Overview

Every method in Topic 5 achieved non-linearity either by partitioning feature space geometrically (margins) or by fixing a similarity function up front (kernels). Neural networks take a third route: stack layers of simple, adjustable non-linear units, and let gradient-based optimisation learn the feature representation jointly with the final prediction, rather than designing or fixing it in advance. This topic builds that machinery at the level a classical-ML course needs — architecture, the training algorithm, practical stabilisation techniques, and a first look at convolutional, recurrent, and attention-based priors — while deliberately stopping at survey depth for frontier material. Deeper architecture labs and generative modelling live in the Neural Networks and AI unit.