一种动态构建深度信念网络模型方法

A dynamic method of constructing deep belief network models

针对深度信念网络无法科学有效地确定网络模型深度和隐层神经元数目等问题,根据贪心算法思想,提出了一种动态构建深度信念网络模型的新方法.即从底层逐层构建深度信念网络的过程中,根据验证集错误分类率调整当前层神经元数目,使当前模型达到最优后,固定当前层神经数目,网络深度增加一层;继续调整下一层神经元数目,直至整个模型构建完成.最后,根据重构误差微调各层神经元数目.结果表明,与依据重构误差构建的深度信念模型相比,利用此方法构建的深度信念网络模型的分类准确率更高.

In view of the fact that deep belief networks could not determine the depth of the network and the number of hidden units, a dynamic method of constructing deep belief networks was proposed according to the idea of the greedy algorithm. In the process of constructing deep belief networks from the bottom layer, the number of hidden units in the current layer was adjusted according to the error classification rate of the v erification set. After the current model was optimized, the number of hidden units of the current layer was fixed, and the depth of the network was increased by one layer. The number of hidden units in the next layer was adjusted as the same ways until the whole model was completed. Finally, the number of hidden units in each layer was tuned according to the reconstruction error. The experimental results show that the model constructed according to this method is more accurate than the model constructed based on the reconstruction error.