支持OpenCL的GPU加速人工神经网络训练

摘要

The computation quantity in artificial neural network training will get more and more with the increase of neurons quantity, it is time-consuming for training a neural network with too many neurons.A method that accelerates artificial neural network training is to optimize the training algorithm, so as to reduce the computation quantity.Since there is too much matrix and vector computation in artificial neural network training algorithm, the optimized training algorithm implemented by OpenCL C language on GPU, compared to the conventional CPU-based implementation, the training speed will be increased a lot.Based on parallel computing ability of hardware, accelerating of artificial neural network training by GPU with OpenCL Support is researched in this paper.%人工神经网络训练所包含的运算量随着网络中神经元的数量增多而加大,对于神经元较多的网络训练很耗时.提高人工神经网络训练速度的一个方法是对训练算法优化以减少计算量.由于人工神经网络训练算法包含大量的矩阵和向量运算,如果把优化的算法用运行在GPU上的OpenCL C语言实现,则训练速度相比传统基于CPU计算的实现会提高很多.从硬件的并行计算能力着手,以RPROP算法为例,对其运行在GPU上的OpenCL C语言实现作一些研究.