Reinforcement learning and its variants have emerged as some of the most successful artificial intelligence (AI) algorithms in recent years. For example, Google Deepmind’s AlphaGo, the first computer program to defeat a Go world champion, is largely based on reinforcement learning from games of self-play. However, the program had to be initially trained with a supervised dataset of human expert games on which reinforcement could build. More recently, AlphaGo Zero achieved a long-standing and much more challenging goal of AI by learning to play Go, tabula rasa, entirely relying on reinforcement learning with no human data or guidance. Biological learning systems that use reinforcement face a similar challenge. Is reinforcement learning a valid paradigm for motor skill acquisition very early in training?
We showed that midbrain dopamine neurons provide a performance prediction error signal in adult zebra finches. However, predictions are needed before the brain can generate a reinforcement learning signal based on prediction errors. Does the very young brain first need to learn the mapping between motor output and sensory feedback before a prediction can even be made?
