JMSNAS: Joint Model Split and Neural Architecture Search for Learning over Mobile Edge Networks

Deep neural networks (DNNs) enable the booming intelligent mobile applications owing to the effectiveness and reliability. With the rapid development of fifth generation (5G) communication networks, edge devices can be deployed with computation capability, which makes the distributed implementation of DNN over a wireless network possible. However, the simplified model splitting method cannot guar-antee the accuracy and latency performance. They call for the iterative design of model splitting and parameter updating. In this paper, a joint model split and neural architecture search (JMSNAS) framework is proposed to deploy the generated DNN model over mobile edge networks. Considering both computing and communication resource constraints, we re-formulates the multi-split problem to a computational graph search problem and optimizes the objective function to realize the trade-off between model accuracy and completion latency. The experiment results reveal the superiority of the proposed JMSNAS over state-of-the-art split machine learning model designs.