The Little Book of Deep Learning

This book is a short introduction to deep learning for readers with a STEM background, originally designed to be read on a phone screen. This is the v1.2 updated on May 19th, 2024.

It is distributed under a non-commercial Creative Commons license and was downloaded 500'000 times in eight months.

Download the phone-formatted pdf Buy a $9 paperback copy on lulu.com Download a printable A5 booklet pdf

Note that the version sold on Amazon and other on-line platforms for $40 is "unauthorized": someone stole the content and sells it. Preventing this would require to sue and I have better things to do with my time.

Four
pages of the book in low-resolution

Updates

V1.2 (May 19, 2024)

Chapter 8. New chapter on low-resource methods (prompt engineering, quantization, low-rank adapters, model merging).
Miscellaneous. Changed "meta parameter" to "hyper parameter".
Section 3.6. Added a sub-section about fine-tuning.
Section 4.8. Added the note about the quadratic cost of the attention operator
The missing bits. Added a note about the O(T) of standard RNNs vs. the O(log T) of methods that leverage parallel scan.

V1.1.1 (Sep 20, 2023)

Section 4.2. Added a paragraph about the equivariance of convolution layers.
Section 5.3. Fixed the description of the original Transformer, and modified Figures 5.6, 5.7, 5.8, and 5.9 accordingly.

V1.1 (Sep 8, 2023)

Miscellaneous. Fixed minor typos and phrasings.
Section 1.3. Reformulated the text to clarify that overfitting is not particularly related to noise, but to any properties specific to the training set, as it is the case on the Figure 1.2.
Section 3.2. Clarified the phrasing and changed the Figure 3.1.
Section 3.4. Fixed the indexing of the mappings in the example of a composition.
Section 3.7. Fixed the label "1TWh" in Figure 3.7, that should be "1GWh".
Section 4.5. Added a figure to illustrate the functioning of 2D dropout.
Section 4.6. Changed the Figure 4.8 so that in the top part illustrating the re-scaling / translating after normalization, the highlighted sub-blocks correspond to groups of activations that are re-scaled / translated with the same factor / bias.
Section 6.6. Restricted the Figure 6.4. to three sub-images to make the text more legible.
Section 7.1. Added two paragraphs to introduce the notion of Reinforcement Learning from Human Feedback.
The missing bits. Removed the fine-tuning sub-section, since most of it was moved to Section 7.1.