Machine Learning: A First Course for Engineers and Scientists

andreas-lindholm, niklas-wahlstrom, fredrik-lindsten and thomas-schon's undergraduate machine-learning textbook, published by cambridge-university-press. Read from the online draft of 4 March 2026, 348 pages, hosted at smlbook.org — "free to view and download for personal use only. Not for re-distribution, re-sale or use in derivative works."

This is the spoke's second textbook and its second T1 book, after MacKay — and where MacKay comes at learning through information theory, this one is the plain first course the spoke has been missing.

It is the founding gap, as a syllabus

The spoke's founding complaint is that it holds the top of the stack (agents, fine-tuning, systems) and the bottom (information theory), with nothing in the middle. Chapters 2 to 10 are the middle:

k-NN and decision trees as the first two methods, before any parametric model.
Linear regression, logistic regression, regularisation, generalised linear models, with the normal equations derived in an appendix.
Understanding, evaluating and improving performance — an entire chapter, discussed below.
Learning parametric models: loss functions and likelihood, regularisation, parameter optimisation, optimisation with large datasets, hyperparameter optimisation.
Neural networks and deep learning: the model, training, CNNs, dropout, with backpropagation derived in an appendix.
Ensembles: bagging, random forests, AdaBoost, gradient boosting.
Kernels: non-linear input transformations, kernel ridge regression, support vector regression and classification, kernel theory, the representer theorem.
The Bayesian approach and Gaussian processes, with the multivariate Gaussian in an appendix.
Generative models and unlabelled data: Gaussian mixtures and discriminant analysis, clustering, deep generative models, representation learning and dimensionality reduction.

Gradient boosting and random forests get a full chapter here, which matters for the spoke's tabular thread: the method practitioners actually reach for is taught, not just named.

The two chapters that do not usually exist

Chapter 11, "User Aspects of Machine Learning." Not the algorithms but the work: defining the problem, improving a model, what to do when you cannot collect more data, practical data issues, and "Can I Trust my Machine Learning Model?" The improvement chapter carves a training-validation set out of the training data, so the hold-out error splits into three named terms — training error, generalisation gap, and a train-val mismatch — and the size of the third tells you whether your problem is the model or a data distribution that does not match the one you care about. That is a diagnostic procedure, which is the shape of the practice this spoke keeps failing to find in papers.

§11.5 is unusually blunt for a textbook: "The main argument for supervised machine learning is, frankly, that it works well empirically," and the design philosophy "is to deliver good predictions rather than to explain them."

Chapter 12, "Ethics in Machine Learning" — contributed in full by david-sumpter, per the acknowledgements. Fairness and error functions, limitations of training data, and §12.2, "Misleading Claims about Performance."

§12.2 is about this corpus's problem

The spoke's standing complaint, written into growth edge 2, is that nearly every number it holds is a vendor's own. §12.2 teaches that as a property of the field:

"Much of the research in machine learning is conducted by large private companies such as Google, Microsoft, and Facebook… they also have public relations departments whose goal it is to engage the wider general public in the research conducted. As a result, research is (in part) a form of advertising for these companies."

It then names cases — AlphaGo/AlphaZero promoted with a company-financed film watched nearly 20 million times on a company-owned platform, and Musk's repeated level-5 autonomy claims — and says the point is not whether the technical work was good but that "the scope of the results can potentially be exaggerated for commercial gain."

So the spoke's evidence spectrum now has a T1 textbook stating the mechanism, rather than the corpus inferring it one vendor blog post at a time. It does not close growth edge 2 — that edge wants a neutral benchmark that states its own label quality, and a chapter about publication incentives is not a benchmark. It does explain why the edge keeps failing.

What it does not answer

Nothing on fine-tuning, PEFT or adaptation of pretrained models (open question 2), nothing on time-series foundation models (question 7), and no end-to-end account of a labelling operation (edge 1a) — §11.4's practical data issues are a first course's treatment, not a documented operation. The book predates none of that; it is a first course, and the spoke's open questions live above it.

Licence and tier

T1 — the authors' own text for a CUP-published book. Not open-licensed: personal use only, no redistribution, no derivative works, which is a stricter grant than any mathematics text this hub holds and worth noting beside the free-to-read framing. freshness: volatile because the draft is dated and revised; the March 2026 version is what this page rests on.

information-theory-inference-learning-algorithms · cambridge-university-press · andreas-lindholm · niklas-wahlstrom · fredrik-lindsten · thomas-schon · david-sumpter · synthesis