Is this really just Hebbian learning? That's often stated as "cells that fire together wire together", and I can see how that can line up with FF increasing the 'goodness' of actual data. But decreasing the 'goodness' of negative data seems like a qualitatively different mechanism.