Attacking Pseudorandomness With Deep Learning