, every discovery is significant. Supervised and Unsupervised Pattern Recognition gives you access to many notable findings in one convenient volume.

Section I

Overviews of Neural Networks, Classifiers, and Feature Extraction Methods—Supervised Neural Networks

1 Classifiers: An Overview

Woogon Chung and Evangelia Micheli-Tzanakou

1.1 Introduction

One way to better understand a subject is to classify or categorize it among related subjects. Many classifiers result from different approaches to classification problems. The purpose of this article is to categorize the well-known classifiers in the literature according to how they learn to classify.

Lippmann’s tutorial paper¹ described various classifiers as well as neural networks in detail after his first discussion² on the general application of neural networks. Another general overview on this subject is found in a paper by Hush and Horne³ in which neural networks are reviewed in the broad dichotomy of stationary vs. dynamic networks. Weiss and Kulikowski’s book⁴ generally touches the classification and prediction methods from the point of view of statistics, neural networks, machine learning, and expert systems.

The purpose of this article is not to give a tutorial on the well-developed networks and other classifiers but to introduce another branch in the growing classifier tree, that of nonparametric regression approaches to classification problems. Recently Hastie, Tibshirani, and Buja⁵ introduced the Flexible Discriminant Analysis (FDA) in the applied statistics literature, after the unpublished work by Breiman and Ihaka.⁶

Canonical Correlation Analysis (CCA) for two sets of variables is known to be a scalar multiple equal to the Linear Discriminant Analysis (LDA). Optimal Scaling (OS) is an alternative to CCA, where the classical Singular Value Decomposition (SVD) is used to find the solutions. OS brings the flexibility obtained via nonparametric regression and introduces this flexibility to discriminant analysis, hence the name Flexible Discriminant Analysis.

A number of recently developed multivariate regressions are used for classification, in addition to other groups of classifiers for a data set obtained from handwritten digit images. The software is contributed mainly from the authors or active researchers in this area. The sources are described in later sections after the description of each classifier.

1.2 Criteria For Optimal Classifier Design

We start with a general description of the classification problem and then proceed to a discussion of simpler cases in which assumptions are made. Which criterion should be used is application specific. Expected Cost for Misclassification (ECM) is applied to problems in which...

Table of contents

• Cover
• Title Page
• Copyright
• Contents
• Section I — Overviews of Neural Networks, Classifiers, and Feature Extraction Methods—Supervised Neural Networks
• Section II Unsupervised Neural Networks
• Section III Advanced Neural Network Architectures/Modular Neural Networks
• Section IV General Applications
• Index