You have finished the nine technical modules of the course: you know the full machine learning workflow, you have trained supervised and unsupervised models, you can evaluate them rigorously, and you have taken one all the way to production through the MercaFresh projects. Now it is time to consolidate and go deeper, and for that books remain irreplaceable: they offer a depth and coherence that no stray tutorial can match. In this lesson you will not find an endless list, but a short selection of well-established works, with clear guidance on who each one is for, what it adds after this course, and in what order to tackle them.

1. How to choose a technical book (and how not to)
2. Practical introductions: the two essentials
3. Mathematical and statistical foundations
4. Deep learning
5. Machine learning in production
6. Summary table: which book extends which module
7. How to read a technical book
8. Reading paths by profile

Before the list, one criterion: the right book for you is not the most cited one, but the one sitting exactly one step above your current level — not two steps, not zero. After this course you are no longer an absolute beginner: you can skip the "ML for everyone" books and go straight to works with real code or real mathematics.

One practical advantage you already have: virtually all of the reference literature in machine learning is written in English, and translations into other languages tend to run an edition or two behind. You will be reading everything in its original, most up-to-date form — no waiting, no translation lag.

• Edition: make sure you get the latest one — in ML, editions matter, and older printings lag behind the current APIs.
• Who it is for: exactly someone who has just finished this course. It is the natural companion.
• What it adds after this course: it covers the same stack we used (scikit-learn and Keras) but with more depth on every algorithm. The first part extends modules 3, 4, 6 and 7 (pipelines, trees, ensembles, hyperparameter tuning); the second part extends the deep learning we only sketched in module 7 and in the CNN project of module 9.
• How to approach it: this is a book to read with your computer open. The author's notebooks are available on GitHub; reproduce them, then adapt them to a dataset of your own (for example, the MercaFresh data you built in the projects).

• Access: free as a PDF on the authors' official site (statlearning.com). There is an edition with Python labs (ISL with Applications in Python), alongside the classic R edition.
• Who it is for: anyone who wants to understand the statistical why behind what we did, without needing a PhD.
• What it adds after this course: it is the perfect bridge between module 2 (statistics) and modules 4 and 7. It explains linear and logistic regression, cross-validation, regularization (ridge and lasso, which we saw in 07-01) and trees with a rigor the course's hands-on approach could not afford, yet with approachable mathematics.
• How to approach it: after, or in parallel with, Géron. Géron gives you the "how"; ISL gives you the "why".

When ISL starts to feel too light, there are two reference works. An honest warning: they are not the immediate next step after this course; they are books you will keep coming back to for years.

• Access: free as a PDF on the authors' Stanford page.
• Who it is for: readers with a solid grounding in linear algebra, calculus and probability. It is ISL's "big brother", written in part by the same authors.
• What it adds after this course: the full theory behind almost the entire syllabus — from the regularized regression of module 7 to boosting (the reference chapter on gradient boosting, the foundation of the XGBoost and LightGBM we used) and the unsupervised methods of module 5.

• Access: Microsoft Research has distributed it as a free PDF for years.
• Who it is for: anyone who prefers a rigorous probabilistic/Bayesian treatment.
• What it adds after this course: it formalizes what you intuited in Naive Bayes (module 4) and in the distributions of module 2; it goes deep into mixtures and latent-variable methods that connect with the clustering of module 5.

• Edition: pick up the current edition — as with Géron, the APIs move fast.
• Who it is for: anyone who enjoyed the CNN project (09-02) and the NLP project (09-03) and wants to specialize in neural networks.
• What it adds after this course: written by the creator of Keras, the framework we used. It directly extends 07-04 and projects 09-02 and 09-03, with the same practical philosophy as this course. It is the natural next step in deep learning.

• Access: free online at deeplearningbook.org.
• Who it is for: mathematically inclined readers who want the theory of neural networks.
• What it adds after this course: the foundations (optimization, regularization in networks, architectures) that Chollet treats practically. Read it after Chollet, not before.

• Who it is for: anyone left wanting more after module 8, aspiring to an ML engineer role.
• What it adds after this course: everything module 8 covered at an introductory level — deployment, monitoring, data drift (08-03), design decisions, streaming data — treated with the experience of real systems at scale. It is a book of judgment, not of code: it complements, rather than repeats, what we did with FastAPI and Docker in 08-02.

An ML book is not read like a novel. Concrete advice:

• Type the code by hand. Do not copy and paste, and certainly do not just read the code "straight through". Typing it means making mistakes, and debugging them is where the learning happens.
• One chapter, one notebook. Create a notebook per chapter with your notes and experiments. By the end of the book you will have your own reference.
• Swap the dataset. Reproducing the author's example is 40% of the value; applying it to a different dataset (one from MercaFresh, one from Kaggle) is the remaining 60%.
• Give yourself permission to skip chapters. Reference works (ESL, Bishop) are not meant to be read cover to cover, in order. Use the table of contents as a map, not a contract.
• Reread. ISL read right after module 4 and reread after a year of practice are two different books.

In every case: one book at a time. Starting three books in parallel is the most effective way to finish none.

• Mistake: buying books as a substitute for studying. Piling up PDFs and shelf copies produces a false sense of progress. Practical rule: do not get a new book until you have made substantial progress in the previous one.
• Mistake: starting with the advanced books "to do it right from the start". Attacking ESL or Bishop without the necessary mathematical maturity only produces frustration and abandonment. ISL exists precisely for this: the same conceptual content, approachable mathematics.
• Mistake: reading away from the computer. If you are 30 pages into a book with code and have not run anything, you are reading, not learning.
• Tip: take advantage of the free ones. Four of the seven books in this lesson are legally free as PDFs by their authors' own decision. There is no financial barrier to the fundamentals.
• Tip: check the edition. In ML, editions matter (APIs change). Always look for the latest edition, especially with Géron and Chollet.

Pick your profile in the reading-paths table and design a 6-month reading plan for the first book on your path: chapters per week, what you will do with each chapter's code, and which dataset of your own you will use to adapt the examples.

Download the free PDF of An Introduction to Statistical Learning from statlearning.com. Read chapter 2 ("Statistical Learning") and compare its explanation of the bias-variance trade-off with what we saw in module 6 about overfitting. Write down two ideas the book explains better than this course, and one that this course explained in a more practical way.

Get hold of the first code-heavy chapter of the book you have chosen (Géron's notebooks are public on his GitHub) and apply the method from this lesson: type the code by hand into a notebook of your own, note down whatever you do not understand, and swap the example's dataset for a different one.

Exercise 1 (guideline): a realistic plan for Géron could be 1 chapter every 1-2 weeks (the first part moves at a manageable pace; the deep learning chapters deserve 2 weeks each). Commitment per chapter: reproduce the notebook, solve at least one of the proposed exercises, and adapt the example to a dataset of your own (for example, the customer segmentation data from project 09-05). Warning sign: if 3 weeks go by and you have not opened the book, shrink the goal — do not abandon it.

Exercise 2 (guideline): it is common to find that ISL formalizes the bias-variance decomposition better (with its mathematical expression) along with the difference between training and test error; and that this course was more practical in showing overfitting with validation curves in scikit-learn on MercaFresh data. There is no single right answer: the goal is that you actively compare sources.

Exercise 3 (guideline): expect to make 3-5 mistakes while typing (forgotten imports, parentheses, variable names), and debugging them will force you to understand lines you would otherwise have skimmed past. If something breaks when you swap the dataset (columns with different names, missing values), congratulations: you are facing exactly the kind of real-world problem module 3 taught you to solve.

Books are the densest, most coherent form of knowledge available in machine learning, and after this course you have the level to make the most of the good ones: Géron and ISL as the natural continuation, ESL/Bishop and Goodfellow when you want theory, Chollet for deep learning and Huyen for production. The key is not how many you read, but how: one at a time, with your computer open, adapting every example to your own data. In the next lesson we look at the dynamic complement to books: online courses — where to find the ones worth taking, and how to avoid becoming a collector of certificates.