Q: initialisation for local likelihood approximations

Hi,

Just wondering how are the approximate factors initialised? 

For example, in the neural network experiments, the Xavier init scheme is used presumably for the mean of the global q. How about the variances of the global q, and the mean/natural parameters of the local approximate factors?

Many thanks,
Thang.