TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI Applications.

Area of Science:

• Computer Science
• Artificial Intelligence
• Embedded Systems

Background:

• Internet of Things (IoT) devices increasingly utilize machine learning (ML) for intelligent decision-making.
• Resource limitations on IoT devices hinder the deployment of complex ML models, including deep learning (DL).
• Cloud-based processing for IoT introduces latency, privacy concerns, and increased communication costs.

Purpose of the Study:

• To provide a comprehensive overview of the Tiny Machine Learning (TinyML) revolution.
• To review existing TinyML studies, analyzing ML models, datasets, and device characteristics.
• To clarify the state-of-the-art in TinyML and identify future development requirements.

Main Methods:

• Review and analysis of scientific literature focusing on TinyML applications.
• Categorization of ML models employed within TinyML studies.
• Examination of datasets and hardware specifications relevant to TinyML implementations.

Main Results:

• TinyML facilitates on-device data processing and ML model inference, including DL, on microcontrollers.
• Local processing addresses challenges of latency, data privacy, and communication costs associated with cloud-dependent IoT.
• Analysis reveals diverse ML models, datasets, and device types are utilized in current TinyML research.

Conclusions:

• TinyML is a transformative technology enabling intelligent edge computing for resource-constrained IoT devices.
• The study provides a foundational understanding of TinyML's current landscape and future potential.
• Further research is needed to address development requirements for broader TinyML adoption.