Model-driven semantic communication
-
Information source side
① AI model is used to extract semantic information from the original input information, and then send the sum提取的语义信息
and information together . ②In addition, factors such as physical environment, spectral environment, and electromagnetic environment are taken into consideration and extracted and integrated to facilitate semantic information transmission. ③ After the above processing, the original input information is converted into information generated based on the AI model, which is the fusion of , , and , and is transmitted through the communication system based on SHANNON .用于提取语义信息的AI模型
环境联合设计模块
AI模型
环境信息
提取的语义信息
比特级
-
At the sink end
①Since background knowledge such as context and environment will affect the restoration of semantic information,环境表示模块
the model will be used to eliminate the influence of background knowledge to facilitate the restoration of extracted semantic information.
②In addition, information obtained from环境表示模块
and will be sent to . ③In the semantic restoration and decoding module, the received semantic information will be restored and decoded in a way using the restoration and decoding model, which plays a role by combining with .模型
语义恢复和解码模块
模糊
AI模型
Model-based semantic communication will
用于语义信息处理的 AI 模型
also be transmitted over physical channels instead of relying on pre-trained decoders as in Ref. [1,2]. For example, the receiver in a proposal system can fuse models from the transmitter and other nodes according to quality of service (QoS) requirements.
In the intelligent communication system, the analysis, encoding, decoding and recombination of information at the semantic level are all driven by the model, and the sending and receiving end can realize the sharing of knowledge base and the flow of wisdom at the sending and receiving end by transferring the model. If the receiving end does not have a matching semantic decoding and recovery model, the receiving end can request the sending end or the network node to transmit the latest intelligent model. In addition, the intelligent-driven communication system compresses the model, or only sends part of the model, that is,model groupingetc. to further reduce the amount of data transmission.
Reference: Research Report on Key Technologies of Intelligent-Driven Communications
Metrics of Intent-Driven Communication System
-
Semantic communication metrics
①text object,
bilingual evaluation understanding (BLEU), etc.
②image or video object,
peak signal-to-noise ratio (peak-signal-to-noise ratio, PSNR), etc.
③speech object
signal-to-distortion ratio (signal-to- distortion ratio, SDR), etc.
See the paper for details, and it is also described in the Zhijian white paper -
Metrics Related to the Complexity of AI Models
Time Complexity and Space Complexity
Realization of Model-driven Semantic Communication in Intelligent-Driven Communication System
- Modulation
Modulation is applied to物理层
the signal in the medium. Different signals exist in different forms in the physical layer. Image signals are displayed in the form of pictures, and voice signals exist in the form of sound waves. Modulation of the model converts any form of signal into a number向量
or矩阵
expression. The modulation of the entropy reduction model extracts the semantic features and properties of the physical signal. For images, the entropy reduction model uses numeric vectors to express semantic features through modulation. - Encoding
Information can be expressed in different ways, and encoding is the process of converting information between different formats. Entropy-reduced model coding is an editor for signal semantics and attribute features, which can perform semantic and attribute features修改、替换、压缩和删除
. Encoding tries to preserve as much meaningful original information as possible. Likewise, decoding is not only the inverse process of encoding, but also the key to semantic and attribute features. It can vaguely restore the original signal and preserve the meaningful original information as much as possible.
The model can extract signal semantic features and change or adjust semantic features, transforming signals from one dimension to another for expression. At the same time, it can be expressed in reverse, and the original signal can be recovered. The amount of original information retained by the restored signal is related to the code length.