Simpler Neural Networks by Fan-In Reduction