decomposing knowledge into interchangeable pieces promises a generalization
advantage when there are changes in distribution. A learning agent interacting
with its environment is likely to be faced with situations requiring novel
combinations of existing pieces of knowledge. We hypothe