You know what sucks? When AI appears smart but its explaination is so over your head you don't even fully grasp if it is bullshitting or not.

For reference, what the following does is decomposes several runs of a network, takes them as samples, then generates a distribution with those samples. It then applies a fourier transform on the samples, to get the frequency components of the networks derivatives (first and second order), in order to find winning subnetworks to tune, and enforces a gaussian distribution in the process.

I sort of understand that, but the rest is basically rocket science to me.

Starts with an explanation of basic neural nets and goes from there. Most of the meat of the discussion is at the bottom.

https://pastebin.com/DLqe70uD