Credit Assignment Problem

The challenge in training DNNs to figure out how much each weight in the hidden layers contributed to the overall error in the output, in order to adjust them correctly. Also called "blame assignment problem".