Library · book

Understanding Deep Learning

Simon J.D. Prince
2023·MIT Press

Source: https://udlbook.github.io/udlbook/

The rigorous companion to any hands-on LLM build.

Prince's freely available textbook does the thing that separates understanding from familiarity: it makes the mathematics of deep learning visible, from gradient descent to self-attention, with figures that turn abstractions into objects you can reason about.

For directing technology around AI, that matters because the consequential decisions — about cost, context, precision, capability — bottom out in this mathematics, and borrowed intuition eventually fails you.

It pairs naturally with a from-scratch build: construct the model to own the mechanics, read Prince to own the reasons.

Central argument

A modern, comprehensive textbook on deep learning that prioritizes intuition and clean mathematics over encyclopedic coverage. It moves from supervised learning, loss functions and gradient descent through convolutional and residual networks to transformers, self-attention, and the practical realities of training deep models, with a consistent figures-first pedagogy designed to make abstract objects visible.

Critique

It is a foundations text, not a systems or frontier one: it explains why a transformer works far better than it explains how a lab trains and serves one at scale (data pipelines, RLHF, distributed compute, inference economics). Read for the mathematics of the mechanism, not for the operational last mile.

Why it matters for product

For a CTO/CPO who studies fundamentals close to the material, this is the rigor track that keeps intuition honest. When a decision hinges on why quantization degrades a model, why context length is quadratic, or what finetuning actually moves, Prince gives the load-bearing mathematics in a form you can hold in your head — the 'why' beneath the 'how'.