as an experiment, try this:
go on https://color.adobe.com and click on one of the color rules (it will give you a random palette based on the rule)
now compare with a palette from https://coolors.co/ or https://color.adobe.com/explore/ (user-uploaded)
if color theory + regression fully solved this problem, none of these color sites would need to exist.
I really liked colormind.io, and made a little toy Vue site when I was looking for a job (I grabbed your colors and pushed them to state so my website colors changed). See here: http://q8z8p.net/#/color. I just wanted to thank you because I think that was helpful in getting my first job!
That got me wondering if the process going on for colourmind could be turned into a new theory on colour. Is there a way to boil the process down into a deterministic one, or do you feel like the neural network is accomplishing something that couldn't be refined into a "rule" or guidline to use in a regular theory?