【摘要】 神经网络架构搜索主要解决人工设计神经网络难度大的问题，针对该算法的研究在自动化机器学习领域有着深远的意义。搜索算法的主要流程包括设计搜索空间、设计搜索策略、网络评估，针对搜索空间，应用通道随机重排技术和上下文信息融合技术进行高效特征提取，同时在搜索策略上，针对联合搜索优化困难且消耗时间长的问题，设计了修正网络的reduction cell，仅搜索normal cell。结果表明，该网络的搜索效率得到了提高搜索时间减少了31%左右，并且在cifar10数据集和tiny-imagenet数据集上的验证精度分别提高了0.21%和2.26%。更多还原

【Abstract】 Neural architecture search mainly solves the difficult problem of designing neural networks by human. Research on this algorithm has far-reaching significance in the field of automated machine learning. The process of search algorithm includes designing search space,designing search strategy,and network evaluation.For search space,channel shuffle and global context fusion technology are used for efficient feature extraction. Meanwhile,to solve the problem that joint optimization is difficult and takes a long time,this paper designs a reduction cell to modify the network. Results shows that the search efficiency of the network is improved and the search time is reduced by about 31%. The verification accuracy on the cifar10 and tiny-imagenet dataset has been improved by 0.21% and 2.26% respectively.更多还原