A Systematic Literature Review on Distributed Machine Learning in Edge Computing

Area of Science:

• Computer Science
• Artificial Intelligence
• Distributed Systems

Background:

• Edge computing enables processing data near its source, reducing latency and bandwidth usage.
• Distributed edge intelligence (DEI) applies machine learning (ML) and deep learning (DL) algorithms on these edge devices.
• Edge devices present unique challenges due to their limited resources and heterogeneity compared to cloud environments.

Purpose of the Study:

• Investigate the challenges and adaptations required for running distributed ML/DL on resource-constrained edge devices.
• Analyze techniques for optimizing ML/DL processes including caching, training, inference, and offloading in edge settings.
• Explore the advantages and disadvantages of various strategies for DEI.

Main Methods:

• Review and analysis of existing and novel techniques for distributed ML/DL on edge devices.
• Focus on adaptations for restricted device capabilities in caching, training, inference, and offloading.
• Comparative assessment of different approaches' benefits and drawbacks.

Main Results:

• Identified key challenges in implementing distributed ML/DL on heterogeneous edge devices.
• Detailed examination of adapted techniques for caching, distributed training, efficient inference, and intelligent offloading.
• Evaluation of trade-offs associated with different DEI strategies.

Conclusions:

• Successfully running ML/DL on edge devices requires specialized techniques to overcome resource limitations.
• Optimized caching, distributed training, efficient inference, and strategic offloading are crucial for effective DEI.
• Understanding the benefits and drawbacks of these strategies is essential for successful deployment.