Maxout网络的剪枝
首发时间:2015-11-23
摘要:近年来,Maxout网络在计算机视觉和语音识别等领域均取得了很好的成果。与Dropout方法类似,Maxout网络通过对神经网络每层单元进行分组并选取每组最大输出,达到了合并子网络的效果。但由于其每层单元数量较多,模型参数量较大,导致网络训练缓慢,某些单元训练不充分,且参数量增加容易导致过拟合。本文针对Maxout网络存在冗余单元这一现象进行了分析,并提出了对应的剪枝方法。实验表明,该剪枝方法在保持原Maxout网络的识别效果的基础上,减少网络参数50%~60%,且剪枝后的模型效果优于与其参数量相当的未剪枝的Maxout网络模型,说明对复杂模型进行剪枝既能压缩参数量至合理的范围,其效果也好于随机初始化的规模较小的模型训练后的效果。
For information in English, please click here
Pruning Maxout Network
Abstract:Recently the maxout network is widely applied in computer vison and speech recognition and proved to be a state of the art method. Like Dropout, a maxout unit takes the max output of its candidates, so the maxout network aggregates actually a set of its sub-network. However, the large number of units in maxout layers leads to the increase of training time, and certains units aren't well-trained. Besides, the increase of the parameters could cause the over-fitting problems. In this paper we analyze the redundancy of the Maxout network and propose a method to prune the redundant units. The experiments prove that the pruned model can keep the performance of the original model, while reducing the number of parameters about 50% to 60% . In addition, the pruned model performs better than the model which has the same numbers of parameters as the pruned model, which means that pruning a complex model can not only reduce the storage space, but also give a good initialization of parameters for smaller models.
Keywords: convolutional neural network maxout network pruning method
基金:
论文图表:
引用
No.4658959110188914****
同行评议
共计0人参与
勘误表
Maxout网络的剪枝
评论
全部评论0/1000