Joint Channel Estimation and Classification for Task-Oriented Communications

Task-oriented communications, which focus on the effective execution of the task of interest rather than the recovery of the transmitted symbols, have great potentials in future Internet of Things (IoT) systems. In particular, the fundamental change in the design philosophy brings new opportunities for the physical-layer technique. In this paper, we consider a new joint channel estimation and classification (JCEC) receiver design for a multi-device edge inference system. By fully exploiting both the feature and the channel distributions, we establish a maximum $a$ posteriori (MAP) based JCEC formulation, which turns out to be a challenging non-convex problem. A majorization-minimization (MM) algorithm is proposed to solve the difficult problem, where only a closed-form expression needs to be calculated in each iteration. Simulations validate that our scheme outperforms the benchmark with decoupled channel estimation and MAP classifier on both a synthetic feature dataset and the widely used CIFAR-10 dataset.