Machine Learning Concepts

A four-level handbook from intuition to research and systems

A four-level handbook

Machine learning, from first principles to deployed systems.

This handbook explains the mechanisms behind modern machine learning without asking readers to choose between intuition and technical precision. Each chapter develops one idea through explanation, mechanics, mathematics and code, and its consequences for research or production systems.

Start with Chapter 1 → Browse the concepts appendix

data representation model prediction

A way through a large field

The field is easier to follow when its dependencies are visible. The first part establishes how a model makes a prediction and how training changes its parameters. The next sections develop neural networks, representations and transformers before turning to large-language-model training, reasoning methods, multimodal architectures, deployment and responsible use.

01Foundationsmodels, loss, optimisation

02Neural networkslayers, gradients, regularisation

03Transformersvectors, attention, tokens

04Language modelspretraining, tuning, retrieval

05Reasoning & systemsinference, deployment, safety

Four levels, one argument

Every chapter begins with the question the concept solves. The same concept is then developed at progressively greater technical depth. Readers can stop at the level that matches their purpose or continue through the full sequence.

Level 1

Intuition

What the concept is, what problem it addresses, and what would fail without it.

Level 2

Mechanics

The steps, information flows and design choices that make the method work.

Level 3

Mathematics & code

The notation, objective functions and compact implementations behind the mechanism.

Level 4

Research & systems

What changes at scale, where evidence is limited, and how the method is used in practice.

What the book covers

Part I

Learning from data

Predictions, losses, gradient descent and backpropagation.

Chapters 1–4 → Part II

Neural networks

Layers, activation functions, generalisation and efficient training.

Chapters 5–15 → Part III

Representations & transformers

Embeddings, sequence modelling, attention and tokenisation.

Chapters 16–24 → Part IV

Large language models

Next-token prediction, pretraining, inference, tuning and retrieval.

Chapters 25–34 → Part V

Reasoning models

Evaluation, test-time compute, search, reinforcement learning and distillation.

Chapters 35–42 → Part VI

Architectures & deployment

Mixture-of-experts, vision, diffusion, multimodality, monitoring and safety.

Chapters 43–49 →

Reference tool

Find a concept without losing its context.

The appendix collates the book’s core concepts and distinctions, with links back to the chapters where each one is developed.

Open the appendix →

How to use the handbook

Read the chapters in order if the mechanisms are new to you: later chapters assume the conceptual sequence established earlier. If you already know the basics, use the table of contents to enter at the relevant part, then use the linked concepts appendix to recover any prerequisite. Code examples stay deliberately small. Their purpose is to expose the operation that a library would otherwise hide.