基于拟神经突触的概率计算神经网络加速器

Neural network accelerator based on probability calculation of pseudo neural synapses

深度卷积神经网络是人工智能应用中最有前途的神经网络类别之一。但是与其性能相关的计算量和计算复杂度也呈现爆炸式增长,导致了实现的高功耗和巨大的硬件成本。因此,研究基于概率计算空间编码的思想,使用电路中信号线密度模拟神经突触可塑性的工作原理,将传统乘法器复杂的运算变成了简单线与的低复杂度方式,实现了一种拟神经突触概率乘法器。并以此为基础,实现了基于多段分解的空间并行高精度概率乘法器,并提出了一种以突触可塑性计算单元为基础的低复杂度、低开销的神经网络加速器。实验使用该概率乘法器基于LeNet-5网络在MNIST上和基于VGG13的CIFAR-10上分别达到了97.99%和85.21%的正确率,与传统乘法器相比,降低了面积和功耗,提升了运行速度。同比其他神经网络加速器,提高了吞吐量、面积效率、功率效率。

Deep convolution neural network is one of the most promising neural networks in the application of artificial intelligence.However,the amount and complexity of computing related to its performance also show an explosive growth,resulting in high power consumption and huge hardware costs.Therefore,based on the idea of probability calculation space coding,the research uses the signal line density in the circuit to simulate the working principle of neural synaptic plasticity,and turns the complex operation of the traditional multiplier into a low complexity mode of simple lines and realizes a pseudo neural synaptic probabilistic multiplier.On this basis,a high precision spatial parallel probability multiplier based on multi segment decomposition is implemented,and a low complexity and low overhead neural network accelerator based on synaptic plasticity computing cell is proposed.In the experiment,the probability multiplier based on LeNet 5 network achieves 97.99%accuracy on MNIST and 85.21%accuracy on CIFAR 10 based on VGG13,respectively.Compared with traditional multiplier implementation,it reduces area and power consumption,and improves operating speed.Compared to other neural network accelerators,it has increased throughput,area efficiency,and power efficiency.