Pattern Recognition and Neural Networks

Reader Reviews
This review is from: Pattern Recognition and Neural Networks (Hardcover) If you want a nice up-to-date treatment on neural networks and statistical pattern recognition with lots of nice pictures and an elementary treatment, I recommend the new edition of Duda and Hart. However, neural networks were basically started by the computer-science / artificial intelligence community using analogies to the human nervous system and the perceived connections to the human thought processes. These connections and arguments are weak. However, a statistical theory of nonlinear classification algorithms shows that these methods have nice properties and have mathematical justification. The statistical pattern recognition research is well over thirty years old and is very well established. So these connections are important for putting neural networks on firm ground and providing greater acceptability from the statistical as well as the engineering community. Ripley provides a theoretical threatment of the state-of-the-art in statistical pattern recognition. His treatment is thorough, covering all the important developments. He provides a large bibliography and a nice glossary of terms in the back of the book. Recent papers on neural networks and data mining are often quick to generate results but not very good at providing useful validation techniques that show that perceived performance is not just an artifact of overfitting a model. This is an area where statisticians play a very important role, as they are keenly aware through their experience with regression modeling and prediction, of the crucial need for cross-validation. Ripley covers this quite clearly in Section 2.6 titled "How complex a model do we need?" It is nice to see the thoroughness of this work. For example, in error rate estimation, many know of the advances of Lachenbruch and Mickey on error rate estimation in discriminant analysis and the further advances of Efron and others with the bootstrap. But in between there was also significant progress by Glick on smooth estimators. This work has been overlooked by many statisticians probably because some of it appears in the engineering literature (but one important paper was in the Journal of the American Statistical Association [JASA] in 1972). To some extent this oversight may be due to the fact that it was not mentioned in Efron's famous 1983 JASA paper and hence is usually missed in the bootstrap literature. Bootstrap methods and cross-validation play a prominent role in this text. This is an excellent reference book for anyone seriously interested in pattern recognition research. For applied and theoretical statisticians who want a good account of the theory behind neural networks it is a must.