Feature Request: Implementation of SET training method

[DPR-Sanchez]

Could Spare Evolutionary Training be implemented as a training method? From reading up on it, from my understanding it is somewhat similar to Adam, but I likely am grossly oversimplifying things.

I have included some reading material if there is interest:
https://phys.org/news/2018-06-ai-method-power-artificial-neural.html

I am up for trying to tackle this for a pull request if you would like assistance.

[itdxer]

That's an interesting method, before saying anything I'll need to read the paper first. I did only a quick look through it and I think there might be a few problems with integration. For example,

With SET, the bipartite ANN layers start from a random sparse topology (i.e. Erdös–Rényi random graph24), evolving through a random process during the training phase towards a scale-free topology.

NeuPy doesn't support efficient sparse connections and the whole implementation might be quite problematic and very inefficient.

It might work as a standalone network with a fixed architecture like in case of the SET-RBM (since in that case it's much easier to make it efficient), but then again, I have to read through the paper in order to be sure.