What training changed

The network began with layers and connections arranged in a fixed structure. Inside it were adjustable numbers.

Before training, those numbers had unhelpful starting values. The network made poor predictions.

Each correction changed the numbers slightly. The layers stayed in place. The routes through the network stayed in place.

After many corrections, the same structure held a useful arrangement of numbers.

Together, the structure and its current settings form a .

Training changed the settings, not the shape of the machine.