Efficient Multi-dimensional Compression for Network-edge Classification

Efficient Multi-dimensional Compression for Network-edge Classification

연구 분야: Infrastructure

학회: MOBIHOC '24: Proceedings of the Twenty-fifth International Symposium on Theory, Algorithmic Foundations, and Protocol Design for Mobile Networks and Mobile Computing

The widespread adoption of low-cost resource-constrained edge devices and high-performance expensive servers necessitates shifting the complexity burden from edge devices to servers. However, in many applications such as image classification, it is often impractical and communication expensive to transmit full information without any form of compression. To address this issue, this paper introduces a neural network (NN)-based compression technique tailored for resource-constrained edge devices for classification at the network edge. The core idea involves simultaneously training a shallow neural network to-be-implemented by the devices and a deep neural network to-be-implemented by the server. To adapt to the time-varying channel conditions, the compression algorithm at the device side must be able to handle multiple output dimensions. To address this issue, we develop two multi-dimensional compression strategies: the multiple codebook approach, using separate NNs for various dimensions, and the single codebook approach, utilizing one NN for all dimensions. The single codebook approach substantially reduces the storage demands on the device, offering a viable solution for low-cost edge devices. Our analysis offers a theoretical performance guarantee, highlighting that the accuracy of the single codebook approach is comparable to that of the multiple codebook strategy. Through empirical evaluations on real-world datasets, we demonstrate that the single codebook approach achieves near-equivalent performance to the accuracy multiple codebook alternative.