“Gradient descent, when training a neural network, may be implicitly searching for and finding an efficient sparse model within the larger parameter space.”