摘要
为了改善传统标准dropout无法根据特定数据集确定合适尺度的不足,提出了多尺度融合dropout（MSF dropout）方法.利用验证数据集对多个不同尺度的网络模型进行训练,通过学习得到符合该数据集特征的最佳尺度组合,让MSFdropout具备自适应数据集的能力,从而使网络能够使用最佳尺度来进行高精确度的预测.首先训练若干组不同尺度的网络模型,使用遗传算法求出各网络模型的最优尺度;然后通过最优尺度对对应的网络参数进行缩小得到预测子模型;最后以一定的权重将这些子模型融合成为最终的预测模型.使用MSFdropout在标准数据集MNIST和CIFAR-10中进行实验,实验表明：当选择了合适的尺度数量和尺度梯度后,预测精度获得了明显的提升,同时很好地控制了计算时间,验证了多尺度融合方法的有效性.
In order to improve the shortage of the traditional standard dropout which cannot find the appropriate scale according to the specific dataset,a novel approach called multi-scale fusion dropout（MSF dropout） was proposed.The validation dataset was used to train multiple network models with different scales so as to acquire the ability to adapt datasets by learning to get the best scale combination that meets the characteristics of the dataset.Therefore,the network can perform high-precision prediction with the optimal scale.A number of network models were first trained with different scales and genetic algorithm was used to find the optimal scale of each network model.Then the prediction sub-models was obtained by scaling down the corresponding network coefficients with the optimal scale,which was fused into a final prediction model with a certain weight.Experiments on benchmark image classification datasets MNIST and CIFAR-10 show when the appropriate scale quantity and scale gradient are selected,not only has the predicting accuracy been significantly improved,but also controls the calculation time well,which prove the effectiveness of the multi-scale fusion dropout.
作者
钟忺
陈恩晓
罗瑞奇
卢炎生
Zhong Xian;Chen Enxiao;Luo Ruiqi;Lu Yansheng(School of Computer Science and Technology,Wuhan University of Technology,Wuhan 430072,China;School of Computer Science and Technology,Huazhong University of Science and Technology,Wuhan 430074,China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2018年第9期35-39,共5页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金资助项目(61003130)
国家科技支撑计划资助项目(2012BAH33F03)
湖北省自然科学基金资助项目(2015CFB525)
