Learning From Data

About the book

Learning From Data is a compact textbook by Yaser S. Abu-Mostafa on the principles behind machine learning. It treats learning from examples as a problem of inference: a system sees finite data and must perform well on cases it has not seen. The book builds its argument around three questions about feasibility, implementation and generalization, and returns to them as the mathematics develops.

The treatment leans on probability and linear algebra, with the bias-variance tradeoff, the VC dimension and regularization among its central ideas. Worked examples and exercises carry the reasoning forward. The emphasis stays on why methods work and when they fail, rather than on recipes for a particular library. The subject is machine learning as a body of theory, presented for readers who want the foundations beneath practical techniques.

How to read it

Read with paper and pencil. The book moves through derivations and inequalities that reward slow, active reading; skimming the formulas leaves the argument without its support. Pause at each definition and check it against a simple case before continuing.

The examples and exercises do most of the teaching. Attempt them before reading solutions, and expect some sections to require a second pass. The prose is economical, so a paragraph may carry several steps of reasoning. Keep the three organizing questions in view; they give the technical material a spine.

Who it suits

This suits students and practitioners with some comfort in probability and linear algebra who want to understand generalization, overfitting and model complexity at a principled level. Readers preparing for more advanced study or research will find the conceptual scaffolding useful. Those seeking a hands-on guide to building systems in a specific framework should look elsewhere.

Reading considerations

The book is theoretical and deliberately spare in engineering detail, so it will not teach you to deploy a model. Its examples and framing reflect machine learning before the deep-learning era; the core theory remains relevant, but readers wanting coverage of modern architectures will need supplementary material.