Optimal Partitioning of Distributed Neural Networks for Various Communication Environments

Abstract

Recently, it is increasingly necessary to run high-end neural network applications on top of resource-constrained embedded systems, such as wearable or Internet-of-Things devices. To cope with their high computation overheads on low-end systems, the distributed neural network approach in which multiple small neural networks separately and cooperatively operate on multiple devices has been proposed. While the computational overhead could be effectively alleviated by this approach, the existing techniques still suffer from large traffics between the devices, making it vulnerable to communication failures. This drawback hinders the application of the distributed neural network techniques to wearable devices, which may be connected with each other through unstable and low data rate communication medium like human body communication. Therefore, in this paper, we propose to improve the distributed neural network by adopting a partitioning method that can adapt to given communication environments. To validate the effectiveness of the proposed portioning technique, we compare the inference accuracies of the distributed neural networks that are partitioned differently for various communication environments.