negative training sets?

[JohnyDL]

I was thinking that some colours are really difficult to classify, and there's a lot of human perception and monitor differences. A good eye coming at a colour from an unbiased location might spot a very pale grey with a blue tint and say 'blue-ish' while a less discerning person with a worse monitor, who'd just seen a vivid blue, might ignore an even stronger tint and say that colour is 'grey-ish'. The line Grey/Pink/Red or Orange/Brown/Dark Red might similarly be hard to categorize,

So my thought was Negative training

255, 255, 000 is Yellow and so is definitely Yellowish but 192, 192, 128, the numbers tell me that's yellow but it's certainly not as yellow as 255, 255, 000, and looking at it well it's more grey, and maybe more green and I wouldn't be confident in saying one way or another between the options but while we/I can't agree on what it is, pretty much everyone will agree it's not white, blue, red, pink etc.

And maybe it's a better way to teach a ML colour picker the boundaries between colours anyway by allowing those ambiguous colours (or all colours) to be marked as "not colour-ish" for as many or as few options as the person feels confident in saying, the NN can learn and limit what they chose to include in colours from the whole space on a per colour level rather than the whole colour space just being divided up by ridged lines and there may be some points that the NN says "I don't know" for because multiple colours overlap or because it's not confident any colour it has access to fills that spot but that's fine cause humans can't tell some colours either, and some humans might say that a particular RGB value doesn't correspond to anything given.

You also get the ability to limit options by saying "not colour-ish" You can say give an RGB colour ask the question "is this Bluish?" have the user answer Yes, No, It's Unclear, And in the case of unclear have them pick from one of not Yellowish, not Pink-ish and not grey-ish allowing you to force more responses where you're struggling to get data.