Neural NetworksComputational Graphs

A computational graph draws a calculation as a network of small steps. Each node does one simple job — multiply, add, or apply the activation — and the arrows carry numbers from node to node. Here is Riya's neuron as a graph. A node can be worked out only when every number flowing into it is ready. Tap the nodes in an order that works, then do the same for a school day.

Working through a graph from the inputs to the output is called the forward pass. A whole neural network is simply a much bigger graph built from the same three kinds of node: multiply by a weight, add, activate. Chain neurons together, so that the outputs of one layer become the inputs of the next, and the network can learn complicated patterns — for numbers (regression) as well as for kinds (classification).