01 — Books Canon¶
The books listed here are not “recommended reading.” They are the structural skeleton of applied ML knowledge. Twelve books, ranked by phase. You don’t need to read all of them cover to cover — but you need to know all of them well enough to know which chapter to open when you’re stuck.
What most people get wrong: They buy all the books in month one, read the first three chapters of each, and finish nothing. The phase mapping below tells you exactly when to open each book and what to read. Start thin. Go deep when the roadmap reaches that domain.
Reading Strategy¶
Active books (months 1-4): Read chapter by chapter, run every code example
Reference books (always): Keep open in a browser tab, consult when stuck on a concept
Phase-specific books: Open only when the roadmap says to — reading PRML in month one is a mistake
The Canon¶
1. Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow¶
Author: Aurélien Géron | Year: 3rd edition, 2022 | Level: Beginner → Intermediate
What it covers: End-to-end applied ML and deep learning. Classification, regression, SVMs, random forests, gradient boosting. Then neural networks, CNNs, RNNs, attention, autoencoders, GANs, RL. All with actual code. Uses Scikit-learn for classical ML, TF/Keras for deep learning.
Community verdict (Reddit/Discord 2024-2026): Consistently the top recommendation for “I want to understand AND implement, not just read theory.” The 3rd edition updated to TF 2.x. Widely described as “the book that finally made things click.”
Cost: ~$55-65 print (O’Reilly). O’Reilly online subscription (~$500/year) includes it — check if your employer covers this. No free legal PDF, but a free preview is available.
⚠️ Caution: Uses TensorFlow/Keras, not PyTorch. For a PyTorch-first learner, treat the deep learning sections as conceptual understanding and re-implement in PyTorch separately.
Roadmap phase: Months 1-3 (Chapters 1-9 on classical ML, Chapters 10-18 on deep learning basics)
2. Understanding Deep Learning¶
Author: Simon J.D. Prince | Year: 2023 | Level: Intermediate
What it covers: Modern deep learning from scratch — supervised learning, loss functions, neural network theory, CNNs, transformers, generative models (VAEs, GANs, diffusion), RL. Written to be the modern replacement for Goodfellow et al.
Community verdict: Released free online, quickly adopted by courses. Goodreads: 4.64/5. Called “the book Goodfellow should have been updated to.” Specifically praised for covering diffusion models and modern architectures that Goodfellow (2016) predates entirely.
Cost: Free PDF legally available at udlbook.github.io/udlbook. Print: ~$80 Cambridge University Press.
Roadmap phase: Months 2-5 (replace or supplement Goodfellow for modern architectures)
3. Deep Learning¶
Author: Ian Goodfellow, Yoshua Bengio, Aaron Courville | Year: 2016 | Level: Intermediate → Advanced
What it covers: The academic foundation — linear algebra, probability, information theory, regularization, optimization, CNNs, RNNs, practical methodology. Dense and comprehensive.
Community verdict: The canonical academic reference. Reddit consensus: “Don’t use it as a textbook — use it as a reference.” Primary complaint is that it predates transformers entirely. Use Prince (2023) as the primary and Goodfellow for historical depth on optimization, regularization theory, and the mathematical treatment of backprop.
Cost: Free PDF legally at deeplearningbook.org. Print: ~$80 MIT Press.
Roadmap phase: Months 3-6 (reference for optimization/regularization chapters; not primary text)
4. Build a Large Language Model (From Scratch)¶
Author: Sebastian Raschka | Year: 2024 | Level: Intermediate
What it covers: Building a GPT-2-scale LLM from scratch in PyTorch. Tokenization, embedding layers, multi-head attention, training loops, fine-tuning with instruction data, RLHF overview. Extremely hands-on.
Community verdict: Raschka is one of the most respected ML educators writing in 2024-2026. His newsletter (Ahead of AI, 73K+ subscribers) signals his credibility. This book is the PyTorch-native code-first path to understanding modern LLMs. Reddit ML communities call it “the Karpathy course in book form.”
Cost: ~$40-50 Manning Publications. No free PDF but Manning has discount codes and MEAP access.
Roadmap phase: Months 7-9 (LLM phase — pair with Karpathy’s Zero to Hero course)
5. Pattern Recognition and Machine Learning (PRML)¶
Author: Christopher Bishop | Year: 2006 (still canonical) | Level: Advanced
What it covers: Bayesian probability, graphical models, mixture models, EM algorithm, kernel methods, SVMs, sequential models, sampling methods. The mathematical backbone of classical ML.
Community verdict: “Requires graduate-level math to appreciate.” Not for beginners. Universally respected. Reddit: “If you can read PRML comfortably, you understand ML mathematics.” Free PDF was made available by Microsoft Research.
Cost: Free PDF legally — microsoft.com/en-us/research/people/cmbishop/ — link to free PDF is on Bishop’s MSR page. Print: ~$80.
⚠️ Prerequisite: Linear algebra, probability, calculus at undergraduate level. Not month-1 material.
Roadmap phase: Months 5-8 (Bayesian methods, advanced ML theory — specific chapters as needed)
6. The Elements of Statistical Learning (ESL)¶
Author: Trevor Hastie, Robert Tibshirani, Jerome Friedman | Year: 2nd ed. 2009 | Level: Advanced
What it covers: The statistical theory of ML — linear methods, regularization (ridge, lasso), SVMs, ensemble methods (boosting, bagging, random forests), unsupervised learning, neural networks (covered mathematically).
Community verdict: The data science bible for people who want to understand why algorithms work, not just how to use them. Frequently cited in industry interviews for ML scientist roles. “ESL + ISL is the complete statistical learning stack.”
Cost: Free PDF legally at stanford.edu/~hastie/ElemStatLearn/. Print: ~$80.
Roadmap phase: Months 4-7 (statistical ML theory, alongside classical methods work)
7. An Introduction to Statistical Learning (ISL)¶
Author: James, Witten, Hastie, Tibshirani | Year: 2nd ed. (ISL with Python) 2023 | Level: Intermediate
What it covers: The accessible version of ESL. Regression, classification, cross-validation, tree methods, SVMs, unsupervised learning. Python edition uses sklearn examples. “ESL’s friendlier sibling.”
Community verdict: Where ESL is a reference text, ISL is a teachable textbook. The 2023 Python edition is a genuine update for the modern practitioner.
Cost: Free PDF legally at statlearning.com. Print: ~$60.
Roadmap phase: Months 2-4 (statistical ML foundations, read before ESL)
8. Mathematics for Machine Learning¶
Author: Deisenroth, Faisal, Ong | Year: 2020 | Level: Beginner → Intermediate
What it covers: Linear algebra, analytic geometry, matrix decompositions, vector calculus, probability and distributions, continuous optimization. Not applied ML — pure math prerequisites with ML motivation.
Community verdict: “Finally a math book written for ML practitioners.” Bridged the gap between “I know what a matrix is” and “I understand why PCA works.” Recommended by Reddit for anyone who needs to solidify mathematical foundations before PRML or ESL.
Cost: Free PDF legally at mml-book.github.io. Print: ~$50 Cambridge.
Roadmap phase: Months 1-2 (use as math primer alongside core courses — don’t try to “finish” it, use as reference)
9. Deep Learning with Python¶
Author: François Chollet | Year: 2nd edition, 2021 | Level: Beginner → Intermediate
What it covers: Deep learning through the lens of Keras (now TF/Keras). Chollet is the creator of Keras and the inventor of depthwise separable convolutions. Excellent on the why of architectural decisions. Covers CNNs, RNNs, transformers (chapter 11 in 2nd ed.), GANs, and model interpretation.
Community verdict: “Chollet explains concepts better than anyone. The book reads like a conversation with someone who actually understands what’s happening.” Strong for intuition building, weaker on PyTorch (it’s Keras-first).
Cost: ~$50-60 Manning. No free PDF. Manning often has 50% off sales.
Roadmap phase: Months 2-4 (deep learning intuition — read alongside Géron for a second perspective)
10. Deep Learning for Coders with fastai and PyTorch¶
Author: Jeremy Howard, Sylvain Gugger | Year: 2020 | Level: Beginner (top-down)
What it covers: Practical deep learning using the fast.ai library. Image classification, segmentation, NLP, collaborative filtering, tabular data. Top-down: build working systems first, understand theory second.
Community verdict: “Best book for getting something working fast.” Companion to the fast.ai course. Reddit: “If you’ve been stuck in tutorial hell, this book breaks you out.” PyTorch under the hood. 2020 is the edition — still valid, but some fast.ai API changes since publication.
Cost: Free online at book.fast.ai. O’Reilly print: ~$60.
⚠️ Note: fast.ai library API has changed since 2020. Some code examples require version pinning. Check the fast.ai forums for updated notebooks.
Roadmap phase: Months 1-3 (parallel to Géron — fast.ai for intuition, Géron for depth)
11. Designing Machine Learning Systems¶
Author: Chip Huyen | Year: 2022 | Level: Intermediate → Advanced
What it covers: Not a math book — an engineering book. Data engineering for ML, feature engineering, model selection, training, deployment, monitoring, MLOps. The gap between “I can train a model” and “I can ship a model.”
Community verdict: “The book nobody writes but everybody needs.” Reddit ML/MLOps communities consistently recommend this for anyone going into production. Chip Huyen is one of the most credible ML practitioners writing about production systems.
Cost: ~$55-65 O’Reilly. No free PDF. O’Reilly subscription access available.
Roadmap phase: Months 9-13 (MLOps and production — do NOT read this before you’ve built real systems)
12. Natural Language Processing with Transformers¶
Author: Lewis Tunstall, Leandro von Werra, Thomas Wolf (HuggingFace team) | Year: 2022 | Level: Intermediate
What it covers: Practical NLP using HuggingFace Transformers. Text classification, NER, QA, summarization, translation, generation, fine-tuning, building custom pipelines. Written by HuggingFace core team.
Community verdict: “The HuggingFace docs aren’t enough — this book fills the gap.” Highly recommended for anyone doing NLP or working with LLMs in production via HuggingFace.
Cost: ~$60 O’Reilly. Partial free online read at transformersbook.com.
Roadmap phase: Months 6-9 (NLP/LLM phase)
Quick Reference: Books by Phase¶
Phase |
Months |
Books |
|---|---|---|
Foundations |
1-3 |
Géron (active), MML (reference), fast.ai book (active), ISL (reference) |
Deep Learning Core |
2-5 |
Prince (active), Chollet (active), Goodfellow (reference) |
Statistical Theory |
4-7 |
ISL → ESL → PRML (sequential) |
LLM/NLP |
6-9 |
NLP with Transformers, Raschka LLM book |
Production |
9-13 |
Chip Huyen (active), PRML/ESL (reference) |
Verified Free PDFs Summary¶
Book |
Free PDF? |
URL |
|---|---|---|
Deep Learning (Goodfellow) |
✅ Yes |
|
Understanding Deep Learning (Prince) |
✅ Yes |
|
PRML (Bishop) |
✅ Yes |
Bishop’s MSR page |
ESL (Hastie et al.) |
✅ Yes |
|
ISL (James et al.) |
✅ Yes |
|
Mathematics for ML |
✅ Yes |
|
fast.ai book |
✅ Yes |
Return to README.md · Next: 02_courses_canon.md