Book Review: The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World

Interest in data science is at an all-time high. With buzzwords like big data and deep learning creeping into non-technical spheres and algorithms playing a central role in blockbuster movies like Captain America, computing is becoming a topic with mainstream appeal. The average person might not need to (or care to) learn how to program, but a high level understanding of computing and machine learning is becoming necessary to understand how the world works. The Master Algorithm is a new book by University of Washington professor Pedro Domingos that attempts to foster this sort of high level understanding by surveying key ideas of machine learning with largely non-technical, allegorical prose.


As the name implies, the "master algorithm" is the key concept of the book that drives its chapters. The master algorithm as Domingos calls it is a universal learner: a single algorithm capable learning everything that there is to be learned from any data you give it. In the author's words the book hypothesizes: "All knowledge--past, present and future--can be derived from data by a single, universal learning algorithm." Domingos is convinced that such an algorithm exists, which is debatable, but even if it does, we aren't likely to find one any time soon. When figuring out how to eke out an extra 10% accuracy for recommendations on Netflix captivates some of the world's greatest minds for over two years, we'd be lucky to find anything resembling the universal learner Domingos describes before the next millennium. Although the central hypothesis of the book may be a pipe dream, its subtitle is not. Regardless of whether a master algorithm exists, the push to create more powerful machine learning techniques and apply them to every conceivable area of our lives will have far-reaching consequences.

"The Master Algorithm" does a great job explaining the basics of machine learning in terms that an intelligent layman can understand. The prologue and first two chapters explain the basics of machine learning and the concept of the master algorithm and some of the big problems that machine learning can help solve. The following five chapters are devoted to the "five tribes" of machine learning. Each tribe has its own philosophy about how to learn from data and a gold-standard algorithm doing it. Symbologists create logic-based models (decision trees), connectionists make models inspired by the human brain (neural networks), evolutionaries use models inspired by evolution (genetic algorithms), Bayesians make models based on statistics and probabilistic inference (naive Bayes) and analogizers describe the world in terms of similarities (K-nearest neighbors.). The book explains concepts in enough depth to understand how the algorithms work at a conceptual level, but it doesn't get into technical details.

The final chapters pull concepts together, culminating in an extended allegory about unifying the various tribes to make the master algorithm. Domingos goes on to describe Alchemy, a machine learning system he helped create that attempts to incorporate aspects of the various tribes to arrive at a general-purpose learner. It concludes with a glimpse of what the future might look as data science advances toward the master algorithm.


The Master Algorithm is well-written book that tells some nice stories and explores fascinating concepts, while eschewing technical details. The discussion is sometimes more opinion and speculation than science, but it rarely ceases to entertain. If you are looking for a practical book, look elsewhere. If you're interested in data science and looking for a fun, though-provoking read, this book is for you.


I give The Master Algorithm 4 out of 5 stars: Very Good.




*Note: Mr. Domingos has an archived machine learning course available on Coursera. You can view the lectures by clicking "Preview Lectures" on the course page.