A complete set of professional, textbook-quality teaching PDFs covering a depth-first path from mathematical foundations to production deep learning — written for someone who already codes well but wants real understanding, not just API calls.
The teaching material is authored in LaTeX and compiled to polished PDFs with xelatex
(full Unicode + CJK support, so the Chinese-language tokenization examples render natively).
What's inside
Each phase is a self-contained PDF in pdf/ with its own title page, table of
contents, math derivations, worked examples, diagrams, code, common-pitfall callouts,
a checkpoint project, and a self-check.
| Phase | Pages | |
|---|---|---|
00_overview.pdf |
Curriculum map, study method, two routes, reference library | 6 |
phase0_math_tooling.pdf |
Math & Tooling Foundations (linear algebra, calculus, probability, NumPy) | 13 |
phase1_core_ml.pdf |
Core Machine Learning (classical models, evaluation, leakage, tuning) | 10 |
phase2_dl_foundations.pdf |
Deep Learning Foundations (backprop by hand, optimization, PyTorch) | 8 |
phase3_computer_vision.pdf |
Computer Vision (convolutions, CNNs, transfer learning, ViT) | 7 |
phase4_nlp_transformers.pdf |
NLP & Transformers (attention, tokenization/CJK, LoRA, RAG) | 8 |
phase5_generative_models.pdf |
Generative Models (VAEs, GANs, diffusion) | 6 |
phase6_reinforcement_learning.pdf |
Reinforcement Learning (MDPs, DQN, PPO, RLHF) | 6 |
phase7_mlops_deployment.pdf |
MLOps & Deployment (tracking, serving, drift, CI/CD) | 7 |
phase8_capstones.pdf |
Capstone Projects (domain-tied, end-to-end) | 6 |
ML_DL_Curriculum_complete.pdf |
All of the above merged into one volume | 77 |
Suggested pace: ~9–14 months at 8–12 hrs/week going deep with projects. Compressed high-value path: Phase 0 → 1 → 2 → 4 → 7.
Standalone deep-dive books
| Topic | Pages | |
|---|---|---|
linear_algebra.pdf |
Linear Algebra: From Beginner to Expert — a complete, visual, application-driven course (vectors & geometry, linear systems, matrices, determinants, vector spaces & the four fundamental subspaces, linear transformations, orthogonality/least squares/QR, eigenvalues, the spectral theorem & quadratic forms, the SVD & Eckart–Young, numerical linear algebra, tensors & matrix calculus) with a capstone chapter synthesizing how every concept powers Machine Learning, Deep Learning, and AI. | 60 |
calculus.pdf |
Calculus: From Beginner to Expert — a complete, visual, application-driven course (functions & limits, the derivative & differentiation rules, applications of the derivative, the integral & the Fundamental Theorem, techniques & applications of integration, sequences/series/Taylor, multivariable partial derivatives & the gradient, multiple integrals & vector calculus, optimization & differential equations) with a capstone chapter synthesizing how calculus powers Machine Learning, Deep Learning, and AI (gradients, backpropagation, optimizers, Hessians, entropy/KL/ELBO, diffusion & neural ODEs). | 57 |
probability.pdf |
Probability & Statistics: From Beginner to Expert — a complete, visual, application-driven course (sample spaces & axioms, conditional probability & Bayes' theorem, random variables & expectation, the distribution zoo, joint distributions/covariance/the multivariate Gaussian, inequalities & limit theorems (LLN/CLT), estimation & maximum likelihood, confidence intervals & hypothesis testing, Bayesian inference, regression & GLMs with the bias–variance tradeoff, Markov chains & MCMC, information theory) with a capstone chapter synthesizing how probability and statistics power Machine Learning, Deep Learning, and AI (MLE↔losses, MAP↔regularization, calibration & uncertainty, generative models). | 54 |
regression.pdf |
Regression: From Beginner to Expert — a complete, visual, application-driven course (what regression is & the conditional mean, simple linear regression, multiple regression & the projection geometry, assumptions/diagnostics/inference, feature engineering, the bias–variance tradeoff & cross-validation, regularization (ridge/lasso/elastic net), logistic & softmax regression, generalized linear models, nonlinear & nonparametric regression (splines/kernels/GAMs/trees/Gaussian processes), robust/quantile/Bayesian regression, evaluation/pitfalls/causality) with a capstone chapter synthesizing how regression is the conceptual core of Machine Learning, Deep Learning, and AI (neural nets as flexible regression, losses↔likelihoods, gradient descent, weight decay, LLM regression heads, diffusion, distributional RL, scaling laws). | 52 |
bayes.pdf |
Bayesian Inference & Naive Bayes: From Beginner to Expert — a complete, visual, application-driven course (Bayesian thinking & probability as belief, Bayes' theorem & the base-rate fallacy, the prior/likelihood/posterior engine, conjugate priors (Beta–Binomial/Gamma–Poisson/Normal–Normal), choosing priors, posterior summaries & decision theory, Naive Bayes classifiers (Gaussian/Multinomial/Bernoulli + Laplace smoothing), Bayesian vs frequentist inference, Bayesian networks & graphical models, MCMC & variational inference, Bayesian regression & hierarchical models, model comparison & Occam's razor) with a capstone chapter synthesizing how Bayes powers Machine Learning, Deep Learning, and AI (MAP↔regularization, Bayesian neural networks, MC dropout, Bayesian optimization & Thompson sampling, VAEs/diffusion, in-context learning as implicit Bayes). | 53 |
svm.pdf |
Support Vector Machines: From Beginner to Expert — a complete, visual, application-driven course (the maximum-margin idea & widest-street intuition, the maximal-margin classifier & hyperplane geometry, soft margins/slack/the $C$ parameter, hinge loss & the regularized primal, Lagrangian duality/KKT/support vectors, the kernel trick & RKHS/Mercer, kernels in depth (linear/polynomial/RBF/$\gamma$, string & graph kernels, random Fourier features), support vector regression & the $\epsilon$-tube, optimization (QP/SMO/LIBSVM/LIBLINEAR/Pegasos), multiclass/probabilities/one-class SVM/practical pipelines, margins & VC dimension & generalization theory, SVMs vs logistic/trees/deep learning) with a capstone chapter synthesizing how SVMs and kernel methods power Machine Learning, Deep Learning, and AI (hinge & large-margin losses, ArcFace/contrastive metric learning, deep features + SVM, the NTK = kernel-SVM equivalence, contrastive learning as one-class SVMs, generalization & robustness certificates). | 53 |
trees.pdf |
Tree-Based Methods: From Beginner to Expert — a complete, visual, application-driven course (recursive partitioning & the CART/ID3/C4.5 lineage, classification trees with Gini/entropy/information gain, regression trees & variance reduction, growing/splitting/cost-complexity pruning, the bias–variance of trees, bagging & random forests with OOB error, boosting from AdaBoost to gradient boosting (gradient descent in function space), the modern libraries XGBoost/LightGBM/CatBoost (second-order objectives, histogram & leaf-wise growth, GOSS/EFB, ordered boosting), feature importance & SHAP, practical modeling (categoricals/missing/imbalance/tuning/leakage), specialized trees (ExtraTrees, Isolation Forests, LambdaMART ranking, survival/quantile/BART/EBM, stacking), and trees vs linear/deep learning) with a capstone chapter synthesizing how trees power Machine Learning, Deep Learning, and AI (tabular dominance & inductive bias, GBDT+neural hybrids, differentiable/deep neural decision forests, TabNet/FT-Transformer/TabPFN, TreeSHAP interpretability, anomaly detection). | 54 |
Reading the PDFs
Just open anything in pdf/. Start with 00_overview.pdf, or read the single
merged ML_DL_Curriculum_complete.pdf.
Teaching-box legend
The documents use color-coded callouts consistently:
- Key idea (blue) — the one thing to remember from a section
- Definition (navy) — precise statements and formulas
- Intuition (green) — the mental model / "why it works"
- Common pitfall (red) — mistakes that bite practitioners
- Worked example (gray) — a computation done step by step
- Checkpoint project (orange) — the hands-on deliverable for the phase
- Resources (purple) — curated books, courses, papers
Rebuilding the PDFs
Requirements
- A TeX Live (or MacTeX) distribution with
xelatexand the packagestcolorbox,tikz,pgfplots,listings,fontspec,hyperref(all standard). - Fonts: TeX Gyre family (bundled with TeX Live) and a CJK font
(the style auto-detects
Heiti SC/PingFang SC/STHeition macOS). - Optional:
pdfunite(from poppler) to produce the merged volume.
Build
./build.sh # build every PDF + the merged volume
./build.sh phase4_nlp_transformers # build a single document (by src stem)
Output PDFs land in pdf/. Intermediate files go to .build/ (git-ignorable).
Repository layout
Machine_Learning/
├── README.md
├── build.sh # XeLaTeX build script (single or all docs + merge)
├── style/
│ └── mlcourse.sty # shared style: palette, fonts, teaching boxes, math macros
├── src/ # one .tex per phase (edit these)
│ ├── 00_overview.tex
│ ├── phase0_math_tooling.tex
│ ├── … phase8_capstones.tex
└── pdf/ # compiled output (the deliverables)
To edit content, change the .tex files in src/; all visual styling lives in
style/mlcourse.sty so the look stays consistent across every phase.