|Artificial Intelligence, Machine Learning, and Deep Learning (Mercury Learning)
Author: Oswald Campesato
Another AI/ML book - is there room for another one?
The AI/ML book market is more than adequatly catered for and another book in the same area needs something to make it worth buying. This one is interesting because it is small, slim and well written. If you want an in-depth detailed discussion of the mathematics and implementation then there are lots of very thick volumes that you can consider. Some of them are written by experts so close to their subject that they explain every single twist and turn on the way to mastery of the subject. These are not good if you want a quick overview or a "get me started" guide, which is where this book comes in.
This book is reasonably short and doesn't spend too much time going over history and management issues. It is a book that aims to tell you about AI/ML and get you started with some practical work. The implementation is via Python, Keras, TensorFlow and Pandas which is perfectly reasonable as they are what most people use right up to the cutting edge - only the really bleeding edge uses anything more custom.
The book is a little slow to start and Chapter 1 spends too long on musings about what AI is. If you are keen to get on to the subject matter then skip to Chapter 2. This introduces the very basics of machine learning, but in a very tangential way at first. I'm not sure I'd introduce dimension reduction at this point but it is a matter of taste. Much of the chapter is about testing and gauging how good an ML method is - again better left for later.
When we do eventually meet a real ML method it is linear regression, which alway brings me to a halt because I think of it as a classical statistical technique hardly worthy of being called ML. What is more, the form introduced isn't even particular sophisticated - isn't stepwise regression closer to ML than simple linear multivariate regression? This book isn't alone in ignoring the sophisticated forms of linear regression - it seems to be a modern trend.
Chapter 3 moves on to classification - the usual stuff kNN, decision trees, random forests, SVMs, Bayes and logistic regression. No mention of discriminant analysis - yet another important technique ignored by most.
Chapter 4 gets going on deep learning which I suspect what most readers want to know about. It is a reasonably practical and understandable account and we get as far as convolutional networks.
Chapter 5 moves on to recurrent networks which are far to advanced for most real applications and mostly overshadowed by the idea of "attention" added to a basic network. The chapter ends with a look at autoencoders, which are very important and should have a chapter on their own and GANs, which again deserve a chapter on their own.
Chapter 6 also should be split - Natural Language Processing and Reinforcement Learning are two topics not one. However if my advice was followed the book might morph into something undesirably large. Personally I'd drop the language part of the content as it is specialized enough not to be missed from a general book on AI/ML.
This is not cutting edge. It doesn't even mention some of the cutting edge ideas that are around at the moment and many of the up-to-date ideas that it does touch on are not covered in enough detail. This is the tradeoff in writing a reasonably short book. As long as you don't expect this book to tell you everything, you should be reasonably happy with its mix of informal explanation and basic practical projects.
|Last Updated ( Tuesday, 18 May 2021 )