Joint Resource Allocation and Computation Offloading for DNN Inference with Model Partition and Early Exit in MEC Networks

The development of artificial intelligence and edge computing gives rise to the edge intelligence, where the intelligent tasks such as deep neural network (DNN) based inference can be cooperatively performed by edge server and devices. Existing methods often treat DNN based tasks as indivisible units, optimizing computation offloading and resource allocation without exploiting the structural flexibility of DNN models. In this work, we investigate the joint optimization problem of resource allocation and computation offloading in the multi-access edge computing (MEC) network for DNN inference tasks. The objective is to maximize the inference accuracy while minimizing the task latency, through the joint design of DNN model partition, early exiting selection, device association, bandwidth and computation resource allocation. In addition, we consider the long-term optimization so the random generation of tasks and queue stability at devices and servers are also taken into account. Since the formulated problem is hard to solve, we opt to the deep reinforcement learning approach and leverage the transformer aided deep deterministic policy gradient algorithm to find the solution. Simulations confirm the promising performance of the proposed algorithm, which outperforms the other benchmark schemes.