The traditional clock-driven algorithm is very time-consuming when performed on large-scale neuronal networks due to the huge number of synaptic currents computation and low performance of the parallel implementation of the algorithm. We find in this paper that the conductance coefficients of all the synapses coming from the same presynaptic neuron (neuron i for example) does not need to be computed one by one, rather only one common conductance coefficient needs to be computed for all synapses from this neuron. We then propose an idea of virtual synapse for neuron i to compute this common conductance coefficient and thereby have N ( N is the number of neurons in the network) virtual synapses for all presynaptic neurons in the network. Since each common conductance depends on only the spiking activity of the presynaptic neuron i and is irrelevant of postsynaptic neurons, the computation of the different virtual synapses can be deployed to different computer processing unit efficiently. By introducing a circular data structure for the virtual synapses, we present a novel parallel clock-driven algorithm based on graphics processors for simulation of neuronal networks. It is demonstrated by test results that the proposed algorithm reduces memory and time consumption greatly, and improves the performance of the parallelization for large-scale neuronal network simulations effectively.

中文翻译：

---

一种基于虚拟突触的神经网络仿真并行时钟驱动算法

传统的时钟驱动算法在大规模神经元网络上执行时，由于突触电流计算量巨大，算法并行实现的性能低下，非常耗时。我们在本文中发现，来自同一个突触前神经元（例如神经元 i）的所有突触的电导系数不需要一一计算，而只需要为所有突触计算一个公共电导系数。神经元。然后我们提出了神经元 i 的虚拟突触的想法来计算这个公共电导系数，从而为网络中的所有突触前神经元拥有 N 个（N 是网络中的神经元数量）虚拟突触。由于每个公共电导仅依赖于突触前神经元 i 的尖峰活动，与突触后神经元无关，因此不同虚拟突触的计算可以有效地部署到不同的计算机处理单元。通过为虚拟突触引入循环数据结构，我们提出了一种基于图形处理器的新型并行时钟驱动算法，用于模拟神经元网络。测试结果表明，该算法大大减少了内存和时间消耗，有效提高了大规模神经网络模拟的并行化性能。通过为虚拟突触引入循环数据结构，我们提出了一种基于图形处理器的新型并行时钟驱动算法，用于模拟神经元网络。测试结果表明，该算法大大减少了内存和时间消耗，有效提高了大规模神经网络模拟的并行化性能。通过为虚拟突触引入循环数据结构，我们提出了一种基于图形处理器的新型并行时钟驱动算法，用于模拟神经元网络。测试结果表明，该算法大大减少了内存和时间消耗，有效提高了大规模神经网络模拟的并行化性能。