Back to index
OtherAround 350 pages across 13 chapters, three to five weeks·Around $50 in print, with frequent Manning ebook discounts

Grokking Machine Learning (Luis Serrano)

4.3

The best on ramp I know for someone who has been defeated by a textbook once already. Serrano is unusually good at the moment before understanding, and the price of that is limited depth.

What We Liked

  • Genuinely accessible without being dishonest about what the algorithms do
  • Every method is built from a picture first, so the formalism arrives with meaning attached
  • The chapters on how models fail and how to evaluate them are better than the level suggests
  • Serrano's explanatory instinct is exceptional and his YouTube work is a free preview of it
  • Code is simple enough to read line by line rather than being an opaque library call

What Could Be Better

  • Deliberately shallow, so you will outgrow it within a few months of serious study
  • Almost nothing on deep learning, which is what most readers arrive wanting
  • The slow pace becomes frustrating if your maths was fine to begin with
  • Manning's print pricing is high for a book you will use as a stepping stone
  • Skips enough mathematical detail that some explanations are true but incomplete

Detailed review

Most machine learning books are written by people who find the mathematics easy and cannot fully remember what it was like not to. This produces material that is technically correct and pedagogically hopeless for a large group of readers, who conclude they are not clever enough for the field when what actually happened is that nobody explained the first step. Serrano is unusually good at that first step. If you have watched his YouTube videos you already know the quality, and the book is the same instinct at length.

The method is consistent and it works. Before any equation appears, you get a picture and a story. What problem are we solving, what would a reasonable person try, why does the obvious approach fail, and only then, here is the mathematical object that fixes it. By the time notation shows up you already know what it is for, so it reads as a compact way of saying something you understand rather than as an obstacle.

This ordering is the whole book and it is the right one for the intended reader. Linear regression arrives as fitting a line and adjusting it when it is wrong, and the loss function appears as a way of measuring wrongness rather than as a formula to accept. Logistic regression comes from wanting a probability rather than a number and needing to squash the output. Support vector machines are introduced as the question of which separating line is best, which makes the margin concept obvious rather than arbitrary.

Naive Bayes is built from counting. Decision trees from asking questions in a good order. Each of these is a topic where a standard textbook opens with the objective function and loses half the room. Two sections are better than the book's level would suggest.

The first is on evaluation. Accuracy is a bad metric for imbalanced data, precision and recall trade against each other, ROC curves mean something specific, and choosing a metric is a decision about what kind of error you are willing to make. Many readers get several chapters into a course without meeting this properly, and it is more consequential than any individual algorithm. The second is on overfitting and regularisation, where the geometric explanation of what a penalty term actually does to the solution is clearer than treatments in far more advanced books.

The code is deliberately simple and that is a strength here. You can read it and see the algorithm, rather than watching a library call do the interesting part invisibly. For a reader at this stage that transparency matters more than efficiency. Now the limitations, which are mostly deliberate and should still shape your decision.

This is a shallow book by design. It covers classical machine learning at an introductory level and stops. If you are serious about the field you will outgrow it in a few months, and you should treat it as a first step rather than a destination. That is a fine thing to be, and it does mean you are paying Manning's fairly steep print price for something with a short useful life.

Get the ebook when it is discounted. The deep learning coverage is thin, essentially one chapter, and it is the thing most people arrive wanting. Neural networks are introduced and the treatment does not go far. Anybody buying this to understand how modern AI systems work will be disappointed, and the honest framing is that this is a foundations book that makes those later resources comprehensible rather than one that competes with them.

The pace is the other issue and it cuts by audience. If you are rusty at maths, the slow careful build is exactly what you need. If your calculus and linear algebra are solid, you will find yourself skimming, and by chapter three you may be irritated by how long each idea takes to arrive. This book is not for you and there is no shame in that, it is simply aimed elsewhere.

There is also a cost to the accessibility that deserves naming. Some explanations are true but incomplete, simplified to the point where you have an intuition that will need correcting later. That is a legitimate teaching decision and it means a reader moving on to a rigorous treatment will occasionally have to unlearn a picture. Serrano is generally honest about where he is simplifying, which mitigates it, and the effect is still real.

My 4.3 is for a book that knows exactly who it is for and serves them extremely well. The specific reader is someone who has tried a proper machine learning course, hit the mathematics, and concluded the field was closed to them. It is not, and this is the book that demonstrates it. Read it, gain confidence, then move to something harder within the year.

[ final ]

The verdict.

Perfect if a textbook has already beaten you and you want to try again. Skip it entirely if you are comfortable with calculus and linear algebra, because you will be bored by chapter three.