MLModelsFine-Tuning

The Little Book of Deep Learning

by François Fleuret (University of Geneva)

IntermediateBookFree~185-page phone-format PDF; about 8-12 hours of focused reading

A dense, free ~185-page deep learning primer that takes you from losses and backprop to transformers, diffusion, quantization and adapters.

Start LearningAdded Sep 25, 2026 · Updated Sep 25, 2026

Overview

The Little Book of Deep Learning is a compact deep learning textbook by François Fleuret, Full Professor and head of the Machine Learning group at the University of Geneva. It is formatted to be read on a phone screen and has been downloaded more than a million times since its April 2023 release. Version 1.3, published May 12, 2026, corrects typos and phrasing and adds a new section on large-scale parallel training (Section 3.8). The book has about 185 pages in three parts. Part I, Foundations, covers learning from data, basis function regression, under- and overfitting, GPUs/TPUs and tensors, losses, autoregressive models, gradient descent, backpropagation, the value of depth, training protocols, the benefits of scale and large-scale parallel training. Part II, Deep Models, covers layers (linear, activations, pooling, dropout, normalization, skip connections, attention, token embeddings, positional encoding) and architectures: MLPs, convolutional networks and attention models including GPT and ViT. Part III, Applications, covers image denoising, classification, object detection, segmentation, speech recognition, CLIP-style text-image representations and reinforcement learning. It continues with text and image generation, including denoising diffusion, and a 'Compute Schism' chapter on prompt engineering, quantization, adapters and model merging. It closes with 'The missing bits', a bibliography and an index. Jargon is underlined and linked to its definition. The book is released under a non-commercial Creative Commons license as a phone PDF and a printable A5 booklet, with a $9 paperback also sold on Lulu.

At a Glance

Topic: ML
Level: Intermediate
Format: Book
Cost: Free
Duration: ~185-page phone-format PDF; about 8-12 hours of focused reading
Provider: François Fleuret (University of Geneva)
Hands-on: No
Certificate: None

What You’ll Learn

✓How losses, gradient descent and backpropagation fit together when training deep networks
✓Why depth and scale help, including scaling laws and the costs of training models
✓How large models are trained in parallel across many GPUs, from the new v1.3 section
✓How attention layers, token embeddings and positional encodings combine into GPT and ViT
✓How normalization layers, skip connections and dropout keep deep networks trainable
✓How denoising diffusion and autoregressive text generation work at the level of the math
✓What quantization, adapters, model merging and prompt engineering change once a model is trained
✓How the same building blocks carry over to detection, segmentation, speech and CLIP-style models

Highlights

•Downloaded more than 1,000,000 times, and still maintained: v1.3 shipped May 12, 2026 with a new large-scale training section
•Unusually short for its breadth: foundations through transformers, diffusion, quantization and adapters in about 185 pages
•Built for reading on a phone, with jargon underlined and linked to its definition
•Free under a non-commercial CC license. Hacker News readers note it leans mathematical, so it suits readers with a CS-theory background

Who It’s For

Best For

✓Software engineers building on LLMs who want the underlying deep learning theory in one short read
✓ML practitioners who want a concise refresher before an interview or a paper-heavy course
✓Students with a STEM background moving into machine learning
✓Engineers who want to understand what quantization and adapters actually do to a model

Prerequisites

•A STEM background: comfort with linear algebra, calculus and basic probability notation
•Some programming exposure helps for the tensor and GPU sections, although the book contains no code exercises

FAQ

What is The Little Book of Deep Learning?

The Little Book of Deep Learning is a free, short book by François Fleuret, a machine learning professor at the University of Geneva. It is for engineers and students with a STEM background who want a compact, math-literate overview of deep learning. Afterwards you should be able to read transformer, diffusion and fine-tuning papers without getting lost.

Is The Little Book of Deep Learning free?

The Little Book of Deep Learning is free to access.

What level is The Little Book of Deep Learning for?

The Little Book of Deep Learning is aimed at a intermediate audience. Recommended background: A STEM background: comfort with linear algebra, calculus and basic probability notation, Some programming exposure helps for the tensor and GPU sections, although the book contains no code exercises.

How long does The Little Book of Deep Learning take?

Expect roughly ~185-page phone-format PDF; about 8-12 hours of focused reading. Most learners work through it at their own pace.

What will I learn from The Little Book of Deep Learning?

You'll learn: How losses, gradient descent and backpropagation fit together when training deep networks; Why depth and scale help, including scaling laws and the costs of training models; How large models are trained in parallel across many GPUs, from the new v1.3 section; How attention layers, token embeddings and positional encodings combine into GPT and ViT; How normalization layers, skip connections and dropout keep deep networks trainable; How denoising diffusion and autoregressive text generation work at the level of the math; What quantization, adapters, model merging and prompt engineering change once a model is trained; How the same building blocks carry over to detection, segmentation, speech and CLIP-style models.

Topics

deep learningtransformersfree bookneural networksdiffusion modelsquantization

Sources

This page was written from 3 sources, 1 on domains other than fleuret.org.