I was thinking about the semantics thing and as I often do - I looked at how the brain does it. The map to map connections includes a significant number of connection that skip a level. See figure 2 in this paper and note the map “skipping” in semantic learning.
One possibility is: I have been leaning towards the “top boss” idea that this upper level is a boss that learns quickly with the long loops first. As the “middle” layer learns the same input (with counter-flowing feedback from the boss) it adds its influence to the lower resolution input and the two signals together form a new input to the boss that may change it to a more nuanced interpretation of the input.
Another idea that was offered is: So perhaps the middle layer would provide some workspace to ease moving a frequently encountered concept down the hierarchy. The more informed higher level guy can provide smarter feedback based on having more information, then the middle guy would start to get the hang of it and replace the higher guy when he becomes more certain.