Customer Reviews

A Probabilistic Theory of Pattern Recognition (Stochastic Modelling and Applied Probability)

5 star:		(5)
4 star:		(0)
3 star:		(0)
2 star:		(0)
1 star:		(0)

 

 

In giving this book a second read, its importance finally dawned on me: it is one of the few if only books that provides a well-rounded theoretical (i.e. mathematical definitions and proofs) perspective on pattern recognition. Although other books, such as Duda et al's "Pattern Classification", have a significant degree of mathematical rigor, very few can claim to be based on the solid mathematical foundations of Lesbesgue measure theory, as this book is. This book has been a big inspiration for me, in that most pr papers I come across provide some method X, and show how experimentally it is more efficient or effective than methods Y and Z. Such papers, although possibly generating interest in the subject or method, do little if anything to advance the theory which in the end will have the final say of how, when, and why something works or when it doesn't. On the other hand, by making the assumption that the data comes from an unknown (i.e. nonparametric) probability measure space which induces an inherent optimal Bayesian error on the classification problem, this book shows how the theory of probability can be used to prove some very interesting results.

As an example, the authors define what it means to have a universally consistent classifier; i.e. a classifier which converges to the optimal Bayesian classifier as the amount of training data approaches infinity in the limit (irregardless of the data distribution). Moreover, one of the important results is that such classifiers exist and are often quite easy to devise (e.g. nearest-neighbor methods). And to be able to mathematically prove this is indeed inspiring.

In closing, I would highly recommend this book to anyone who has the mathematical prereqs (probability from an abstract measure-theory point of view)
and is interested in doing high quality mathematical research in pattern recognition. For that audience this book will provide a good foundation for literally an unlimited number of interesting questions; many of which remain unanswered.

For those who are more interested in the practice of pattern recogition, the above mentioned book by Duda et al. (ISBN 0471056693) will do just fine as a reference. The book "Pattern Recognition" by Theodoridis et al. is also of high quality (ISBN: 0126858756).

This is an awesome book, the best in-depth book on statistical classification to date. Filled with theorems and proofs on classical nonparametric techniques plus neural networks and learning. Standard reference for anybody doing serious pattern recognition and learning. Destined to become a classical reference in the field.

The book is great but the notations the authors employ will make you want to drop it on a first reading. Despite the generic title, it is really a reference book for the experts.

Issues in generalization are presented better in the book by Anthony and Bartlett but overall it is the best book available (for learning theorists).

I am a machine learning/ data mining student and bought this because my advisor recommended it to me and this is my second favorite book after Christopher Bishop's Pattern Recognition and machine learning... !!!
If i were wondering about some inequality, this book has it.
if i want some sort of risk bounds this book has it. wow. i just skimmed through the contents lists and read parts but i am already a fan.
probably not used in many courses (spanning: CMU, MIT, STANFORD, etc where machine learning is big) because this is more like a reference than a straight learning textbook.
but THIS IS A GEM and a must have.
BUY THIS!!!

This book provides a solid theoretical foundation for pattern recognition and statistical learning. If you consider yourself and expert, or want to be an expert in this field, this book is a must read. It will make you think hard about the concepts (and may be question whether you are or want to become an expert!).