Back to index
OtherAround 540 pages across 21 chapters, roughly a semester at a serious pace·Free PDF from the author's site, with a hardcover edition from MIT Press

Understanding Deep Learning (Simon J.D. Prince)

4.8

If I could hand one book to someone who wants to genuinely understand deep learning, this is it. The diagrams alone do work that three other textbooks fail to do in twice the pages.

What We Liked

  • Free PDF, officially, with no email wall, alongside a proper MIT Press print edition
  • The illustrations are the best in the field and carry real explanatory weight rather than decorating the text
  • Short focused chapters make it genuinely readable in single sittings, which almost no textbook manages
  • Covers transformers, diffusion models, graph networks and reinforcement learning rather than stopping at 2018
  • Python notebooks accompany nearly every chapter, so the maths has somewhere to land

What Could Be Better

  • It is a textbook and the maths is not optional, so you need calculus, linear algebra and probability going in
  • Light on engineering practice, with little on data pipelines, deployment or debugging real training runs
  • Reinforcement learning and graph network chapters are introductions rather than treatments
  • No video component, which will lose people who need a lecture to stay engaged
  • The curation that makes it tight also means some topics you might want are simply not here

Detailed review

Deep learning has had a textbook problem for a decade. Goodfellow, Bengio and Courville was the standard and is now substantially out of date, having been written before transformers changed what the field looks like. Bishop is rigorous and forbidding. A great many recent books are either shallow tours with no maths at all or thin wrappers around library documentation that will be obsolete in eighteen months.

Simon Prince's book solves the problem, and the fact that he gives it away free is genuinely unusual. Prince is an honorary professor at Bath who spent years in industry research, and the book reads like it was written by someone who has taught this material to people who did not immediately get it. The structure is the first thing you notice. Twenty one short chapters rather than twelve enormous ones, each covering one idea and stopping.

Supervised learning, shallow and deep networks, loss functions, gradients and initialisation, then measuring performance and regularisation, then convolutional networks, residual networks, transformers, graph networks, unsupervised models, diffusion, reinforcement learning, and a closing pass on why deep learning works at all. The chapters are short enough that you can finish one in a sitting, and that changes the completion odds enormously. Most people who abandon technical books do so in the middle of a forty page chapter. The second thing you notice, and the thing I would actually buy the book for, is the diagrams.

They are extraordinary. Prince's approach is to present each concept in plain language, then give it precisely in mathematics, then show it visually, and the visual step is not decorative. The illustrations of how network depth partitions input space, of what residual connections do to the loss surface, of how attention routes information, each convey something the equations alone leave abstract. I have read explanations of self attention in maybe fifteen places and this is the one where the mechanism finally sat still.

If you are a visual thinker who has bounced off deep learning maths before, this is the book to try before you conclude the problem is you. The coverage is current in a way that matters. Transformers get a full treatment rather than a late addition. Diffusion models get their own chapter, which most competing texts still lack.

There is a chapter on why deep learning works, covering the loss landscape and generalisation questions that the field has still not fully answered, and Prince is refreshingly willing to say when something is not understood rather than papering over it with confident hand waving. There is also a substantial ethics chapter written with more seriousness than the genre usually gets. The notebooks matter too. Nearly every chapter has accompanying Python material, so when you have just read about backpropagation you can go and watch gradients flow rather than take it on trust.

This closes the loop that pure textbooks leave open and it is the reason this works as a self study resource rather than only as a course companion. What it is not. It is not an engineering book. You will not learn how to build a data pipeline, how to serve a model, how to debug a training run that is silently producing garbage, or how any of this survives contact with production.

Pair it with something applied. The reinforcement learning and graph network chapters are introductions and you will need dedicated resources if either becomes your area. And the maths is genuinely required. Not graduate level, but you need to be comfortable with matrix multiplication, chain rule differentiation and basic probability, and the book moves at a pace that assumes you are.

If you are not, spend a few weeks elsewhere first, because the failure mode here is concluding you are not clever enough when the actual problem is a missing prerequisite. My 4.8 is the highest I have given a written resource and the deduction is only for scope, since the engineering and applied side is absent by design. That it is free is close to absurd. There is a paid ecosystem of AI courses charging thousands of dollars for material that is thinner, less current and less clearly explained than a PDF you can download in ten seconds.

Download it, work through it with the notebooks open, and if it serves you well, buy the hardcover so that this model of publishing keeps being viable.

[ final ]

The verdict.

The best deep learning textbook available and it costs nothing. Read it alongside a practical course and you have a better foundation than most people who paid five figures for one.