Slides
Lecture slides for each day
The deck used in each lecture, as a PDF. Each one is released on the morning of its teaching day. Until then, the practical notebook and dataset for that day are already on the practicals page.
Parts I to IV. A clothing catalogue needs 70,000 photographs sorted. The day builds the network that does it, from one unit to a forward pass that ends in a single number, the loss.
Backpropagation and optimisation turn Day 1's loss into an update, and generalisation and evaluation ask whether the result holds on unseen data. The day then gives the network an image: why images need convolution, and how a convolutional network reaches a class decision.
From the CNN class decision to language: what a language model receives and predicts, how text trains it and generation uses it, attention over earlier tokens, and the Transformer block. The web slides cover AI in research practice.
Model inputs and tiers, agents and orchestration, reproducible documents, hosted and local models, development workflow, voice input, APIs, browser automation, and four applied notebooks.
