ML-Based Task-Oriented Offloading: A Survey

In the Industrial Internet of Things (IIoT), machine learning (ML) plays a crucial role in driving intelligent, data-driven processes, especially in applications that require computation-intensive tasks, such as real-time defect detection, predictive maintenance, and quality inspection. To meet the demands of these ML tasks, effective offloading strategies are essential for reducing latency and ensuring high inference accuracy. This paper surveys recent advancements in ML-based task-oriented offloading strategies, focusing on two primary directions: (1) enhancing timeliness and quality in computation-intensive tasks and (2) ensuring accuracy-aware offloading for high-reliability applications in IIoT. The first direction addresses the need for rapid processing by minimizing task delays and optimizing real-time scheduling, while the second focuses on maintaining inference accuracy across resource-constrained environments, essential for industrial applications. By categorizing and analyzing the latest strategies, this paper highlights the evolution of ML-based task offloading in IoT and IIoT systems, emphasizing the balance between timeliness, accuracy, and resource management.