Cerebellar circuits for reward-based learning

点击翻译按钮获取中文摘要

ABSTRACT The cerebellum plays a key role in motor coordination and learning. Classic models posit that cerebellar learning is instructed by teaching signals from climbing fibers (CFs) that act according to the principles of supervised learning. While such models work well to describe CF activity and learning in some behaviors, they are not sufficient to explain CF activity in others. By developing an operant, reward-guided cerebellar-dependent task for the mouse, as well as a modified classical conditioning task, we used calcium imaging of CF input to Purkinje cell dendrites to demonstrate that CFs can be driven by reward-related task parameters. Our data suggested the possibility that CFs might engage in reinforcement learning to report predictions about expected rewards (reward prediction errors) in a similar manner as dopaminergic neurons of the ventral tegmental area (VTA). Importantly, however, our data also show significant differences from some predicitons of leading reinforcement learning models, and many other properties of cerebellar reward-based learning remain unclear. Thus, it remains largely unknown how the cerebellum operates in reward-based learning. Here will rigoursly test the hypothesis that CFs instruct cerbellar learning according to reinforcement learning rules: In the first aim, we will test whether CF activity obeys the many diverse requirements of reward prediction error signals, for example by scaling with both the probability and size of an expected reward. To do so, we will use two-photon calcium imaging to monitor CF input to Purkinje cell dendrites while manipulating reward contingencies during a classical conditioning paradigm. We will also determine the contribution of behavioral context, learning, and motor output to CF activity. In the second aim, we will test whether reward-predictive CF activity is generated by reward- responsive CF activity. This is a key property of reinforcement learning because it binds activity driven by an unconditioned simulus (US) to activity driven by a conditioned stimulus (CS). We will use classical blocking experiments and optogenetic manipulations to determine the neccessity and sufficiency of US-linked CF responses to generating CS-linked CF responses. Finally, in the third aim, we will determine whether CS-linked CF activity drives learned changes in behavior and cerebellar output. Thus, we will use a combination of optogenetics and extracellular electrophysioloigcal recordings test the function of reward-related CF activity. Together, these experiments will reveal the key principles that govern reward-based cerebellar learning, and how this learning alters cerebellar output and behavior.