LO-SC: Local-only Split Computing for Accurate Deep Learning on Edge Devices

Luigi Capogrosso1, Enrico Fraccaroli1,2, Marco Cristani1 Franco Fummi1, Samarjit Chakraborty2
1 Department of Engineering for Innovation Medicine, University of Verona, Italy 2 Department of Computer Science, The University of North Carolina at Chapel Hill, USA
Paper arXiv Code
🎉 Accepted @ VLSI Design 2025 🎉
Teaser
Traditional SC approach versus the proposed LO-SC architecture.

Abstract

Split Computing (SC) enables deploying a Deep Neural Network (DNN) on edge devices with limited resources by splitting the workload between the edge device and a remote server. However, relying on a server can be expensive, requires a reliable network, and introduces unpredictable latency. Existing solutions for on-device DNNs deployment often sacrifice accuracy for efficiency. In this paper, we study how to use the concepts from SC to split a DNN for executing on the same device without compromising accuracy. In other words, we propose Local-Only Split Computing (LO-SC), a new approach to split a DNN for execution entirely on the edge device while maintaining high accuracy and predictable latency. We formalize LO-SC as a Mixed-Integer Linear Problem (MILP) problem and solve it using a multi-constrained ordered knapsack algorithm. The proposed method achieves promising results on both synthetic and real-world data, offering a viable alternative for accurately deploying DNNs on resource-constrained edge devices.

BibTeX

TBA.