High-accuracy optical convolution unit architecture for convolutional neural networks by cascaded acousto-optical...

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Computer Engineering

Background:

• Optical neural networks (ONNs) offer advantages in speed and power efficiency for AI acceleration.
• Existing optical computing hardware primarily focuses on fully-connected networks, limiting broader application.
• Convolutional neural networks (CNNs) are crucial for many AI tasks but require specialized hardware for efficient optical implementation.

Purpose of the Study:

• To propose and experimentally validate a novel optical convolution unit (OCU) architecture.
• To demonstrate the OCU's capability to perform convolutions for CNNs through hardware reuse.
• To assess the OCU's performance in terms of signal-to-distortion ratio (SDR) and inference accuracy.

Main Methods:

• Designed a modular optical convolution unit (OCU) architecture.
• Implemented the OCU using cascaded acousto-optical modulator arrays for proof-of-concept.
• Conducted experiments with both ex-situ and in situ training of neural network parameters.
• Evaluated OCU performance on standard CNN tasks.

Main Results:

• The OCU architecture allows convolutions with arbitrary input sizes by reusing the unit.
• Ex-situ training yielded an SDR of 28.22 dBc, with successful inference on CNN tasks.
• In situ training improved the SDR to 36.27 dBc, demonstrating potential for refinement.
• The OCU architecture is simplified, scalable, and suitable for integration into chip-scale optical CNNs.

Conclusions:

• The proposed OCU architecture effectively performs convolutions for optical CNNs.
• The OCU demonstrates high accuracy and signal quality, adaptable through training methods.
• The modular design facilitates the development of future large-scale, integrated optical neural networks.