Information Theory, Inference & Learning Algorithms

I find it interesting that most of the people reviewing this book seem to be reviewing it as they would any other information theory textbook. Such a review, whether positive or critical, could not hope to give a complete picture of what this text actually is. There are many books on information theory, but what makes this book unique (and in my opinion what makes it so outstanding) is the way it integrates information theory with statistical inference. The book covers topics including coding theory, Bayesian inference, and neural networks, but it treats them all as different pieces of a unified puzzle, focusing more on the connections between these areas, and the philosophical implications of these connections, and less on delving into depth in one area or another.

This is a learning text, clearly meant to be read and understood. The presentation of topics is greatly expanded and includes much discussion, and although the book is dense, it is rarely concise. The exercises are absolutely essential to understanding the text. Although the author has made some effort to make certain chapters or topics independent, I think that this is one book for which it is best to more or less work straight through. For this reason and others, this book does not make a very good reference: occasionally nonstandard notation or terminology is used.

The biggest strength of this text, in my opinion, is on a philosophical level. It is my opinion, and in my opinion it is a great shame, that the vast majority of statistical theory and practice is highly arbitrary. This book will provide some tools to (at least in some cases) anchor your thinking to something less arbitrary. It's ironic that much of this is done within the Bayesian paradigm, something often viewed (and criticized) as being more arbitrary, not less so. But MacKay's way of thinking is highly compelling. This is a book that will not just teach you subjects and techniques, but will shape the way you think. It is one of the rare books that is able to teach how, why, and when certain techniques are applicable. It prepares one to "think outside the box".

I would recommend this book to anyone studying any of the topics covered by this book, including information theory, coding theory, statistical inference, or neural networks. This book is especially indispensable to a statistician, as there is no other book that I have found that covers information theory with an eye towards its application in statistical inference so well. This book is outstanding for self-study; it would also make a good textbook for a course, provided the course followed the development of the textbook very closely.

This is fantastic book. Really takes an intuitive approach to the material. The explanation of occam's razor is worth the price of the whole book. Highly recommended.

I have used this to get a good background in the topics covered, especially inference theory, and in general I found it to be great book which fills a market gap. The only sins I see are sins of omission. I personally would have enjoyed seeing a more task driven organization. I seem to need these methods periodically but I never seem to need the same method twice. Also, many of the techniques are heavily iterative, i.e., monte carlo, neural networks, etc. This is fine but much of what I do is in the context of simulations where 100,000 step iterative methods don't work so well because of resource constraints. Historically, that has been the problem with many of these methods. They are useful for relatively small domains but don't necessarily work that well for "real" problems. That is probably why more task oriented books are not available. Of course the author is following the outline of the current research into the subject manner which in turn is largely determined by "interesting" and "doable" problems. The real progess in this field will come when the problems are formulated more by what is needed in the nontraditional domains of application. A good example of a useful compression (and identification in some cases) technique that is not covered is Principal Component Analysis. Technically, it is in none of the technique domains covered in this book, but it would have been nice to see some of the methods in the book compared with PCA. The author does make the statement at one point that image recognition is an interesting problem for which the method being discussed at the time is used. Nevertheless, this is a great overview of the subject manner and is very entertaining. That in the long run probably explains the problem: it is a textbook.

Uniting information theory and inference in an interactive and entertaining way, this book has been a constant source of inspiration, intuition and insight for me. It is packed full of stuff - its contents appear to grow the more I look - but the layering of the material means the abundance of topics does not confuse.

This is _not_ just a book for the experts. However, you will need to think and interact when reading it. That is, after all, how you learn, and the book helps and guides you in this with many puzzles and problems.

I am a PhD student in computer science. Over the last year and a half this book has been invaluable (and parts of it a fun diversion).

For a course I help teach, the intoductions to probability theory and information theory save a lot of work. They are accessible to students with a variety of backgrounds (they understand them and can read them online). They also lead directly into interesting problems.

While I am not directly studying data compression or error correcting codes, I found these sections compelling. Incredibly clear exposition; exciting challenges. How can we ever be certain of our data after bouncing it across the world and storing it on error-prone media (things I do every day)? How can we do it without >60 hard-disks sitting in our computer? The mathematics uses very clear notation --- functions are sketched when introduced, theorems are presented alongside pictures and explanations of what's really going on.

I should note that a small number (roughly 4 or 5 out of 50) of the chapters on advanced topics are much more terse than the majority of the book. They might not be of interest to all readers, but if they are, they are probably more friendly than finding a journal paper on the same topic.

Most importantly for me, the book is a valuable reference for Bayesian methods, on which MacKay is an authority. Sections IV and V brought me up to speed with several advanced topics I need for my research.