A Review of Model Lightweighting Techniques Based on Edge Computing

Abstract

With the booming development of the Internet of Things (IoT) and the wide application of artificial intelligence at the edge, it has become an inevitable trend for deep learning models to be deployed on resource-constrained edge devices. Model lightweighting techniques have made significant progress in the past five years as a solution to the problem of limited computation, storage and power consumption resources in edge devices. In this paper, we systematically review the research on model lightweighting techniques for edge computing in the past five years (2020-2025). In this paper, the mainstream lightweighting methods are classified into two categories: parametric compression and structural compression based on the core idea of compression. Parameter compression focuses on reducing redundant parameters, and focuses on the latest breakthroughs in the directions of parameter pruning, parameter quantization, and parameter sharing, which significantly reduce the model size and computation volume while effectively balancing accuracy and hardware friendliness. Structural compression techniques, on the other hand, focus on designing more efficient network architectures, with an in-depth discussion of the core ideas and representative work on compact network design and knowledge distillation. Then, examples are given on the application scenarios and key challenges of lightweight models. Finally, future research directions for the above challenges are envisioned.