信号的语义刻画与度量  

Semantic Characterization and Measurement of Signals

作　　者：石光明 高大化 杨旻曦 谢雪梅 董明皓 李雷达 于凯 SHI Guang-ming;GAO Da-hua;YANG Min-xi;XIE Xue-mei;DONG Ming-hao;LI Lei-da;YU Kai(School of Artificial Intelligence,Xidian University,Xi’an,Shaanxi 710071,China;Peng Cheng Laboratory,Shenzhen,Guangdong 518055,China;School of Life Science and Technology,Xidian University,Xi’an,Shaanxi 710071,China)

机构地区：[1]西安电子科技大学人工智能学院,陕西西安710071 [2]鹏城实验室,广东深圳518055 [3]西安电子科技大学生命科学技术学院,陕西西安710071

出　　处：《电子学报》2022年第9期2068-2078,共11页Acta Electronica Sinica

基　　金：国家重点研发计划课题(No.2019YFA0706604);国家自然科学基金(No.61871304,No.61976169,No.61875157);鹏城实验室重大攻关项目(No.PCL2021A12)。

摘　　要：相比于基于比特数据的信息处理及通信技术,人类通过语义处理和传递信息的方式,在面对智能体间传递处理海量信息这一问题时显得更为高效和自然.然而由于目前缺乏关于语义度量和刻画的数学描述,涉及语义的应用无法兼顾可解释性和泛化性,无法发挥语义的高效自然的优势.本文围绕语义的度量和刻画,首先依据信息科学和神经科学相关结论,讨论了语义的内涵,并指出语义具有模块化、多模态、层级化的特点;接着提出了一种多模态信号的语义刻画和度量的数学描述;然后为了验证所提信号语义的刻画和度量的可行性和有效性,在MNIST(Mixed National Institute of Standards and Technology database)手写数字识别和水声目标识别两个应用中进行了实验,获得比传统深度学习更好的性能;最后将语义用于视频编码,实现了远超传统方法的压缩比,展现了语义在通信领域的实用价值.这为未来建立以语义为基础的新型信息处理与通信技术奠定了理论和实践基础.Compared with the modern information processing and communication technology based on bits,the se⁃mantic-based way of processing and transmitting information used by humans is more efficient and natural in the face of the massive information that needs to be sent between agents.However,due to the lack of mathematical description of semantic measurement and characterization,semantics'applications cannot consider both interpretability and generalization.There⁃fore,they cannot give play to the advantages of efficient and natural semantics.This paper focuses on the measurement and characterization of semantics.Firstly,we discuss the connotation of semantics based on the relevant conclusions of informa⁃tion science and neuroscience,and concludes that semantics has the characteristics of modularity,multi-mode,and hierar⁃chy.Then,a semantic description of multimodal signals and a mathematical description of their measurement are proposed.Next,to verify the feasibility and effectiveness of the characterization and measurement of the proposed signal semantics,experiments are carried out in two applications:MNIST(Mixed National Institute of Standards and Technology database)handwritten digital recognition and underwater acoustic target recognition,and the results are better than those of the tradi⁃tional deep learning.Finally,the semantics is used for video coding,and a compression ratio far exceeding that of tradition⁃al methods is achieved.This lays a theoretical and practical foundation for establishing new information processing and communication technology based on semantics.

关 键 词：语义 语义刻画 语义度量 语义基元 语义计算 语义识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]