In situ trainable optical convolution enabled by an AWG core and gradient-based model-free optimization

Area of Science:

• Photonics
• Artificial Intelligence
• Optical Computing

Background:

• Photonic neural networks (PNNs) offer potential for high-speed, low-power computation.
• Existing PNNs face challenges in scalability and efficient training, especially with physical non-idealities.

Purpose of the Study:

• To propose and validate a novel optical tensor processing unit (OTPU) architecture.
• To develop an effective training framework for PNNs leveraging optical components.

Main Methods:

• Utilized an arrayed waveguide grating (AWG) for parallel multiply-accumulate (MAC) operations.
• Implemented a gradient-based model-free optimization (G-MFO) framework for training.
• Encoded signed weights using Mach-Zehnder modulators (MZMs) and wavelength tuning.

Main Results:

• Achieved 98.80% accuracy on MNIST classification with 7.40 effective number of bits (ENOB) precision.
• Demonstrated versatile edge extraction on handwritten digits and breast ultrasound images.
• Showcased scalability through the use of additional wavelength channels.

Conclusions:

• The proposed AWG-based OTPU, trained with G-MFO, provides a scalable and training-efficient solution for PNN accelerators.
• This co-designed system addresses practical challenges in photonic computing for AI applications.
• The approach enables efficient in-situ training of non-differentiable optical weights.