"After training on 12,000 patterns it becomes obvious that the new network will outperform the old… If our predictive method gives a good quantitative estimate of the network's test error, we can decide whether three weeks of training should be devoted to the new architecture"

→ View original post on X — @jxmnop,