A comprehensive guide to pattern recognition methods, the most difficult problem for machine learning, is provided in this book. The author presents practical material with lots of examples, and covers almost all of the methodologies used and accepted today. Although the presentation is structured toward novices, professionals might also enjoy the examples, the outline of the algorithms, and the presentations of less common methods. Readers will find coverage of radial basis function networks or the use of splines, interesting techniques to boost the performance of a classifier or to combine several classifiers, and search and optimization algorithms.

The book is conceived rather like a manual of applied computer learning. Each chapter ends with an inventory of supplementary reading and a range of questions for review.

Instead of providing exhaustive details on every topic, the author stresses the basic ideas or elements concerning particular issues throughout the book. In the introductory chapters, the author clarifies some specific notions (such as learning and classification), outlines the different types of learning, describes basic concepts (such as regression, Hebb’s rule, the simple model of a neuron, and the concept of the perceptron), and discusses linear separability issues. Subsequently, the author develops the framework outlined in the introduction--supervised, unsupervised, evolutionary, and reinforcement approaches to pattern recognition--and presents these techniques in this order.

Chapter 3 is devoted to the multi-layer perceptron, discussing specific problems such as the back-propagation of error, output activation functions, and the number of hidden layers. The next chapter presents radial basis functions and spline interpolation, proving their role in defining a specific network. The subsequent section is an introduction to support vector machines. The sequence of supervised methods ends with the presentation of decision trees.

Chapter 7 is very interesting, as it describes techniques to boost the performance of a particular classifier, the Adaboost algorithm, and also techniques to combine several expert algorithms in order to improve a recognition system’s performance.

Chapter 8 is an introduction to statistical methods, describing the essential background for statistical approaches to machine learning, basic statistic measures and rules, and the Bayes classifier.

Chapter 9 introduces some unsupervised learning techniques. The reader will find information on k-means, vector quantization, a weighting approach to vector quantization, and self-organizing maps.

The following two parts are about dimension reduction, optimization, and search. The author covers some popular techniques--linear discriminative analysis, factor analysis, principal component analysis, and independent component analysis--and how they relate to neural networks.

In the framework of evolutionary learning, the author describes many attractive examples providing the background for genetic algorithms. The last sections discuss reinforcement learning and graphical models. The hidden Markov model approach is classified as a probabilistic graphical model, based on its relation to graph theory.

The last part is an introduction to Python, the programming language chosen to implement the algorithms and illustrate their results. According to Marsland, it was chosen because it is “freely available, multi-platform, and becoming a default for scientific computing.”

The author gathered, synthesized, and integrated a great deal of information, and provided many interesting, illustrative examples and formalized algorithms. The book is well suited for anyone interested in machine learning issues, but students and lecturers are the main audience.