Knowledge partitioning in humans and neural networks

Natural agents are localised in space and time, so that states, values and goals are encountered in sequence. A major challenge for representation learning is to partition new knowledge in ways that avoid mutual interference and allow for flexible reuse. In humans, slowly changing sensory signals may give contextual cues that facilitate this partitioning process. I describe a theory of how knowledge partitioning mitigates catastrophic interference, allows for rapid retrospective re-evaluation of knowledge structures (transitive hierarchies), and accounts for both successes and failures of human compositional generalisation. The theory is cast in terms of a connectionist model, and discussed with reference to the study of human behaviour, geometries of neural representation measured with neuroimaging methods, and analysis of single cell data from NHPs.