NEURAL BASIS OF PERFORMANCE EVALUATION IN BIRDS

Abstract

Reinforcement learning (RL) is one of the most successful theories in the intersection between animal behavior, neural basis of learning, and artificial intelligence. RL-like computation in the brain is evidenced by representation of a key variable in the dopaminergic system, the reward prediction error. Yet most studies of the dopaminergic system have been in situations where an animal is motivated by external rewards, such as food or juice. Does RL apply to learning of complex behavior without immediate external reward? What are the neural circuits responsible for learning a motor sequence such as birdsong? Here I introduce work in the songbird zebra finch demonstrating RL applies to performance learning based on a flexible internal benchmark. Dopaminergic neurons in the midbrain were shown to report a performance prediction error. I expand the song system involved in song learning to neural circuits responsible for RL. I found ventral pallidum (VP), a major input area to the dopaminergic midbrain, to be involved in song learning. I discovered a rich representation in VP including performance prediction related activity that is signaled to the dopaminergic midbrain. I contextualize my findings in the songbird system and hypothesize a consolidator-actor-critic architecture for song learning. I discovered neural representations in VP and the dopaminergic midbrain that switch between body movement and song timing depending on the animal’s behavioral states of singing and non-singing. I also report the first neural recordings from a parrot. Surprisingly, representations of both vocalization and body movements were found in the parrot primary vocal motor cortex. This finding sets up the foundation for future neurophysiological studies of social behavior in the parrot.