Text size
Chapter 05 · Section II · 16 min read
How neural networks are built
Layers, weights, activation functions, and the back-propagation algorithm that made deep learning practical.
Stacking neurons into layers is easy. Choosing the right layers, the right number of neurons, and finding the right weights — that’s the engineering.
Back-propagation is the algorithm that does the weight-finding. It works by computing how wrong the network’s prediction is, then walking that error backwards through the layers, adjusting each weight a tiny bit in the direction that would have made the error smaller. Repeat a million times.
This section is a stub. The full version will cover layer types, activation functions, and a worked back-propagation example.