I would have thought the "few parameter" high-error region corresponds more naturally to the part of language learning where the learner overgeneralizes and thinks everything is regular, and the "many parameter" high-error region corresponds to knowing the irregular forms for each word that you've encountered. But this blog seems to think of it in the other way around. Maybe I'm missing something.
It can be argued that the human case correlates better with the "data double descent" model, assuming the brain network size does not change. The thing is... it does! It goes up and down as connections grow and get pruned. Biology is messy. Biological analogies can never be perfectly clean, especially for a system as giant as a human.
