Disaggregated machine learning via in-physics computing at radio frequency.

Area of Science:

• Computer Engineering
• Wireless Communication
• Machine Learning

Background:

• Edge devices require efficient machine learning for intelligent applications.
• Traditional digital computing architectures face memory and power constraints for edge AI.
• Resource-constrained edge devices limit real-time machine learning inference.

Purpose of the Study:

• To present WISE, a novel computing architecture for wireless edge networks.
• To enable efficient and simultaneous machine learning inference on multiple edge devices.
• To reduce the energy consumption of machine learning computations at the edge.

Main Methods:

• Developed WISE, a computing architecture for wireless edge networks.
• Implemented disaggregated model access via over-the-air wireless broadcasting.
• Utilized in-physics computation of complex-valued matrix-vector multiplications at radio frequency.
• Employed a software-defined radio platform for experimentation.

Main Results:

• Achieved 95.7% image classification accuracy and 97.2% audio classification accuracy.
• Demonstrated ultralow energy consumption: 6.0 fJ/MAC for image classification and 2.8 fJ/MAC for audio classification.
• Showcased over a 10x improvement in energy efficiency compared to traditional digital computing (e.g., GPUs).

Conclusions:

• WISE offers a significant advancement in edge computing for machine learning.
• The proposed architecture overcomes the limitations of traditional digital computing for edge AI.
• WISE enables highly accurate and energy-efficient AI inference on resource-constrained wireless edge devices.