In the first unsupervised set, the objective was defined as preserving the relative distances amongst examples in the same batch. In other words, for a given set of examples, the model would compare how far apart each pair of colors were in RGB space and in the discovered 1-dimensional space. It tried to keep these pair-wise distances similar to each other.
At some point during its discovery, the model comes across the concept of luminance (brightness), before settling into a discovery path that blends this concept in combination with grouping like-colors. Many attempts turned out with blue-brown-yellow characteristics, while some landed closer to a sorting-by-brightness result.
