Nature's fingerprint: neural plasticity, in man and machine.
Interviewee, Nawab Hussain: https://www.linkedin.com/in/nawab-hussain/
His thesis is on unsupervised entity mapping.
"So does this mean you can create SQL from an NSQL dataset?"
"Exactly", he says. Enabling not just self training models, but self structuring data.
His breakthrough was from another paper (Lample et al; Neural Architecture...), which removed words of their context and built the model based essentially at grapheme granularity.
"Hmmm interesting that a more primitive language segmentation improved accuracy... that reminds me of something", I told him.
A study measured the frequency of lines and curves as they appeared in alphabets. It found an almost exact frequency graph when overlaying all Latin based alphabets...then languages of the orient... then even photos of architecture.
Uncanny. Line orientation frequencies the same everywhere! Why?
Finally, they plugged in pictures of nature. It mapped over the other frequency graphs almost exactly. Astonishing?
Perhaps obvious. Nature's fingerprint as captured through five senses, stored in our 'wetware', then manifesting back in language. An echo.
You're using nature's fingerprint to train your machines, but remember they aren't confined to our senses, so the structure will be different.