Machine Learning

This section is a deep, self-contained curriculum covering machine learning from first principles through to production deployment: the mathematical foundations, classical algorithms, deep learning, sequence models and LLMs, computer vision, generative models, reinforcement learning, and the systems engineering needed to run any of it in production.

How to use this section

Follow one of the paths below rather than reading the sidebar top to bottom. Each section builds on Foundations; most later sections only need Foundations plus one or two others, not the whole tree.

Three learning paths

Path	Section sequence	What you can build at the end
Foundations first	Foundations → Classical ML → Deep Learning	Train and evaluate models on tabular data, from a linear baseline through a tuned deep network
Deep learning / LLM	Foundations → Deep Learning → Sequences & NLP → LangChain	Fine-tune and serve a language model, and build an LLM-powered application on top of it
Applied practitioner	Foundations → Classical ML → Production & MLOps	Ship a model to production and keep it monitored, versioned, and safe to roll back

How the sections depend on each other

Foundations underlies everything else. Deep Learning is the shared base for Sequences & NLP, Computer Vision, and Generative Models. Time Series and Recommender Systems are applied domains that lean mostly on Classical ML. Production & MLOps sits downstream of all of them — it's about deploying and operating whatever model the earlier sections produced.

The eleven sections

Section	Pages	Description	Start here
Foundations	17	Math, statistics, and the core concepts every other section assumes	What Is Machine Learning
Classical ML	20	Regression, classification, trees, ensembles, clustering, and dimensionality reduction	Linear Regression
Deep Learning	18	Neural networks from a single perceptron through training at scale on GPUs	From Perceptron to MLP
Sequences & NLP	16	Tokenization through transformers, pretraining, fine-tuning, and decoding	Text Preprocessing and Tokenization
Computer Vision	13	Convolutions, CNN architectures, detection, segmentation, and vision transformers	Images as Tensors
Generative Models	11	Autoencoders, VAEs, GANs, normalizing flows, and diffusion models	What Is a Generative Model
Reinforcement Learning	12	MDPs, value-based and policy-gradient methods, PPO, and RLHF	The Reinforcement Learning Problem
Time Series & Forecasting	5	Stationarity, ARIMA, ML-based forecasting, and backtesting without leakage	What Makes Time Series Different
Interpretability	4	Feature importance, PDP/ICE, SHAP, LIME, counterfactuals, and how they mislead	Why Interpretability Matters
Recommender Systems	4	Collaborative filtering, matrix factorization, hybrids, and ranking metrics	The Recommendation Problem
Production & MLOps	16	Data pipelines, experiment tracking, serving, monitoring, and responsible AI	From Notebook to Production

See the LangChain reference for building LLM-powered applications on top of the models this section covers.

Conventions used across these docs

• Figures over diagrams. Concepts are illustrated with real plots wherever one can be drawn — decision boundaries, loss surfaces, gradient flow, attention maps. Every figure is generated from source by the scripts in tools/figures/, so it matches the page it sits on, and several are genuine computations rather than illustrations: the gridworld values come from running value iteration, the cliff-walking paths from running Q-learning and SARSA. The full inventory is in static/img/ml/CREDITS.md.
• Mermaid is kept for flows and state machines, where there is nothing to plot.
• Tables are preferred over prose for comparisons.
• Admonitions flag the important parts: info for the key idea, warning/danger for genuine foot-guns — leakage, silent miscalibration, metrics that mislead.
• Every page ends with See also links, and most carry a runnable, dependency-light code block that demonstrates the idea rather than wrapping a library call.