Directly modeling and generating neural network parameters as a distribution rather than training networks end-to-end.