Quantization-aware training for low precision photonic neural networks.

Area of Science:

• Photonics
• Artificial Intelligence
• Computer Engineering

Background:

• Deep Learning (DL) advances drive demand for efficient neuromorphic hardware accelerators.
• Photonic neuromorphic architectures offer high computational speed and energy efficiency (femtojoule per MAC).
• A key challenge is the reliance on expensive, high-precision analog-to-digital (ADCs) and digital-to-analog converters (DACs) for signal interfacing.

Purpose of the Study:

• To investigate quantization effects from ADCs/DACs as a noise source in photonic DL models.
• To develop a photonics-compliant training framework for DL models with limited precision.
• To reduce the necessity for high-precision, costly ADCs/DACs in photonic neuromorphic systems.

Main Methods:

• Studied quantization phenomena induced by limited-precision ADCs/DACs in photonic models.
• Developed and applied a novel training framework tailored for photonic DL architectures.
• Validated the framework across diverse neural network types: fully connected, convolutional, and recurrent.

Main Results:

• Quantization was characterized as an additional noise/uncertainty source in photonic DL models.
• The proposed framework effectively trains photonic DL models using limited-precision converters.
• Demonstrated successful application across various neural network architectures, confirming effectiveness.

Conclusions:

• The developed framework enables the training of photonic DL models with reduced precision requirements.
• This approach mitigates the bottleneck posed by expensive high-precision ADCs/DACs.
• Facilitates the development of more practical and energy-efficient photonic neuromorphic hardware.