neural network

A neural network typically starts out with random coefficients (weights); hence, it produce essentially random predications when presented with its first case. What is the key ingredients by which the net (neural network) evolves to produce a more accurate predication? (Please answer your question as clearly and concisely as possible.)