Deep-Learning Architectures

Two copies of the same tiny network, same data, same everything — except one trains with a learning rate of 0.5 and the other with 10. The first solves its task in a few hundred steps; the second thrashes, its loss curve pinned at the ceiling, and never learns at all. Nothing was "wrong" with the second network. One number killed it. Training is full of silent failures like this — runs that look busy while learning nothing — and telling a healthy run from a sick one is a skill this track teaches deliberately, not one you are left to absorb by accident.

Neural-Network Foundations left you with a wired network and a loss score; this track makes the score go down. Backpropagation shows how blame for an error flows backwards through the layers — which weight, exactly, should change, and by how much. The optimiser modules walk the road from plain SGD to Adam, each step fixing a concrete failure of the last. Then regularisation and normalisation: the disciplines that keep a network learning the pattern rather than memorising the examples, and keep deep stacks trainable at all.

With the training loop in hand, the track widens to the two classic architecture families — CNNs for text and sequence models — the ideas the Transformer was invented to beat, and the reason its design choices make sense. Everything converges in the capstone: a training-dynamics clinic where you read real loss curves, diagnose sick runs (that learning-rate disaster above is one of the patients — as is an unlucky initialisation that plateaus in a local minimum), and meet the platform's first real PyTorch exercises.

The two-lane structure continues throughout. The core modules stay plain-spoken and interactive — everyone can follow the intuition lane end to end. Paired deep-dive modules on backpropagation, optimisers and regularisation (IT/ML lane) are optional detours into the full mathematics. Finish the clinic and the jump to the Transformer is short: P10 trains with exactly the machinery you just debugged.