Edinburgh AI Workshop Resources

Edinburgh AI runs beginner-friendly AI workshops at the University of Edinburgh. This repo contains the notebooks, datasets, helper code, and solution files used across different weekly sessions.

Each workshop focuses on a concrete build or activity: a regression model, a neural-network classifier, CNN visualisation, language models, a podcast generator, object detection, sign-language classification, a Flappy Bird reinforcement-learning agent, a judo sequence model, or a reels recommender. Dataset and setup requirements vary by workshop.

Most sessions were designed around Kaggle because it gave students the easiest shared environment: browser notebooks, dataset attachment, optional GPUs, and minimal local setup.

Workshop Catalogue

Area	Workshop	What students build	Main stack
Foundations	Intro to Machine Learning	Linear regression and a quadratic-feature extension using embedded synthetic data	pandas, scikit-learn
Foundations	Neural Networks	A first PyTorch neural network and MNIST classifier	PyTorch, torchvision
Foundations	Computer Vision with CNNs	Convolution demos, satellite-image classification, transfer learning and face segmentation	PyTorch, torchvision, transformers
Foundations	Language Models	Embeddings, semantic search, sentiment fine-tuning and GPT-2 generation	gensim, sentence-transformers, transformers
Theoretical	CNN Class Visualisation	Images optimised to activate classes and intermediate layers of a frozen ResNet50	PyTorch, torchvision
Theoretical	Single-Object YOLO Detection	A raccoon detector with confidence/box loss and IoU evaluation	PyTorch, pandas, PyYAML
Theoretical	Where's Wally? Detection	Synthetic training images and a CNN with a heatmap and bounding-box head	PyTorch, torchvision, Pillow
Theoretical	Sports Vision with RNNs and LSTMs	Clip-level judo throw-attempt classification, plus a no-code beginner activity	PyTorch, NumPy, scikit-learn; browser app
Practical	Notes to Podcast	A local RAG pipeline that turns university notes into a two-host podcast	FAISS, Ollama, Kokoro TTS
Practical	ASL Interpreter	Sign-language classification using hand landmarks and image models	MediaPipe, OpenCV, scikit-learn, TensorFlow
Practical	Flappy Bird RL	A DQN agent trained to play Flappy Bird	Gymnasium, PyTorch
Practical	Instagram Reels Recommendation	An HSTU sequential recommender, plus a no-code recommendation activity	PyTorch, pandas; browser app and AI chat tool

How To Run A Workshop

Read the selected workshop's README first to choose a notebook and check its data requirements. The usual coding route is Kaggle:

1. Go to kaggle.com and create an account.
2. Create a new notebook.
3. Download the linked student .ipynb file from this repository and import it into the new notebook. If GitHub import is available, select EdinburghAI/workshops and the topic-based notebook path instead.
4. Turn on internet access when the notebook needs package, model or data downloads.
5. Attach the datasets listed in that workshop's README and check that the input paths match. Some workshops embed their data, download it in code or collect experience from a game instead.
6. Enable a GPU where the workshop recommends one.
7. Run the cells in order, complete the TODO prompts and use the solved notebook as a reference.

Each workshop folder has its own README with the exact notebooks, datasets, outputs, and takeaways for that session.

The beginner Sports Vision and Reels notebooks are activity guides using external browser apps; they do not require Python execution. Follow their workshop READMEs for the app links and instructions.

For local coding, use Jupyter or another notebook environment and follow the workshop-specific dependency and path guidance. The root pyproject.toml and uv.lock describe a Python 3.11 environment with a subset of dependencies; they are not a complete environment for every workshop. Notes to Podcast also uses Linux system packages and an Ollama service. Local runs have not been standardised across all topics.

Repository Layout

TheoreticalWorkshops/
  intro-to-machine-learning/
    legacy/v1-intro-to-machine-learning/IntroToML/
    legacy/v1-intro-to-machine-learning/generation/
  intro-to-neural-networks/
    legacy/v1-intro-to-neural-networks/
  intro-to-convolutional-neural-networks/
  intro-to-large-language-models/
  cnn-class-visualisation/
  yolo-single-raccoon-box-bounding/
    workshop-notebooks/
    yolo-workshop-api/
  yolo-wheres-wally/
    notebook-dataset/
  intro-to-rnn-lstm/
    legacy/

PracticalWorkshops/
  Notes-To-Podcast/
  ASL-Interpreter/
  Flappy-Bird/
  Instagram-Reels-Recommendation/

Workshops are organised by topic within theoretical and practical groups, rather than by semester and workshop number. The notebooks at each topic's main level (or its documented workshop-notebooks/ folder) are the main version. Earlier deliveries are retained under legacy/ where present. Difficulty variants within the main version are alternatives for the same workshop, rather than separate deliveries.

Notebook filenames retain their original spelling, including the double .ipynb extension on the CNN visualisation solution. Notebook titles may still show historical session numbers and dates. Use this catalogue and the topic READMEs for navigation.

Earlier Workshop Versions

• Legacy machine learning: CSV-based regression and decision trees; the main version uses embedded data and focuses on regression.
• Legacy neural networks: earlier beginner, intermediate, expert and solved notebooks.
• Legacy RNNs and LSTMs: frame-level judo detection using the older dataset; the main version classifies whole clips using version 2 data.

The raccoon workshop also retains an earlier solved beta notebook alongside its main notebook pair, as explained in its README.

Some notebooks and supporting notes still contain links to the previous repository layout. The relevant workshop READMEs provide current replacement links without changing those teaching materials. External datasets and browser apps are maintained separately; their live availability is not guaranteed by the local files.

Credits

Created by Pierre Mackenzie and Leo Camacho, with help from Valentin Magis, Niall Meagher, Finlay Ross, Conor O'Shea, and Edinburgh AI workshop contributors. Please credit Edinburgh AI if you use or adapt this material.