Integrated multi-operand optical neurons for scalable and hardware-efficient deep learning.

Area of Science:

• Photonics
• Neuromorphic Computing
• Optical Neural Networks

Background:

• Integrated photonic tensor cores (PTCs) face challenges with large area costs and high propagation loss due to single-operand modulators.
• Efficient implementation of large tensor operations in optical neural networks (ONNs) is crucial for advancing neuromorphic computing.

Purpose of the Study:

• To propose a scalable and efficient optical dot-product engine using multi-operand photonic devices.
• To demonstrate the effectiveness of a multi-operand optical neuron (MOON) for image recognition tasks.

Main Methods:

• Development of a multi-operand optical neuron (MOON) based on a multi-operand Mach-Zehnder interferometer (MOMZI).
• Experimental demonstration of a MOMZI-based ONN for image recognition.
• Performance analysis comparing MOMZI-based PTCs with single-operand MZI-based counterparts.

Main Results:

• The MOMZI-based ONN achieved 85.89% accuracy on the Street View House Number (SVHN) dataset with 4-bit voltage control.
• A 128x128 MOMZI-based PTC demonstrated superior performance over single-operand MZIs in propagation loss, optical delay, and device footprint.
• Comparable matrix expressivity was maintained with significant improvements in efficiency.

Conclusions:

• The proposed MOON, implemented with MOMZI, offers a scalable and efficient solution for photonic tensor cores.
• This approach significantly reduces hardware overhead and improves performance for large-scale neuromorphic computing applications.
• The MOMZI-based design represents a substantial advancement in integrated photonic devices for AI hardware.