EdinburghAI Workshops
September 2026
A series of hands-on machine learning workshops designed and delivered for EdinburghAI, the University of Edinburgh's student-led AI society.
A series of hands-on workshops I designed and delivered for EdinburghAI, the University of Edinburgh’s student-led AI society. Three workshops have been delivered so far.
Each session opens with a 15-minute lecture and slides covering the theory, followed by a hands-on Kaggle notebook where attendees work through the core deep learning logic themselves. A provided API handles the boilerplate so that the focus remains on the concepts that matter.
The first two workshops map directly onto the technical stack behind Judo Clipper and were partly motivated by wanting to share what I learned while building it.
Workshop 1: Single Object Detection with YOLO
Attendees built a single-object YOLO detector in PyTorch from scratch, training it to locate raccoons in images.
The session covered:
- Converting images into tensors
- Normalising bounding-box coordinates
- Implementing confidence and coordinate loss
- Evaluating predictions using Intersection over Union
- Drawing predicted bounding boxes on held-out images
Attendees finished the workshop with a trained model capable of locating raccoons in unseen images.
Workshop 2: LSTMs, RNNs and Sequence Modelling
Building on the concepts introduced in Workshop 1, this session explored recurrent architectures for time-series data using an earlier version of the Judo Clipper pose dataset.
Attendees implemented both an RNN and an LSTM in PyTorch and trained them side by side. This allowed them to observe the RNN’s gradient explosion and compare it with the more stable convergence of the LSTM.
The session covered:
- Representing pose data as temporal sequences
- Implementing RNNs and LSTMs in PyTorch
- Understanding exploding gradients
- Using masked loss functions for missing pose detections
- Applying class-imbalance weighting
- Evaluating classification models using F1 score
Workshop 3: Short-Form Content Recommendation Algorithms
Delivered on 18 September 2026. The workshop explored how platforms like Instagram Reels decide what to show you, using Meta’s Hierarchical Sequential Transduction Unit (HSTU) as the core example.
The lecture covered:
- What Meta wants from Reels commercially: retention, ads and user data
- Why recommendation is multi-stage when users view over 200 billion Reels a day
- Lightweight candidate retrieval (two-tower models and user/reel graphs)
- Heavier ranking with HSTU, which treats a scrolling session as an alternating sequence of videos and actions
Both tracks shared a live doomscrolling app with around 300 Instagram reels. Attendees scrolled through it, exported their session as JSON, got personalised recommendations from their model, and then watched those picks back in the app. I ran 2 tracks so the session worked for any background:
- Beginner (no coding required): Attendees pasted their scroll log into a
structured prompt that turned an LLM into an HSTU-style recommendation engine,
complete with sequential reasoning rules and a weighted scoring formula
(
1.5·P(Watch) + 2.0·P(Like) − 3.0·P(Skip)). They then learned how engineers evaluate recommenders with ranking metrics such as Hit@10. - Advanced (PyTorch): Attendees built and trained their own HSTU, covering:
- Item and event encoders combining content embeddings with watch and action features
- HSTU blocks with pointwise SiLU attention instead of softmax, relative position and time biases, gating, and causal and padding masks
- A chronological 80/10/10 train/validation/test split to avoid learning from the future
- Next-item training with cross-entropy loss over a 31,496-video catalogue
- Evaluating with Hit@10, reaching 0.34% on the test set, about 10.7x better than random
- Transferring the trained weights to a separate catalogue of 298 Instagram reels and recommending from their own scrolling history
Attendees on both tracks finished with a working recommender they could test on their own behaviour.
Resources
All workshop materials are available through the following repositories: