Neural circuits for visual stimulus competition in zebrafish

摘要

The need for reinforcement signals other than dopamine Christopher D. Fiorillo KAIST, Daejeon, Republic of Korea The brain must learn to approach reward and avoid punish- ment. Computational and physiological models explain how this could occur through reinforcement signals that indicate good or bad and cause Hebbian reinforcement learning (RL). These mod- els are supported by overwhelming experimental evidence that midbrain dopamine neurons signal a ‘reward prediction error’ (RPE) that causes the positive reinforcement of stimuli and actions. There is far less experimental evidence concerning the relation of the dopamine RPE to learning aversiveness, and absence of reward. In standard RL models, the dopamine RPE is sufficient for all reinforcement. These models assume that one dopamine RPE signals ‘total value,’ in which reward and aversiveness are merely opposites on a single dimension of value (analogous to light and dark on the single dimension of light intensity). In con- trast, I will present the theoretical rationale for why reward and aversiveness should be two distinct dimensions of value, and why the learning of reward and aversive value calls for four discrete types of reinforcement signals, representing evidence for reward, against reward, for aversiveness, and against aversiveness. Care- fully controlled experiments indicate that the well characterized dopamine RPE signals only evidence for reward, and therefore plays amore restricted role in RL than commonly believed. Greater effort is needed to characterize the other three types of reinforcement signals.