Identify what comes from the input and what training changes.
Trace an error back to each adjustable coefficient.
Understand what a training fit says about new data.