On the Discussion of the Semantic Communication with the Transmission of MNIST Dataset

Inspired by the success of artificial intelligence (AI) in various fields, recently, AI-based semantic communication has been studied extensively and achieved great success. However, the existing works may not provide a comprehensive understanding of how channels affect semantic communication from phenomena. In this paper, based on the observations on the MNIST dataset, the following questions are discussed: (1) Why is semantic communication robust? We believe and observe in the experiments that the trade-off learning between the characteristic and common information of data is one of the key properties of semantic communication. The “channel robustness” and “hard generalization” phenomena are explained in this perspective. (2) How does noise affect semantic communication in the AWGN channel? Because of noise, there is a smallest distinguishable unit, called as semantic constellation cluster (SCC), where the features in the SCC are confused by noise and can not be distinguishable. The feature criterion (FC) is analyzed and observed in the experiments, which means the minimum distance of features that the features can just be distinguishable and is also the radius of SCC in the AWGN channel. (3) How does the semantic decoder distinguish noised features under the large noise in the AWGN channel? It is found that the outer edge of the expending space brought by noise is easy to be distinguish for the decoder, thus maintaining fairly good performance under the large noise.