“Limiting the number of training examples is important because it shows that the algorithm is able to generalize like the human brain,” says Phoenix.