It is possible to find a reasonably similar representation of the same points in 2D space, making it possible to see what this looks like on a “map”. This view gives us special insight into how the data is being processed by the model, SDO FM.
Using the visual below, we can see how these points group together, and by coloring them by the # of observed sunspots at that time, we can clearly see that this model’s pre-training methods are implicitly learning such properties, which are good proxies for solar activity.