11 TOPS photonic convolutional accelerator for optical neural networks

Area of Science:

• Artificial Intelligence
• Optical Computing
• Machine Learning

Background:

• Convolutional neural networks (CNNs) are powerful AI tools for feature extraction, crucial for tasks like computer vision and medical diagnosis.
• Traditional electronic CNNs face limitations in speed and power consumption.
• Optical neural networks offer a promising avenue for accelerated computing by leveraging optical bandwidths.

Purpose of the Study:

• To demonstrate a universal optical vector convolutional accelerator.
• To achieve high-speed image recognition using an optical convolutional neural network.
• To explore the potential of integrated optical systems for complex AI tasks.

Main Methods:

• Development of an optical vector convolutional accelerator operating at over ten tera-ops per second.
• Utilized an integrated microcomb source to interleave temporal, wavelength, and spatial dimensions.
• Configured the hardware to form a sequential optical convolutional neural network for image recognition tasks.

Main Results:

• Demonstrated an optical accelerator capable of processing images with 250,000 pixels, suitable for facial recognition.
• Achieved 88% accuracy in recognizing handwritten digit images using the optical convolutional neural network.
• The system operates at speeds exceeding ten tera-ops per second.

Conclusions:

• The developed optical convolutional neural network accelerator offers a scalable and trainable platform for high-performance AI.
• This technology has significant potential for demanding applications like autonomous vehicles and real-time video recognition.
• The integrated approach using microcomb sources enables efficient processing by interleaving multiple dimensions.