So why not do this directly?
->Requires differentiating a graph
Why not use only one Network?
Graph->Vector->pdf
->Finds trivial solution, as $<pdf> \propto \frac{1}{\sigma_{Vector}}$
So regularise the standart deviation of the vector space!
<l2st>
Interplay between encoding and NF
Could also be useful for highdim data
</l2st>