Mathematics of Machine Learning: Master linear algebra, calculus, and probability for machine learning

What was your motivation for writing this book?

When I started in machine learning, I saw most of my colleagues struggle with the basic mathematics behind neural network models. Even though they were great engineers, they made basic modeling mistakes, like using the mean-squared error for classification, having trouble fixing tensor shape mismatch errors, and similar things.

So, I started to educate them about fundamental mathematical topics, like how matrix multiplication works or what the mean-squared error and cross-entropy loss are. Soon, I realized that these intuitive explanations were also missing from the machine learning literature: most books either completely ignored the mathematics behind machine learning or were way too advanced to be useful.

Because of this, I started posting my explanations as Twitter threads, and they instantly became a smash hit. Soon, I had enough material for a book, and four years later, here we are with a ~770-page tome aiming to bring mathematics closer to machine learning professionals.

How is this book different from other machine learning books?

Most books are either written from a purely practical or a purely theoretical perspective. In my experience, the learning curve is extremely steep for most theory-focused books. Even with a PhD in mathematics, there are textbooks where I get stuck in the first chapter, leaving the rest of the book inaccessible to me. I wanted to change that. So, I had three guiding principles in mind.

First, I wanted to “minmax” mathematics in the sense that the knowledge in the book should be maximally useful, but minimal in quantity.

Second, explanations should be intuitive and visual, leaning toward user-friendliness rather than mathematical precision. We have machine learning in mind, not mathematics.

Third, every concept and algorithm should be implemented from scratch in Python. This is essential to understanding mathematics. In the book, we learn NumPy along with the math, implementing algorithms such as the singular value decomposition or the omnipresent gradient descent.