Learning with naturalistic odor representations in a dynamic model of the Drosophila olfactory system

Abstract Many odor receptors in the insect olfactory system are broadly tuned, yet insects can form associative memories that are odor-specific. The key site of associative olfactory learning in insects, the mushroom body, contains a population of Kenyon Cells (KCs) that form sparse representations of odor identity and enable associative learning of odors by mushroom body output neurons (MBONs). This architecture is well suited to odor-specific associative learning if KC responses to odors are uncorrelated with each other, however it is unclear whether this hold for actual KC representations of natural odors. We introduce a dynamic model of the Drosophila olfactory system that predicts the responses of KCs to a panel of 110 natural and monomolecular odors, and examine the generalization properties of associative learning in model MBONs. While model KC representations of odors are often quite correlated, we identify mechanisms by which odor-specific associative learning is still possible.