Learning Compact Neural Networks Using Ordinary Differential Equations as Activation Functions