1.5. Road Map of the Book
The organization of this book is displayed in Figure 1.2.
Part I provides an overview of state-of-the-art biometric authentication applications, including face and speaker recognition. Chapter 2 contains an overview of the design and system requirements pertaining to biometric authentication systems. It also presents a general pattern classification system in which feature extraction and adaptive classifiers play an important role.
Part II establishes the theoretical foundation for the machine learning techniques advocated in this book. To facilitate development of effective biometric authentication systems, several modern machine learning models are instrumental in handling complex pattern recognition and classification problems. Chapter 3 discusses the expectation-maximization (EM) algorithm useful for pattern representation and classification in an unsupervised training environment. In general, machine learning becomes much more effective under a supervised learning framework. Chapters 4 and 5 cover key learning strategies, taking advantage of a teacher's guidance during the training phase. Note, however, that if a model is overtrained for the sake of achieving nearly perfect training accuracy, it could compromise generalization accuracy. For this reason, the tradeoff between training and generalization accuracies is a focus of this research. The fundamental theory of Fisher's linear discriminant analysis (LDA) and support vector machines (SVM) is presented in Chapter 4. Comprehensive coverage of multi-layer learning models and well-known back-propagation (BP) algorithms is provided in Chapter 5.
Machine learning models can be naturally integrated into hierarchical and modular neural networks to provide a unified framework for biometric authentication. Part III presents several flexible network structures. Chapter 6 elaborates a flexible hierarchical information processing structure comprising many expert- and class-based modules. Prominent expert-based modular networks, such as the mixture-of-experts (MOE) and the hierarchical mixture-of-experts (TIME), are presented. Pattern classification and biometric authentication are intimately related, therefore, it is sometimes necessary to adopt interclass learning strategies based on class-based modular networks. In Chapter 7, a probabilistic decision-based neural network (DBNN) is developed based on this design principle.
In Part IV, the learning networks are tailored to special information processing systems for biometric authentication and data fusion. The most important authentication application domains are face recognition and speaker verification. Chapter 8 presents probabilistic neural networks for face-based biometric authentication while Chapter 9 covers voice-based authentication. These two chapters also explore the issue of robustness in the design of practical face or voice recognition systems. Although single modality systems still prevail in current authentication systems, multicue biometric authentication will play a vital role in future security systems. The hierarchical learning model facilitates the adoption of soft-decision strategies, which is vital to several of the multicue data-fusion techniques addressed in Chapter 10.