Neural networks and other forms of machine learning ultimately learn by trial and error, one improvement at a time.