Learning Precoding for Semantic Communications

When knowing the goal of transmission, resources can be used more efficiently in semantic communication systems, where only the information necessary for accomplishing the goal needs to be transmitted. Existing works for semantic commu-nications do not investigate resource allocation. In this paper, we consider a multi-antenna-multi-subcarrier system for trans-mitting images to multiple users, by taking a goal of classifying the images as an example. We propose a semantic information-aware precoding policy to mitigate multi-user interference based on deep learning, where the modulated symbols of the users are input into a graph neural network together with estimated channel matrix for learning the policy. To emphasize the impact of harnessing semantic information on precoding, we apply two convolutional neural networks to learn the mapping from the image of each user to modulated symbols and the mapping from the received symbols of each user to a representation of the image, respectively. A fully-connected neural network is followed for image classification. After training these neural networks jointly, the learned precoding policy operates in a water-filling manner, which allocates more power for transmitting stronger symbols, where the important information for classification is carried. Simulation results show that the learned precoding policy is superior to existing precoding policies in reducing the bandwidth for transmission required to achieve an expected classification accuracy when the signal-to-noise ratio is low, channel estimation error is high, and the number of users is large,