The goal of machine learning theory is to make mathematical guarantees about the performance of machine learning algorithms both generally and on a case by case basis. There are two reasons most data scientists don't learn machine learning theory. The first is that learning it requires prerequisites that most people would only take as part of a graduate program (ex. measure-theoretic probability). The second is that knowing machine learning theory doesn't really change how one uses machine learning in practice. Despite these issues, we think it is worthwhile to study machine learning theory because it offers a richer understanding of the algorithms. You'll also be much more likely to understand theory-oriented machine learning papers if you're familiar with the current state of machine learning theory.
Understanding Machine Learning
Shai Shalev-Shwartz and Shai Ben-David
- In-text exercises
- Errata and free .pdf
Once you have the prerequisites (which is no easy feat), this is a very accessible book on machine learning theory. The book is divided into two parts: theory and algorithms. The pure theory section is only about 100 pages, and is not especially dense. Once you've read the introductory theory, you can more or less jump into any algorithm section you want. The algorithm sections themselves are quite theory-heavy, so you should probably already be familiar with a particular algorithm before you read this book's take on it. This book unified a lot of discordant machine learning concepts for us, so we think it makes for a great capstone book if you have been studying machine learning for some time.