文档介绍：计算机工程
Computer Engineering
ISSN 1000-3428,CN 31-1289/TP

结构，有效地弥补深度鸿沟，
并通过运用贪心指标作为选边准则，增加搜索评估的相关性并提高搜索的稳定性。同时针对网络结构搜索算法消耗计算资源
多的问题，提出了渐进式划分数据集方法，通过分阶段不同比例的划分数据集来减少结构搜索的计算资源消耗，能够在短时
间内设计出性能优越且稳定的网络结构，在 1080Ti 上仅需要个 GPU Days。实验结果证明，以准确率和搜索时间作为指
标，快速渐进式搜索算法搜索出的网络结构搜索相关性高，稳定性得到改进，搜索时间减少，在 CIFAR-10 数据集上的最高
精度达到 %。
关键词：深度学习；卷积神经网络；可微结构搜索；渐进式结构搜索；划分数据集

开放科学（资源服务）标志码（OSID）：

A Fast and Progressive Convolution Neural Network Architecture
Search Algorithm
ZHAO Liang, FANG Wei
(School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, Jiangsu 214122, China)
【Abstract】Convolutional neural network has played a great role in the field of deep learning, but it is very difficult to design
network manually which requires high professionalism. Neural architecture search has become the research focus of deep learning
with its advantages of high efficiency and intelligence. But the search algorithms generally have