Henrik Brink Joseph W. Richards Mark Fetherolf - Real-World Machine Learning

Machine learning (ML) has become big business in the last few years: companies are using it to make money, applied research has exploded in both industrial and academic settings, and curious developers everywhere are looking to level up their ML skills. But this newfound demand has largely outrun the supply of good methods for learning how these techniques are used in the wild. This book fills a pressing need.

Applied machine learning comprises equal parts mathematical principles and tricks pulled from a bagit is, in other words, a true craft. Concentrating too much on either aspect at the expense of the other is a failure mode. Balance is essential.

For a long time, the bestand the onlyway to learn machine learning was to pursue an advanced degree in one of the fields that (largely separately) developed statistical learning and optimization techniques. The focus in these programs was on the core algorithms, including their theoretical properties and bounds, as well as the characteristic domain problems of the field. In parallel, though, an equally valuable lore was accumulated and passed down through unofficial channels: conference hallways, the tribal wisdom of research labs, and the data processing scripts passed between colleagues. This lore was what actually allowed the work to get done, establishing which algorithms were most appropriate in each situation, how the data needed to be massaged at each step, and how to wire up the different parts of the pipeline.

Cut to today. We now live in an era of open source riches, with high-quality implementations of most ML algorithms readily available on GitHub, as well as comprehensive and well-architected frameworks to tie all the pieces together. But in the midst of this abundance, the unofficial lore has remained stubbornly inaccessible. The authors of this book provide a great service by finally bringing this dark knowledge together in one place; this is a key missing piece as machine learning moves from esoteric academic discipline to core software engineering skillset.

Another point worth emphasizing: most of the machine-learning methods in broad use today are far from perfect, meeting few of the desiderata we might list, were we in a position to design the perfect solution. The current methods are picky about the data they will accept. They are, by and large, happy to provide overly confident predictions if not carefully tended. Small changes in their input can lead to large and mysterious changes in the models they learn. Their results can be difficult to interpret and further interrogate. Modern ML engineering can be viewed as an exercise in managing and mitigating these (and other) rough edges of the underlying optimization and statistical learning methods.

This book is organized exactly as it should be to prepare the reader for these realities. It first covers the typical workflow of machine-learning projects before diving into extended examples that show how this basic framework can be applied in realistic (read: messy) situations. Skimming through these pages, youll find few equations (theyre all available elsewhere, including the many classic texts in the field) but instead much of the hidden wisdom on how to go about implementing products and solutions based on machine learning.

This is, far and away, the best of times to be learning about this subject, and this book is an essential complement to the cornucopia of mathematical and formal knowledge available elsewhere. It is that crucial other book that many old hands wish they had back in the day.

B EAU C RONIN

H EAD OF D ATA , 21 I NC .

B ERKELEY , CA