P6 · Deep Learning & Neural Networks

from the neuron to modern architectures

From one artificial neuron to trained networks: forward passes, loss, backpropagation, optimisers and regularisation — the bridge between classical ML and the Transformer.

This is the bridge pillar. Classical machine learning (P5) taught you models whose shape a human chooses; the Transformer (P10) is a model whose shape is neurons all the way down. P6 crosses the gap in two tracks: Neural-Network Foundations builds the parts — one perceptron, activation functions, a two-layer forward pass, a loss to score it — and Deep-Learning Architectures makes those parts learn, from backpropagation through SGD-to-Adam and regularisation to CNNs and sequence models. The finale is a training-dynamics clinic where you diagnose real, misbehaving training runs the way a practitioner does.

Every core module runs on two lanes. The base lane is hands-on and plain-spoken — each concept comes with a live trace you can perturb, no calculus required. Alongside it sit optional deep-dive modules (marked for the IT/ML profile) that carry the full mathematics for readers who want the derivations. Take the base lane end to end and you arrive at P10's attention chapter knowing exactly what a "trained network" is, because you will have watched one train.

The pillar's arc: build the parts, wire them into a network, train it, then debug training like a practitioner.