ML strategy

Question 1

fit training set well on cost function, if not perform well what to do?

Answer 1

bigger network, Adam

Question 2

fit train set well but not dev set?

Answer 2

regularization, bigger train set, early stopping

Question 3

fit dev set well but not test set?

Answer 3

bigger dev set

Question 4

performs well on test set, not on real world?

Answer 4

change dev set or cost function