[SoK] Systematizing Inference Placement For Deep Learning Across Edge And Cloud Platforms: A Multi-Objective Optimization Perspective

The growing number of edge intelligent applications such as virtual and augmented reality and large language model-based chatbots, spurred by the development of the Internet of Things (IoT) and mobile devices, makes it increasingly difficult to deploy complicated deep learning (DL) models on edge devices that have limited resources. To work around such challenges, studies have sought ways to optimize and offload partitions of DL models onto user devices and edge and cloud servers. This distributed approach utilizes the different strengths of each computing environment, such as edge resources which provide low-response latency and cloud resources which offer economical computation for heavy workloads. Unfortunately, the communication that occurs between partitions

of DL models creates transmission bottlenecks and makes data less secure. Current studies attempt to find the optimal balance between inference latency, the cost of computation, the time it takes to transmit, and the privacy of the data, primarily through model compression and distillation, transmission compression, and classifiers that are internal to the model architecture. This project topic guides organizing a research effort that situates leading-edge model offloading and model adaptation methods and their impact on multi-objective optimization, with a primary focus on inference latency, data privacy, and monetary resource constraints. The goal is to determine precise frameworks and assessments to catalyze the development of distributed DL inference systems that

are robust, cost-effective, and secure.

Message: Hi FE, I need help with “[SoK] Systematizing Inference Placement For Deep Learning Across Edge And Cloud Platforms: A Multi-Objective Optimization Perspective” in “Electronics & Communication Engineering”