Neural architecture search (NAS) is a hot topic in the field of automated machine learning (AutoML) and has begun to outperform human-designed architectures on many machine learning tasks. Motivated by the natural representation form of neural networks by the Cartesian genetic programming (CGP), we propose an evolutionary approach of NAS based on CGP, called CGPNAS, to solve sentence classification task. To evolve the architectures under the framework of CGP, the existing key operations are identified as the types of function nodes of CGP, and the evolutionary operations are designed based on Evolutionary Strategy (ES). The experimental results show that the searched architecture can reach the accuracy of human-designed architectures, such as Transformer. The transfer study proves that the searched architectures have the certain ability for dataset transfer. The ablation study identifies the Attention function as the single key function node. In addition, only through the linear transformations, the accuracy of the searched architectures is reduced by 4%, worthy of investigation in the future.
翻译:神经结构搜索(NAS)是自动化机器学习(AutomL)领域的一个热题,并已开始在许多机器学习任务上超越人设计的建筑。受卡尔提西亚基因编程(CGP)自然代表神经网络的驱动,我们提议以CGP(称为CGPNAS)为基础,采用NAS渐进式方法解决判决分类任务。为了在CGP框架内发展结构,现有关键操作被确定为CGP的功能节点类型,而进化操作是根据进化战略设计的。实验结果显示,搜索结构可以达到人类设计的结构的准确性,例如变异器。转移研究证明,搜索结构具有一定的数据集传输能力。缩略图研究将注意功能确定为单一关键函数节点。此外,只有通过线性转换,搜索结构的准确性才能降低4%,值得今后调查。