Has anybody given a serious go at using log-normal initializer for deep neural network parameters? Seems likely to make sense. Additionally, one can also do batch- and layer- log-normalization on layer activations.