Canonical computations for motor learning by the cerebellar cortex micro-circuit

摘要 小脑对于学习和执行协调、适时的动作至关重要。小脑 大脑皮层似乎在学习给动作计时方面扮演着特殊的角色。从上世纪六十年代的S和七十年代的S开始，我们有 了解小脑微电路的结构，但大多数对行为过程中小脑功能的分析 都专注于浦肯野细胞。在这里，我们建议在一个全新的水平上研究小脑皮质。 通过询问全小脑微电路--苔藓纤维、颗粒细胞、高尔基细胞、分子层 中间神经元和浦肯野细胞--在运动行为和运动学习过程中进行神经计算。 我们努力通过识别所有元素，记录它们在运动过程中的电活动，来“破解”电路 和学习，并重建再现生物数据的神经电路模型。我们将使用三个 已建立的学习系统，所有人都可以学习预测时间：眼皮反应的经典条件反射 (小鼠)，前肢运动的预测时间(小鼠)，以及平稳追踪眼的方向学习 运动(猴子)。我们的提案有六个主要特点。首先，光遗传学(在小鼠身上)会将放电联系起来 不同小脑中间神经元在运动和学习过程中对其分子细胞类型的了解。第二，一个 机器学习聚类分析(在老鼠和猴子中)将在细胞群体中找到相似之处 记录在我们的三种准备中，并将根据其假定的细胞类型对神经元进行分类 记录运动和运动学习过程中非浦肯野细胞的多项参数。第三，多元化-- 接触电极将允许我们同时从多个相邻的单个神经元和 计算尖峰计时交叉相关图(CCG)以识别连接的符号；我们还将寻找 CCG的变化提供了学习过程中特定可塑性部位的证据。第四，GCaMP成像 颗粒细胞层的变化将揭示小脑微回路输入的时间结构，以及 确定这些输入是否针对运动学习进行了修改。第五，模型神经网络，具有 真实的小脑结构将揭示一组模型参数，这些参数将改变被测量的 在我们的三个运动系统中，小脑对所有神经元的测量反应的输入 小脑皮层。第六，该模型将阐明突触和细胞可塑性在 小脑微回路中的不同部位共同作用，导致运动学习。

Abstract The cerebellum is critical for learning and executing coordinated, well-timed movements. The cerebellar cortex seems to have a particular role in learning to time movements. Since the 1960's and 70's, we have known the architecture of the cerebellar microcircuit, but most analyses of cerebellar function during behavior have focused on Purkinje cells. Here, we propose to investigate the cerebellar cortex at an entirely new level by asking how the full cerebellar microcircuit – mossy fiber, granule cells, Golgi cells, molecular layer interneurons, and Purkinje cells – performs neural computations during motor behavior and motor learning. We strive to “crack” the circuit by identifying all elements, recording their electrical activity during movement and learning, and reconstructing a neural circuit model that reproduces the biological data. We will use three established learning systems that all can learn predictive timing: classical conditioning of the eyelid response (mice), predictive timing of forelimb movements (mice), and direction learning in smooth pursuit eye movements (monkeys). Our proposal has six key features. First, optogenetics (in mice) will link the discharge of different cerebellar interneurons during movement and learning to their molecular cell types. Second, a machine-learning clustering analysis (in mice and monkeys) will find analogies among the cell populations recorded in our three preparations and will classify neurons according to their putative cell types based on recordings of many parameters of non-Purkinje cells during movement and motor learning. Third, multi- contact electrodes will allow us to record simultaneously from multiple neighboring single neurons and compute spike-timing cross-correlograms (CCGs) to identify the sign of connections; we also will look for changes in CCGs that provide evidence of specific sites of plasticity during learning. Fourth, gCAMP imaging of the granule cell layer will reveal the temporal structure of inputs to the cerebellar microcircuit, and determine whether those inputs are modified in relation to motor learning. Fifth, a model neural network with realistic cerebellar architecture will reveal a single set of model parameters that will transform the measured inputs to the cerebellum in our three movement systems to the measured responses of all neurons in the cerebellar cortex. Sixth, the model will elucidate how mechanisms of synaptic and cellular plasticity at different sites in the cerebellar microcircuit work together to cause motor learning.