D4 · Deep Learning
Double Descent
Read a double descent curve and say where the classical story stops working.
Make the model bigger and the test error gets worse. Keep going, and it gets better than it ever was. The textbook curve has one dip. Real networks have two.