Learning as hill-climbing in weight space