A Scale Mixture Perspective of Multiplicative Noise in Neural Networks