Kavli Affiliate: Stefano Fusi
| Authors: W. Jeffrey Johnston and Stefano Fusi
| Summary:
The brain has a well-known large scale organization at the level of brain regions, where different regions have been argued to primarily serve different behavioral functions. This regional specialization has been explained theoretically as arising from constraints on connectivity and the physical geometry of the cortical sheet. In contrast, the structure of neural representations within a specific brain region is not well understood. Experimental and theoretical work has argued both for and against the existence of specialization at the level of single neurons, where single neurons might be specialized to encode only particular subsets of variables among those represented by the whole brain region. Here, we use artificial neural networks to study the emergence of this local structure and work to reconcile experimental results that show local structure in some cases but not in others. We show that specialization at the level of single neurons – that is, an explicitly modular representation – emerges to support context-dependent behavior, but only when the network starts from a specific representational geometry. We also outline cases in which either unstructured or population-level specialization (i.e., an implicitly modular representation) is learned. We show that both explicitly and implicitly modular solutions are abstract and allow for rapid learning and generalization on novel tasks as well as zero-shot transfer to novel stimuli. We further show that modular geometries facilitate the rapid learning of novel contexts that are related to previously seen contexts, while less structured geometries facilitate the rapid learning of novel unrelated contexts. Together, our findings clarify multiple conflicting experimental results. Further, they make numerous predictions for future experimental work and highlight the important joint roles of task structure and initial representational geometry in shaping learned representations.