Noise-resilient and high-speed deep learning with coherent silicon photonics

Area of Science:

• Photonics and optical engineering
• Artificial intelligence and machine learning
• Computer hardware and architecture

Background:

• Deep learning's computational demands drive innovation in specialized hardware.
• Integrated photonics offer energy-efficient solutions for high-speed computing, particularly for multiply-accumulate operations.
• Existing photonic neural networks face limitations in compute rate and noise resilience.

Purpose of the Study:

• To experimentally demonstrate a noise-resilient coherent photonic neural network.
• To achieve high compute rates and maintain accuracy in photonic deep learning hardware.
• To improve upon state-of-the-art coherent photonic implementations.

Main Methods:

• Fabrication of a silicon photonic chip implementing a coherent photonic neural network.
• Experimental evaluation using the MNIST dataset for classification tasks.
• Operation at compute rates of 5 and 10 Giga Multiply-Accumulate operations per second per axon (GMAC/sec/axon).

Main Results:

• Achieved >99% accuracy at 5 GMAC/sec/axon and >98% accuracy at 10 GMAC/sec/axon.
• Demonstrated 6x higher on-chip compute rates compared to previous coherent implementations.
• Showcased a >7% improvement in accuracy over state-of-the-art coherent photonic neural networks.

Conclusions:

• The developed noise-resilient coherent photonic neural network significantly advances photonic deep learning hardware.
• This technology offers a promising path towards ultra-high compute rates and energy efficiency in AI.
• The demonstrated performance highlights the potential of integrated photonics for next-generation AI accelerators.