Fujita Masahiko
Brain Science Ciel Laboratory, Kodaira, Tokyo 187-0021, Japan.
Neural Netw. 2016 Mar;75:173-96. doi: 10.1016/j.neunet.2015.12.012. Epub 2016 Jan 6.
Lesions of the cerebellum result in large errors in movements. The cerebellum adaptively controls the strength and timing of motor command signals depending on the internal and external environments of movements. The present theory describes how the cerebellar cortex can control signals for accurate and timed movements. A model network of the cerebellar Golgi and granule cells is shown to be equivalent to a multiple-input (from mossy fibers) hierarchical neural network with a single hidden layer of threshold units (granule cells) that receive a common recurrent inhibition (from a Golgi cell). The weighted sum of the hidden unit signals (Purkinje cell output) is theoretically analyzed regarding the capability of the network to perform two types of universal function approximation. The hidden units begin firing as the excitatory inputs exceed the recurrent inhibition. This simple threshold feature leads to the first approximation theory, and the network final output can be any continuous function of the multiple inputs. When the input is constant, this output becomes stationary. However, when the recurrent unit activity is triggered to decrease or the recurrent inhibition is triggered to increase through a certain mechanism (metabotropic modulation or extrasynaptic spillover), the network can generate any continuous signals for a prolonged period of change in the activity of recurrent signals, as the second approximation theory shows. By incorporating the cerebellar capability of two such types of approximations to a motor system, in which learning proceeds through repeated movement trials with accompanying corrections, accurate and timed responses for reaching the target can be adaptively acquired. Simple models of motor control can solve the motor error vs. sensory error problem, as well as the structural aspects of credit (or error) assignment problem. Two physiological experiments are proposed for examining the delay and trace conditioning of eyelid responses, as well as saccade adaptation, to investigate this novel idea of cerebellar processing.
小脑损伤会导致运动出现较大误差。小脑根据运动的内部和外部环境，自适应地控制运动指令信号的强度和时间。目前的理论描述了小脑皮质如何控制信号以实现精确且定时的运动。小脑高尔基细胞和颗粒细胞的模型网络被证明等同于一个多输入（来自苔藓纤维）的分层神经网络，该网络具有一个由阈值单元（颗粒细胞）组成的单一隐藏层，这些阈值单元接受共同的递归抑制（来自高尔基细胞）。关于该网络执行两种通用函数逼近的能力，对隐藏单元信号（浦肯野细胞输出）的加权和进行了理论分析。当兴奋性输入超过递归抑制时，隐藏单元开始放电。这种简单的阈值特征引出了第一种逼近理论，并且网络的最终输出可以是多个输入的任何连续函数。当输入恒定时，该输出会变得稳定。然而，当通过某种机制（代谢型调制或突触外溢出）触发递归单元活动减少或递归抑制增加时，如第二种逼近理论所示，网络可以在递归信号活动的长时间变化中生成任何连续信号。通过将小脑的这两种逼近能力纳入一个运动系统，在该系统中学习通过伴随校正的重复运动试验进行，可以自适应地获得到达目标的精确且定时的反应。简单的运动控制模型可以解决运动误差与感觉误差问题，以及信用（或误差）分配问题的结构方面。提出了两项生理实验，用于检查眼睑反应的延迟和痕迹条件反射以及扫视适应，以研究这种关于小脑处理的新观点。
