A very interesting idea. I am curious about this sharing and blending of the various nets; I wonder if something as naive as averaging the weights (assuming the neural nets all have the same dimensions) would actually accomplish that?