Book Review: Building Machine Learning Powered Applications

Like many of you may have now, I have recently found myself with a bit of free time while waiting for the COVID-19 storm to pass. Time that was previously used for social activities like high school baseball games and tennis practices is now freed up, and there’s only so many home projects I can tackle at once. I've been using this newly found time to read!

I’ve been looking for a book about machine learning that wasn’t just about a specific topic, but instead took me through the entire process, at some level of detail, around creating a software solution that used machine learning. I’ve read and watched enough about these topics to be dangerous, but wanted to get a more tactical look at a complete working process. I think I may have found a book that does this: “Building Machine Learning Powered Applications: Going From Idea To Product” by Emmanuel Ameisen.  The title describes perfectly what I had been looking for, and below is my take on this new book.

The books is divided into multiple parts (collections of 2-3 chapters) that cover the end to end process of building a ML-powered application, from planning to data pipelining to model iteration to deployment and monitoring.  

Part I: Find the Correct ML Approach

Even if you don’t read the rest of the book, this first section is a great intro to ML-powered applications, and the information you need to make the best decisions about getting started. It covers various types of approaches and the pros and cons of each, while providing some pragmatic examples and practices. It also discusses data needs, and model performance in a very approachable way. There aren’t any code examples yet, and I found these first few chapters easy to digest and understand. It also lays out the example application that the rest of the book takes you through. The example app is an editor that provides writing recommendations for questions to be posted to online forums (such as Stack Overflow).  You feed it your question (in the form of one or more sentences) and it provides suggestions like “use less adverbs".  At first I struggled with this as an example because it wasn’t highly business-focused, but it works for the purposes needed by the rest of the book, and you can apply this same thinking of this type of application to other scenarios. At the end of the day, the model takes some information from the user, and provides suggestions for them on how to improve – that is certainly useful in many business scenarios to support nearly everything that business users do on a daily basis. And while this part doesn’t get technical, if you just read this first section and put the rest of the book down, you’d get value out of it.

Part II: Build a Working Pipeline

Part II dives right into writing code (Python) to start to build the “pipeline” for our model. Now, I haven’t written any real code for several years, but the examples are fairly straightforward enough to follow. If you’ve never written code before, from this part on could be pretty challenging to follow, as it gets technical really quickly. If you are non-ML software developer reading this, I’d encourage you to follow along with the examples on your own machine as you go through it – I didn’t do that, but I would bet I’d have gotten more out of it if I had. 

What I really like about the approach used here is that it starts with a non-ML model – just some hard-coded heuristics – to help test the pipeline process, and then builds on that. In a way, it reminded me of a Test-Driven Development type approach – write just enough code that works, run a test, tweak the code, run another test, and so on. 

This part also covers data in more detail, and really has some great advice and examples. A large amount of effort goes into data preparation and processing to build an ML model, and this covers at least at a high level, what those steps are and also explores data types outside what is needed in the example application. 

Part III: Iterate on Models

This part gets even more highly technical, and I’ll admit to having to re-read paragraphs a few times to make sure I actually understood before moving on. There is some assumption of knowledge of various math topics, but the author does do a good job of straightforward explanations when necessary. There are also many references to tools and examples that you can dive into as you read or follow along in code. 

The three chapters in this part take you through the core steps of iterating on a model, and get you to a reasonably working model. The author encourages the reader to iterate through the steps again, measuring performance and identifying areas for improvement.

Part IV: Deploy and Monitor

By now in the narrative of the book, we have built a working model which gives writing recommendations, and we are ready to deploy it! There’s some extremely good information in here on topics you may not be familiar with but are highly valuable. For example, it covers how someone might defeat a model to preform fraudulent activity (think fraudulent credit card transactions). It also covers various deployment models, and pros/cons of different approaches. The last few chapters cover error management, performance, and monitoring models to improve predictive performance. 

Summary

Although a bit technical for my current skillset at times, I thought this was an excellent overview of looking at the end to end process for building a machine learning powered application. At around 238 pages, it doesn’t go super-deep on any one topic, and that’s a good thing. It’s consumable while giving you an idea of all the pieces you need to know to go from end to end. I’d recommend this book to any software developer or manager who needs to understand how ML applications are built, and needs to start crossing that bridge from traditional software development to ML-focused software development.

More about the book can be found here: https://mlpowered.com/book/

p.s. Aside from work-related books, I’m also reading the Mitch Rapp series by the late Vince Flynn, but that’s purely for some escapism entertainment. I’m enjoying them – a nice distraction.