Machine Learning and Color

Data, source code, documents and apps related to machine learning and color.

This material complements the recent CIC32 short course #09.

→ attribution

First, why machine learn colors?

The MNIST handwritten digits is a widely used machine learning dataset.

But color provides a nice complement to the MNIST data in the following ways :

colors	handwritten digits
A categorization problem	A classification problem
No fixed number of categories	Fixed number of classes
Includes multi-category (bluish green)	One class per digit
5 to 10% adversarial data	No adversarial data
Includes NSFW labels	Entirely SFW labels
Red, green & blue as input	Image input
Upper accuracy of 88%	Upper accuracy of 99.8%
Token label	Character label
Many misspellings	Not applicable
Includes hapaxes	Not applicable
A direct visualization attribute	Not applicable
Categorical context	Nominal context (zip codes)
Compatible with 3D LUTs	Not applicable

---

Some entry points for this repository :

• apps : if you want to try interactive versions of the code
• images : if you want a 'visual index' of the code
• src : if you want a bullet list of the code