Artificial-Intelligence-Courses

The source repository for all KAUST Academy Artificial Intelligence courses — LaTeX/Beamer lecture decks, their built PDFs, and the accompanying Jupyter lab notebooks, homeworks and exams.

Courses	5
Lecture decks	64 (.tex sources + built PDFs)
Slides	~4,800 pages
Notebooks	291 (labs, homeworks, exams)
License	GPL-3.0

---

For students and instructors

Courses

Built PDFs live in Lectures/<Course>/; the LaTeX sources that produce them live in LaTeX/<Course>/. Notebooks for each course live in Labs/<Course>/.

Course	Decks	Pages	Slides	Notebooks
Computer Vision	24	1,964	Lectures/Computer_Vision	56
Natural Language Processing	21	1,530	Lectures/Natural_Language_Processing	46
Reinforcement Learning	10	710	Lectures/Reinforcement_Learning	36
Introduction to AI	8	552	Lectures/Introduction_To_AI	24
Generative AI	1	92	Lectures/Generative_AI	—

Computer Vision — 24 decks

#	Deck	Pages
01	Introduction to Computer Vision & CNNs	100
02	Practical Deep Learning	64
03	Classic Deep CNN Architectures	71
04	Object Detection	141
05	Image Segmentation	93
06	Recurrent Neural Networks	53
07	Transformers	42
08	Vision Transformers	40
09	Deep Unsupervised Learning	117
10	Autoencoders & Variational Autoencoders	147
11	Autoregressive Models	90
12	Normalizing Flow Models	81
13	Generative Adversarial Networks (GANs)	125
14	Diffusion Models	94
15	Advanced Image Generation Models	66
16	Stable Diffusion	82
17	Learning from Videos	112
18	Video Generation & World Models	51
19	Self-Supervised Learning	51
20	Contrastive Learning Methods	71
21	Vision and Text Integration	91
22	Advanced Self-Supervised Learning and JEPA Models	54
23	Foundation Models: Modern Advances and Applications	68
24	World Models	60

Natural Language Processing — 21 decks

#	Deck	Pages
01	Introduction to Natural Language Processing	94
02	Vector Space Models & Word Embeddings	81
03	Recurrent Neural Networks (RNNs)	89
04	Sequence-to-Sequence Models: Intro to Attention	76
05	Attention Mechanism Deep Dive	50
06	Introduction to Transformers	101
07	Large Language Models	66
08	Fine-Tuning LLMs and RLHF	48
09	Prompting & Retrieval-Augmented Generation (RAG)	103
10	Multimodal NLP	33
11	Agentic AI	65
12	AI Safety for Agents	59
13	Mixture of Experts Models	74
14	Large Reasoning Models	80
15	RL Post-Training for Reasoning (GRPO & GSPO)	45
16	Transformers: 2017 vs 2026	47
17	Inference Optimisation for Large Models	56
18	Recent Advancements in NLP	73
19	Audio Processing in NLP	161
20	Speech-to-Text (STT / ASR)	79
21	Text-to-Speech (TTS)	50

Reinforcement Learning — 10 decks

#	Deck	Pages
01	RL Foundations: MDPs & Bellman Equations	76
02	Value-Based Methods: Q-Learning to DQN	40
03	Vanilla Policy Gradient & REINFORCE	79
04	Policy Optimization: Actor-Critic to PPO	89
05	Continuous Control I: Deterministic Policy Gradients & DDPG	58
06	Continuous Control II: Max Entropy RL, SAC	62
07	The Reward Problem: Exploration vs. Exploitation, Bandits, Inverse RL	74
08	Data & Planning: Model-Based & Offline RL	114
09	RL in the Real World: RLHF, Multi-Agent RL & Robotics	54
10	RL Frontiers: Meta-RL, Multi-task RL, Hierarchical RL, Open Problems	64

Introduction to AI — 8 decks

#	Deck	Pages
01	Data Science Foundations	70
02	Machine Learning Algorithms	89
03	Fundamentals of Deep Learning	59
04	Unsupervised Learning	37
05	Data Preprocessing and Data Augmentation	104
06	Decision Trees and their Variants	60
07	Linear Regression	83
08	Support Vector Machines (SVMs)	50

Generative AI — 1 deck

#	Deck	Pages
01	Generative AI for Science — Applications and Techniques	92

Labs

291 Jupyter notebooks. Course folders mirror the lecture tracks; the rest are cross-cutting.

Folder	Notebooks	What it is
Labs/Computer_Vision	56	CNNs through generative and foundation models
Labs/Incomplete_Labs	51	Mixed — see the note below
Labs/Homeworks	54	27 assignment/solution pairs — CV 14, NLP 6, ML 5, RL 2
Labs/Reinforcement_Learning	36	15 exercise/solution pairs, in course order
Labs/Natural_Language_Processing	46	Classical text through agents and RAG
Labs/Introduction_To_AI	24	Classical ML and DL foundations
Labs/Exams	20	10 question/solution pairs across 2025 and 2026 cohorts
Labs/Archive	4	Superseded material

Incomplete_Labs/ is a staging area, not an archive. It holds labs that still need testing or changes before they're finalised.

Notebook naming. A student version is <Name>_Exercise.ipynb and the worked version is <Name>_Solution.ipynb — the same rule in every folder, including Exams/ and Homeworks/. A solution is always its counterpart's name plus the suffix, so pairs match by stem and can be found mechanically. Notebooks carry no numeric prefix; ordering comes from the course.

---

For contributors

Requirements

• A LaTeX distribution — TeX Live or MiKTeX
• latexmk (drives the build)
• pdfinfo (from poppler; used for the page-count report)
• Bash

Shared packages are declared in LaTeX/preamble/packages.tex. Builds pass -shell-escape because preamble/commands.tex defines figure-fetching macros (\fetchimage, \convertimage, …) that shell out to curl and ImageMagick's convert. Each is wrapped in \IfFileExists, so with the figures committed a normal build never invokes them — you only need curl and ImageMagick if you add a macro-fetched figure.

Building

Run from the repository root.

./build.sh                                              # all 64 decks
./build.sh --file Computer_Vision/04_Object_Detection.tex   # one deck
./build.sh --prefix 01                                  # deck 01 of every course
./build.sh --output some/dir                            # choose the destination
./build.sh --keep-logs                                  # keep .aux/.log instead of cleaning up

--file takes a path relative to LaTeX/, i.e. <Course>/<deck>.tex. Those four flags, plus -h/--help, are the complete set; anything else exits with an error.

Per deck you get either

ok  pages=141  overfull_vbox=0  missing_images=0

or a FAILED line, with the full log copied to build/logs/<deck>.log. The script exits non-zero if any deck failed.

build.sh writes PDFs flat into Lectures/, but the committed layout is Lectures/<Course>/. A plain ./build.sh will leave 63 PDFs at the root of Lectures/ alongside the course folders, rather than updating them in place. Build into the right folder explicitly:

./build.sh --file Computer_Vision/04_Object_Detection.tex --output Lectures/Computer_Vision

Repository layout

.
├── build.sh                  # the only build entry point
├── update_version.sh         # semver bump, called by CI
├── VERSION  CHANGELOG.md     # both maintained by CI, not by hand
├── CONFIGURE.md              # environment setup notes
├── Lectures/<Course>/        # built PDFs, committed
├── Labs/                     # 291 notebooks
└── LaTeX/
    ├── Computer_Vision/            # 24 deck main files
    ├── Natural_Language_Processing/#  20
    ├── Reinforcement_Learning/     #  10
    ├── Introduction_To_AI/         #   8
    ├── Generative_AI/              #   1
    ├── preamble/             # packages, commands, beamer_settings (4 files)
    ├── sections/             # 65 topic dirs, ~950 .tex — all slide content
    ├── images/               # 69 dirs, ~2,300 figures
    ├── style_files/          # logos and .sty helpers
    ├── beamerthemeStanford.sty + 2 more, antbrief.cls
    └── references.bib

Naming conventions. Decks are NN_Title.tex, zero-padded, in every course, and the built PDF keeps the same stem in Lectures/<Course>/ — so a deck and its PDF always share a name. Notebooks carry no numeric prefix and use <Name>_Exercise.ipynb / <Name>_Solution.ipynb. No path anywhere in the repo contains a space.

How a deck is assembled

A deck main file is a thin shell — all content lives in sections/. Every one of the 63 decks follows the same shape:

\documentclass[10pt, aspectratio=169]{beamer}
\input{preamble/packages}          % all 63 decks
\input{preamble/commands}          % all 63 decks
\input{preamble/beamer_settings}   % all 63 decks
\begin{document}
\input{sections/cover}             % all 63 decks
\input{sections/toc}               % 28 of 63
\input{sections/<topic>/<file>}    % ~13 of these per deck
\end{document}

So:

• New slide → add or edit a file under LaTeX/sections/<topic>/, then \input it from the deck. Don't put slide content in the deck main file.
• New figure → LaTeX/images/<topic>/, referenced as \includegraphics[width=\linewidth]{images/<topic>/<file>.png}. Image directory names mostly mirror section names, but that's a convention, not a rule.
• New package → LaTeX/preamble/packages.tex, so every deck picks it up. 12 decks currently re-declare a package locally (mostly tikz); treat that as legacy rather than a pattern to copy.

Width note: use \linewidth, not \textwidth or \paperwidth. Inside a list \linewidth accounts for the indent; the others overflow the frame by exactly that amount.

Releases and versioning

.github/workflows/deploy.yml is the only workflow. It does not build LaTeX — no PDF is compiled or published by CI. It bumps the version and regenerates the changelog:

• Triggers on push to any branch without a / in its name, plus manual dispatch (workflow_dispatch ignores the branch filter). There is no pull_request trigger. So a flat branch like nlp_review2 fires it on every push, while user/topic never does.
• Keywords in the commit message drive it:

Keyword	Effect
[major] / [minor] / [patch]	Selects the semver bump. Omitting it defaults to patch.
[skip ci]	Skips the workflow entirely
tag-release	Cuts a GitHub Release — only on main, and not on a merge commit
tag-repo	Tags the repo — only on main, and not on a merge commit

• Merge commits are exempt from all of it. The bump, the changelog, the tagging and the release are each gated on the head commit having fewer than two parents, so a PR merge landing on main produces nothing at all. The bump happens earlier, on the push to the topic branch — which is why the bot commit appears as the branch-side parent of the merge.
• The workflow commits VERSION and CHANGELOG.md back as chore: update changelog and version [skip ci]. Expect a bot commit on your branch and to need a git pull before your next push.
• CHANGELOG.md is generated by conventional-changelog — commits whose subject doesn't parse never appear in it.

Commit convention

Follow Conventional Commits: <type>(<scope>): <description>, imperative mood.

feat(slides): add GRPO derivation to RL post-training deck
fix(cv): correct dead ATIS dataset link
docs(readme): document the CI version keywords

Types in use here: feat, fix, docs, refactor, style, chore. Add [minor] or [major] when the change warrants more than a patch bump.

There is no commit linter, and a malformed prefix fails silently — feat(slides):: … and feat(slides)L … are both in the history and neither reached the changelog.

Things that will catch you out

• Lectures/ is build output but is committed. Because a plain ./build.sh writes flat (see above), it does not update the committed PDFs — it drops 63 new files at the root of Lectures/, next to the course folders. Build with --output Lectures/<Course> and you get the opposite problem: hundreds of MB of binary diffs, most differing from HEAD only by an embedded timestamp. Either way, don't reflexively git add -A; stage the decks you changed.
• The repo is large — roughly 4.5 GB of git objects, ~720 MB of figures and 263 MB of notebooks. A shallow clone (--depth 1) is much faster if you don't need history.
• build.sh skips a SHARED_DIRS list that names assets, which doesn't exist, and does not skip LaTeX/build/. That's harmless today (no .tex there), but a stray .tex dropped in LaTeX/build/ would be compiled as if it were a deck.

License

GPL-3.0 — see LICENSE.