Abstract
A central problem to understanding intelligence is the concept ofgeneralisation. This allows previously learnt structure to be exploited tosolve tasks in novel situations differing in their particularities. We takeinspiration from neuroscience, specifically the Hippocampal-Entorhinal system(containing place and grid cells), known to be important for generalisation. Wepropose that to generalise structural knowledge, the representations of thestructure of the world, i.e. how entities in the world relate to each other,need to be separated from representations of the entities themselves. We show,under these principles, artificial neural networks embedded with hierarchy andfast Hebbian memory, can learn the statistics of memories, generalisestructural knowledge, and also exhibit neuronal representations mirroring thosefound in the brain. We experimentally support model assumptions, showing apreserved relationship between grid and place cells across environments.