A Review of Embedded Machine Learning Based on Hardware, Application, and Sensing Scheme

Area of Science:

• Computer Science
• Artificial Intelligence
• Embedded Systems

Background:

• Machine learning (ML) is increasingly vital across industries, with growing implementation on embedded systems.
• Despite advances in embedded device capabilities, high power and computational demands of ML algorithms pose challenges.
• Embedded machine learning offers tailored solutions based on device processing power and performance.

Purpose of the Study:

• To provide an overview of machine learning implementations on embedded systems.
• To categorize implementations by device (NVIDIA Jetson, Raspberry Pi), application, ML algorithm, and sensors.
• To analyze device suitability for specific applications and common ML models/sensors used.

Main Methods:

• Systematic review of papers from Google Scholar and IEEExplore.
• Selection criterion: use of embedded computing systems for ML (theoretical or practical).
• Data extraction focused on model accuracy, power consumption, and inference time.

Main Results:

• NVIDIA Jetson and Raspberry Pi are prominent embedded platforms for ML.
• Analysis of ML models and sensors commonly paired with specific embedded devices.
• Identified improvements in embedded ML quality, power usage, and overall effectiveness.

Conclusions:

• Embedded machine learning is rapidly expanding due to hardware and algorithm advancements.
• Device selection is critical for optimizing ML performance in specific applications.
• Ongoing improvements in embedded ML systems enhance their quality, power efficiency, and effectiveness.