A brief, but well-written synopsis of neural networks from the viewpoint of a statistician is provided in this book. By merging the two disciplines, the author introduces a new, though well-tested, model to the field of statistics. To neural network practitioners, he shows that there is more to neural networks then just a magical black box.

The book focuses its discussions around two examples common throughout journal literature. The first is a regression example on ground level ozone pollution, and the second is a classification example of loan applications. One of this book’s strengths is that, by using the same dataset as is used in many other journal papers, the author is able to provide comparisons between different approaches and models that have been attempted by other authors.

An introduction to nonparametric models common in the field of statistics starts the book. With this background, the author describes neural networks and how they relate to other models, identifying similarities, and limiting cases. This part of the book reads much like the literature review section of a journal paper. There are many references and models, with few details on any.

The majority of the book is dedicated to the choice of priors and model building, and it is here where the author delivers the most value. Of particular usefulness is the chapter on priors. All too often, random choices are made, and, if the model produces interesting results, a paper is published. The author discusses the theory of proper priors, and how to show that the posterior will be proper as well, while also showing how informative priors are not really suitable for neural networks. A strong overview of noninformative priors is ended with a discussion of model fitting.

The next chapter on model building brings to the fore the Bayesian approach of model over prediction. Finding the right model is where the challenge lies. The author discusses model averaging, automatic relevance determination, and bagging. Finally, the book ends with techniques for searching the model space, covering greedy algorithms, stochastic algorithms, and reversible jump Markov chain Monte Carlo. Comparing the predictive ability of the neural network models discovered using the various methods, as well as comparing them against nonneural network-based models, highlights the view that neural networks are just another statistical model: good for some problems, bad for others.

Lee makes good use of diagrams and figures throughout the text, and these are both relevant and helpful as aids for explanation. The book is intended to be a text for statisticians, and it does presume a significant background in statistics. However, nonstatisticians working with neural networks could potentially benefit even more from this text. If nothing else, just being inspired to acquire the relevant statistical skills would be worth the read.

From the viewpoint of a statistician with no background in neural networks, this book falls short. It is a bit too sparse on the details and intricacies of neural networks, which have evolved into as much an art as a science. However, as an introductory text to neural networks, it will suffice. A comment that I have never made in a review is nevertheless applicable to this book: overall, it was an enjoyable read.