Given the booming popularity of machine learning, a good introductory textbook would indeed be quite valuable. A number of rather advanced textbooks exist, but most of them require a solid background in mathematics, with a heavy focus on probability theory. Is it possible to write a solid introduction to machine learning that would be accessible to, say, second-year undergraduate students?
Kubat’s book is a valiant try. It requires a modest mathematical background. It is superbly organized: each section includes a “what have you learned” summary, and every chapter has a short summary, accompanying (brief) historical remarks, and a slew of exercises. Even better, there are simple exercises, others labeled “give it some thought,” as well as computer assignments. In most of the chapters, there are very clear examples, well chosen and illustrated, that really help the reader understand each concept. While most of the chapters cover a specific technique, others serve as supplements to further illustrate the ideas: one covers case studies, while others distill practical knowledge and discuss evaluation methods and statistical significance. There are frequent warnings that machine learning does not necessarily lead to greater human understanding, as well as a variety of other warnings related to the well-known pitfalls of various machine learning techniques.
Unfortunately, this is marred by a number of defects. First is the amazingly sloppy editing, as seen through obvious grammatical errors, for example: “In as sense” in the footnote on p. 114, “less-then-useful” on p. 145, and “less then 2%” on p. 153. Even more annoying are condescending sentences such as, “A purist may protest: the induction mechanism seeks to create a decision tree that scores zero errors on the training examples, but this perfection may be lost in the pruned tree! But the practically-minded engineer is not alarmed” (pp.126-127). There are references to “practically minded engineers” throughout; this seems to be frequently accompanied by a dismissal of theory and an invitation to “tweak” things until they work. Once in a while, the author seems to quite grudgingly admit (like at the start of chapter 7) that a bit of theory really can be useful. But just as often, fundamental theories (such as Shannon’s amazing work on information theory) are casually introduced; for example, on p. 120 one finds: “When quantifying the information content of such a message, the following formula has been found to be convenient: Ipos = -log2 ppos.” “Convenient” indeed! No less subtle is the repeated use of “mathematicians know that” (see p. 139 for just one example).
More damaging is the underlying assumption, throughout the book, that someone who reads this book will learn enough to actually be able to perform nontrivial machine learning tasks. Even worse, somehow there is an assumption that implementing some of the algorithms described in the book, by a lone programmer on a single machine, might actually lead to real results. This is incredibly naive. And it is already outdated (even for a book published in 2015!), as recently a lot of extremely sophisticated machine learning frameworks, finely tuned by experts for a long time, have been made open source. The accessibility of vast computing resources (such as Amazon’s) for hire means that quite a bit of the warnings in the book about the sizes of the problems that can be tackled in practice are irrelevant.
Having said that, I did learn quite a bit about very basic machine learning by reading this book. I feel like I am now ready to read an intermediate textbook. However, I would not be so arrogant as to think of myself as ready to “do” machine learning, using just that knowledge, in realistic situations.
