Fetching the paper…

Parallel training of DNNs with Natural Gradient and Parameter Averaging · Around