Testing an Explicit Method for Multi-compartment Neuron Model Simulation on a GPU
Testing an Explicit Method for Multi-compartment Neuron Model Simulation on a GPU
复制标题
DOI：
10.1007/s12559-021-09942-6
复制
发表时间：
2021-11-09
影响因子：
5.4
通讯作者：
Yamazaki，Tadashi
中科院分区：
文献类型：
--
作者：
Kobayashi，Taira;Kuriyama，Rin;Yamazaki，Tadashi
多室模型的大规模模拟对于理解单个神经元的形态结构在大脑中信息处理中的作用是重要的。在模拟中，描述神经元动力学的偏微分方程组(PDE)必须针对每个时间步长进行数值求解。在求解偏微分方程组时，为了保证稳定性，采用了称为隐式方法的数值方法。隐式方法需要求解联立方程组，这会使并行计算的图形处理单元(GPU)硬件加速器上的数值模拟速度变慢。为了克服这个问题，我们研究了使用显式方法对多室模型进行模拟。我们将Runge-Kutta-Chebyshev(RKC)方法应用于几种小脑神经元模型，包括浦肯野细胞、颗粒细胞、高尔基细胞和下橄榄细胞。接下来，我们实现了一个由颗粒细胞、高尔基细胞和浦肯野细胞组成的小脑皮质模型，同时对不同的细胞类型使用了不同的数值方法。尽管显式方法对偏微分方程组可能不稳定，但在大多数情况下，使用RKC方法表现出足够的稳定性，在GPU上比隐式方法具有更好的计算性能，并且具有良好的重复性。在网络模拟中，为每种单元类型选择合适的数值方法比单独使用隐式方法模拟速度更快。我们的结果表明，显式方法适用于多室模型，并可以加快模拟的计算速度。此外，要对多隔室模型进行大规模模拟，选择有效的数值方法将更加重要。
Large-scale simulation of multi-compartment models is important for understanding the role of morphological structures of individual neurons for information processing in the brain. In a simulation, partial differential equations (PDEs) that describe the dynamics of neurons have to be solved numerically for each time step. To solve PDEs, numerical methods called implicit methods are used for stability. Implicit methods need to solve simultaneous equations, which can make numerical simulation slow on graphics processing units (GPUs) hardware accelerators for parallel computing. To overcome this problem, we investigated the use of explicit methods for multi-compartment model simulation. We applied the Runge–Kutta–Chebyshev (RKC) method to several cerebellar neuron models including Purkinje cells, granule cells, Golgi cells, and inferior olive cells. Next, we implemented a cerebellar cortical model composed of granule cells, Golgi cells, and Purkinje cells, while using different numerical methods for different cell types. Although explicit methods can be unstable against PDEs, using the RKC method showed sufficient stability for most cases, better computational performance than implicit methods on a GPU, and good reproducibility. In the network simulation, choosing the suitable numerical methods for each cell type achieved faster simulation than that used an implicit method solely. Our results suggest that explicit methods are applicable to multi-compartment models and can accelerate computational speed of simulations. Furthermore, to conduct large-scale simulation of multi-compartment models, choosing efficient numerical methods will be more important.