Deep Unfolding Based Precoding for Task-Oriented Semantic Communications Over MIMO Channels

This paper studies model-driven deep learning techniques for precoding design in a task-oriented semantic communication system. In particular, we consider multi-device cooperative edge inference over a wireless multiple-input multiple-output (MIMO) multiple access channel, where multiple devices transmit extracted features to a server to perform a classification task. We formulate the precoding design as an inference accuracy maximization problem. This objective deviates from the traditional focus on throughput, mean-square error (MSE), bit error rate (BER), etc. By leveraging the deep unfolding technique, we develop a precoding network that effectively incorporates the domain knowledge of the solution to the precoding problem. Compared to the optimization based solution, the deep unfolded precoding network exhibits lower complexity by replacing the computation-intensive operations with some learnable parameters. Meanwhile, it offers extra flexibility for end-to-end (E2E) fine-tuning with the semantic codecs to push the performance limit. Simulation results on the ModelNet10 dataset verify that the proposed method achieves significantly higher classification accuracy compared to various baselines.