基于空间相关各向异性规则的仿真小脑模型的神经网络优化研究

Neural Network Optimization of Simulated Cerebellar Model Based on Spatial Correlation Anisotropy Rule

摘要: 仿真小脑模型的颗粒层神经元通过随机投射网络生成时空活动序列，然而网络中的空间相关性和连通性未能充分反映真实生物组织的结构原则。本文针对该问题提出一种基于新的连通规则的优化方案，借助Perlin噪声的纹理特性让神经元的连接具有优先方向偏好和相邻神经元有相似的偏好，从而模拟生物神经元连接在空间上的不对称性；此外，本文还借助高斯分布设定神经元之间的连接强度，使其随着神经元之间距离的增加而呈指数衰减，这与生物学中兴奋性突触后电流随突触距离神经细胞体中心越远而减弱的现象相吻合为了验证改进的有效性，在相同条件下通过对仿真视动反射任务的对比分析，研究了优化前后的小脑模型在记忆学习能力上的差异，并对两种模型下细胞电生理变化进行了深入探讨。该优化的小脑模型不仅对小脑运动学习记忆形成有着理论意义，而且为小脑控制器提供工程上的应用价值。

Abstract: The granular layer neurons of the simulated cerebellar model generate spatiotemporal activity sequences through a random projection network. However, the spatial correlation and connectivity in the network do not fully reflect the structural principles of real biological tissues. Aiming at this problem, this paper proposes an optimization scheme based on new connectivity rules. By using the texture characteristics of Perlin noise, the connection of neurons has the preference of preferred direction and the similar preference of adjacent neurons, so as to simulate the asymmetry of biological neuron connection in space; In addition, we also use the Gaussian distribution to set the connection strength between neurons, so that it decays exponentially with the increase of the distance between neurons, which is consistent with the phenomenon in biology that the excitatory postsynaptic current decreases with the distance from the synapse to the center of the nerve cell body. In order to verify the effectiveness of the improvement, through the comparative analysis of the simulated optokinetic reflex task under the same conditions, the differences in memory and learning ability of the cerebellar model before and after optimization were studied, and the changes of cell electrophysiology under the two models were discussed. The optimized cerebellar model not only has theoretical significance for the formation of cerebellar motor learning and memory, but also provides engineering application value for cerebellar controller.