基于Im2col的并行深度卷积神经网络优化算法

针对大数据环境下并行深度卷积神经网络（DCNN）算法中存在数据冗余特征多、卷积层运算速度慢、损失函数收敛性差等问题，提出了一种基于Im2col方法的并行深度卷积神经网络优化算法IA-PDCNNOA。首先，提出基于Marr-Hildreth算子的并行特征提取策略MHO-PFES，提取数据中的目标特征作为卷积神经网络的输入，有效避免了数据冗余特征多的问题；其次，设计基于Im2col方法的并行模型训练策略IM-PMTS，通过设计马氏距离中心值去除冗余卷积核，并结合MapReduce和Im2col方法并行训练模型，提高了卷积层运算速度；最后提出改进的小批量梯度下降策略IM-BGDS，排除异常节点的训练数据对批梯度的影响，解决了损失函数收敛性差的问题。实验结果表明，IA-PDCNNOA算法在大数据环境下进行深度卷积神经网络计算具有较好的性能表现，适用于大规模数据集的并行化深度卷积神经网络模型训练。

Parallel deep convolution neural network optimization based on Im2col

In the large data environment, there are many problems in the parallel deep convolution neural network(DCNN) algorithm, such as excessive data redundancy, slow convolution layer operation and poor convergence of loss function. This paper proposed a parallel deep convolution neural network optimization algorithm based on Im2col method(IA-PDCNNOA). Firstly, the algorithm proposed a parallel feature extraction strategy based on Marr-Hildreth operator to extract target features from data as input of convolution neural network, which could effectively avoid the problem of excessive data redundancy. Secondly, it designed a parallel model training strategy based on Im2col method, which removed the redundant convolution kernel by designing the Mahalanobis distance center value and improved the convolution layer operation speed by combining MapReduce and Im2col methods. Finally, it proposed an improved small-batch gradient descent strategy, which eliminated the effect of abnormal data on the batch gradient and solved the problem of poor convergence of the loss function. The experimental results show that IA-PDCNNOA algorithm performs well in deep convolution neural network calculation under large data environment and is suitable for parallel DCNN model training of large datasets.