摘要
当前,基于深度学习的目标识别建模技术面临标注样本不足、模型可解释性不高、稳定性不够等新的挑战,限制了深度学习解决更复杂、更抽象问题的可能性。采用知识与数据联合驱动的方式进行智能模型构建是突破现有瓶颈的一条重要途径。该文以外部经验与认知知识在模型构建中的引入方式为区分准则,提出了模型构建方法的分类标准,包括基于显式知识的建模方法、基于隐式知识的建模方法以及基于融合知识的建模方法;然后围绕每类方法在解决小样本、模型可解释性等问题上的探索进行综述,并总结设想了一种未来的知识与数据联合驱动建模方式。这种方式吸取了不同建模方式的优点,通过解耦知识建模与数据建模,以无监督、弱监督为核心训练方式,可以有效解决小样本条件下模型构建问题,提高模型可解释性。最后,该文总结了需要进一步研究的问题和未来的研究方向,以促进目标识别模型构建技术的发展。
In recent years,object recognition modeling techniques based on deep learning face new challenges such as insufficient annotated samples,low interpretability of models,and insufficient stability.All these challenges limit the possibility of deep learning to solve more complex and abstract problems.Constructing intelligent model jointly driven by knowledge and data is an important way to break through the existing bottleneck.This paper presents a classification standard of model constructing methods according to the way of introduction of external experience and cognitive knowledge during model constructing,including modeling methods based on explicit knowledge,modeling methods based on implicit knowledge and modeling methods based on fusion knowledge.Then,following the proposed classification standard,the explorations in each class of methods about solving the problems of few samples and model interpretability are reviewed.Subsequently,taking advantage of different model constructing methods,a future model constructing method jointly driven by knowledge and data is proposed.The proposed method can effectively solve the problem of model construction under the condition of few samples and improve the interpretability of the model by decoupling knowledge modeling and data modeling and taking unsupervised and weak supervised training as the core training patterns.Finally,some research issues which need further study as well as future research directions are drawn in conclusion for promote the object recognition model constructing.
作者
田晟兆
胡迎茜
谷成
陈端兵
TIAN Shengzhao;HU Yingqian;GU Cheng;CHEN Duanbing(Big Data Research Center,University of Electronic Science and Technology of China,Chengdu 611731;Second Academy of China Aerospace Science and Industry Corporation,Haidian Beijing 100854;Aerospace Science&Industry Defense Technology Research and Test Center,Haidian Beijing 100854;Chengdu Union Big Data Technology Incorporated,Chengdu 610041)
出处
《电子科技大学学报》
EI
CAS
CSCD
北大核心
2023年第6期932-943,共12页
Journal of University of Electronic Science and Technology of China
基金
国家自然科学基金重点项目(T2293771)。
关键词
人工智能
深度学习
知识与数据联合驱动
模型构建
目标识别
artificial intelligence
deep learning
knowledge and data driven
model constructing
object recognition
作者简介
田晟兆(1994-),男,博士生,主要从事人工智能、深度学习、目标识别等方面的研究;通信作者:陈端兵,E-mail:dbchen@uestc.edu.cn。
