Task-Oriented Communication for Multi-Device Edge Inference: A Maximal Coding Rate Reduction Approach

In this paper, we consider a task-oriented communication system for multi-device edge inference over a multiple-input multiple-output (MIMO) multiple-access channel, where the learning (feature encoding and classification) and communication (precoding) modules are designed to achieve the same goal of inference accuracy maximization. End-to-end learning of these modules involves both the task dataset and the time-varying wireless channel, which may result in unaffordable training overhead and complexity. Instead, we advocate a modular design of learning and communication to achieve the consistent goal. Specifically, we leverage the maximal coding rate reduction (MCR
²
) objective as a surrogate to represent the inference accuracy, which allows us to explicitly formulate the precoding optimization problem separated from the learning design. We develop a block coordinate ascent (BCA) algorithm for efficient problem-solving. Moreover, the MCR
²
objective also serves the loss function of the feature encoding network, thus unifying the objectives of learning and communication. Simulation results demonstrate the superior performance of the proposed method compared to various baselines. As such, our work paves the way for further exploration into the synergistic alignment of learning and communication objectives in task-oriented communication systems. Code is available at https://github.com/chang-cai/TaskCommMCR2.