Editing Learning (section)

====Color Representation Ideas====
* See what effect reducing the bits of colour has on accuracy... GreyScale vs R8G8B8 vs R8G4B4, etc...
* Dumb & basic - Red*255*255+green*255+blue. Should reduce the number of channels but still have precision bits left over. Maybe it needs to be offset to the centre though to help future layers...
* Would it be possible for a Float32 to be used to encode 8 bits of the next color?
* [https://allrgb.com/ AllRGB] - Shows images with one of every colour in it...
* The way the human eye splits things up? That's like 6 separate images...
* More natural mapping...
* Maybe try and keep similar colours together distance wise? Hard to do with a single dimension float.
* Hilbert Curve
* Periodic algorithms? - [https://allrgb.com/are-they-parallel Sin,Cos,etc...], [https://allrgb.com/bling Repeating patterns], [https://allrgb.com/color-corners Partial gradient+periodic]. Isn't RGB basically just periodic anyway?
* Float32 using 4d color space (but then some colours would be duplicated...]
* 2xFloats merge red&blue?, green&red? (how does the eye do that?)
* Separate out brightness?
* Try an optimisation function similar to an embedding?
* Maybe try and make similar colours further apart to exaggerate the subtle differences?
* Negatives could cause the 'near' dimensional to change... Like having an extra dimension. Would duplicate colours again. 
* What about that Kaggle competition with the 16 band sate-lights?...
* Maybe this is all useless. The convolution layer should just learn what it needs anyway... It's all probability based so it shouldn't need to be too close. This could all hurt it.