Learning with reward prediction errors in a model of the Drosophila mushroom body

Abstract

Effective decision making in a changing environment demands that accurate predictions are learned about decision outcomes. In Drosophila, such learning is or-chestrated in part by the mushroom body (MB), where dopamine neurons (DANs) signal reinforcing stimuli to modulate plasticity presynaptic to MB output neurons (MBONs). Here, we extend previous MB models, in which DANs signal absolute rewards, proposing instead that DANs signal reward prediction errors (RPEs) by utilising feedback reward predictions from MBONs. We formulate plasticity rules that minimise RPEs, and use simulations to verify that MBONs learn accurate reward predictions. We postulate as yet unobserved connectivity, which not only overcomes limitations in the experimentally constrained model, but also explains additional experimental observations that connect MB physiology to learning. The original, experimentally constrained model and the augmented model capture a broad range of established fly behaviours, and together make five predictions that can be tested using established experimental methods.