>>106560333
http://szudzik.com/ElegantPairing.pdf

>>106560765
It is a Godel encoding, making usage of a compiler/decompiler for nets I wrote, plus the above link.

>>106561159
It being a bijection is not meaningful. Most networks found are topologically invalid. Valid networks require incomprehensibly large numbers, the gap between any two of them may also be incomprehensibly large, and "valid" and "useful" are not the same thing. Using my method to find an MNIST digit classifier (fizzbuzz for backprop) would take longer than the lifespan of the universe.

The research "works" in the sense that you could find any neural network if you had infinite compute, but I had the idea and wrote the code while suffering severe mental illness and the longer I take schizophrenia medication the less interesting I find it. The only really interesting property is that training data is not needed. If a neural network exists, there is already a positive integer which encodes it, but finding that integer is a hopeless task unless there is some major breakthrough in computing.

>t. Eli Selig