Answer 1

I think based on feedback, it adapts the parameters of the neural network to get a better approximation each time and with millions of training samples, the parameters can be optimally set for the task which is what is called learning.

Answer 2

I think that it learns on his past experiences. When it meets new problem, it solves it regarding its past knowleadge. And this way it learns constantly. He gets more and more knowleadge and becomes more and more smarter. But for that it need very large storage places for accumulating memory.