AIfES: A Next-Generation Edge AI Framework

Area of Science:

• Embedded Systems
• Artificial Intelligence
• Machine Learning

Background:

• Edge Artificial Intelligence (AI) integrates Machine Learning (ML) into embedded devices for local processing.
• Traditional edge AI frameworks exhibit limitations in hardware flexibility and custom accelerator integration.
• These limitations hinder the adoption of advanced ML innovations in resource-constrained environments.

Purpose of the Study:

• To introduce the Artificial Intelligence for Embedded Systems Framework (AIfES) designed to address limitations of existing edge AI frameworks.
• To provide a detailed overview of AIfES architecture and its underlying design principles.
• To evaluate AIfES performance against TensorFlow Lite for Microcontrollers (TFLM) on embedded hardware.

Main Methods:

• A detailed architectural overview of AIfES is presented.
• Performance comparison between AIfES and TFLM was conducted on an ARM Cortex-M4 System-on-Chip (SoC).
• Evaluations utilized fully connected neural networks (FCNNs) and convolutional neural networks (CNNs).

Main Results:

• AIfES demonstrated superior performance over TFLM in execution time and memory consumption for FCNNs.
• AIfES achieved up to a 54% reduction in memory consumption for CNNs compared to TFLM.
• AIfES successfully enabled on-device training of CNNs on a resource-constrained device (<100 kB RAM).

Conclusions:

• AIfES offers a flexible and efficient framework for edge AI applications, overcoming limitations of traditional approaches.
• The framework supports advanced ML tasks, including on-device training, on highly constrained embedded systems.
• AIfES represents a significant advancement for deploying sophisticated AI capabilities in real-world embedded applications.