Introduction and Motivation Overview
Chapter 1 of the book is six pages long and contains no mathematics. It is worth reading anyway, because it does two things nothing later does: it fixes the vocabulary, and it tells you how to read the other eleven chapters.
Skipping it is the single most common way to make Chapters 2 through 8 harder than they need to be.
The argument the book is making
Section titled “The argument the book is making”You can call model.fit(X, y) without knowing what happens inside, and for a great many tasks that is
the correct engineering decision. The book’s case for the mathematics is not that the abstraction is
bad — it is that abstraction has a boundary, and knowing the foundations is what lets you:
- understand the fundamental principles the more complicated systems are built on,
- debug an approach that is not working,
- create new solutions rather than only applying existing ones,
- and know the inherent assumptions and limitations of the method you picked.
That last one is the sharpest. Every model encodes assumptions — linear regression assumes Gaussian noise (§9.1), PCA assumes the interesting structure is in the high-variance directions (§10.2), an SVM assumes the classes are separable in some feature space (§12.4). None of those assumptions is printed on the API. They are visible only in the derivation.
What this chapter covers
Section titled “What this chapter covers”flowchart TD W["Finding Words for Intuitions
the vocabulary is overloaded"] R["Two Ways to Read This Book
bottom-up or top-down"] W --> C1["data, model, learning
three core concepts"] W --> C2["predictor vs training
two senses of 'algorithm'"] W --> C3["array, arrow, axioms
three senses of 'vector'"] R --> C4["four pillars,
six foundations"] R --> C5["the dependency graph
what you can skip"] C1 --> CH8["Ch 8 restates all three
mathematically"] C3 --> CH2["Ch 2 adopts the axiom view"] C4 --> ALL["the rest of the module"]
| order | page | what it settles |
|---|---|---|
| 511 | Finding Words for Intuitions | Data, model and learning; both senses of “algorithm”; the three views of a vector; why memorising is not learning. |
| 512 | Two Ways to Read This Book | Bottom-up versus top-down and the cost of each; the four pillars on six foundations; the dependency graph; where the exercises live. |
About an hour for both, and it is the cheapest hour in the module.
The three sentences
Section titled “The three sentences”The book compresses its own contents into three lines at the end of §1.1. They are the shortest honest summary of what machine learning is:
- We represent data as vectors.
- We choose an appropriate model, using either the probabilistic or the optimisation view.
- We learn from available data by numerical optimisation, with the aim that the model performs well on data not used for training.
Read those again after finishing Chapter 12 and every clause will have a chapter behind it. Right now they are a promise; that is fine.
What you’ll be able to do afterwards
Section titled “What you’ll be able to do afterwards”- Spot which of the two or three meanings of “algorithm”, “model” or “learning” a sentence intends.
- Say what the four pillars of machine learning are and which foundations each one needs.
- Pick a reading route for your actual goal, and know what you are giving up by taking it.
- Explain why good training performance is not evidence of a good model.
Prerequisites
Section titled “Prerequisites”None beyond Chapter 0, and even that is optional for these two pages — there is no mathematics here to be rusty at. The NumPy in both pages is short and explained inline.
-
What is the book's strongest argument for learning the mathematics rather than only the libraries?
Every model encodes assumptions — Gaussian noise, high-variance-means-important, separability in some feature space — and none of them appears in the function signature. They are visible only in the derivation.
pch.quizShowAnswer
B — The derivations reveal the assumptions and limitations a model's API does not print — Every model encodes assumptions — Gaussian noise, high-variance-means-important, separability in some feature space — and none of them appears in the function signature. They are visible only in the derivation.
-
Which two things does Chapter 1 provide that no later chapter does?
Section 1.1 fixes the overloaded words and section 1.2 gives the two reading strategies and the structure that makes both possible. Neither is repeated later.
pch.quizShowAnswer
B — The vocabulary, and the instructions for reading the other chapters — Section 1.1 fixes the overloaded words and section 1.2 gives the two reading strategies and the structure that makes both possible. Neither is repeated later.
Drill this chapter
Section titled “Drill this chapter”____ — and ordered by difficulty rather than dependency, so any pillar can be read first.
Recall card
Section titled “Recall card”- Chapter 1 contains no mathematics and is still worth reading, because it fixes the vocabulary and explains how to read the rest.
- The case for the mathematics is assumptions and limitations — every model encodes them, and none of them appears in the API.
- The book reduces itself to three sentences — represent data as vectors, choose a model in the probabilistic or optimisation view, and learn by numerical optimisation aimed at unseen data.
- Two pages carry the chapter: one on overloaded vocabulary, one on reading strategy and the dependency graph.
Next: the words, and why they are slippery — Finding Words for Intuitions.
pch.coffeeTagline
pch.coffeeCtapch.feedbackHeading
pch.feedbackSubheading