Harnessing optoelectronic noises in a photonic generative network

Area of Science:

• Optoelectronics
• Photonic Computing
• Artificial Intelligence

Background:

• Integrated optoelectronics offer efficient in-memory computing for neural network accelerators.
• Photonic systems face challenges due to inherent optoelectronic noise, leading to errors.

Purpose of the Study:

• To develop and demonstrate a noise-resilient photonic generative network for neural network acceleration.
• To mitigate and potentially harness noise in photonic computing systems.

Main Methods:

• Implemented a photonic generative network using a core of programmable phase-change memory cells.
• Performed vector-vector dot multiplication and utilized a generative adversarial network (GAN) framework.
• Incorporated noise-aware training by injecting additional noise into the system.

Main Results:

• Successfully generated a handwritten digit "7" in experimental settings.
• Achieved generation of all 10 digits in simulations.
• Demonstrated the network's resilience to hardware nonidealities and optical noise.

Conclusions:

• Photonic generative networks show significant resilience and potential for complex applications.
• The developed noise-aware training strategy enhances robustness in photonic computing hardware.
• This work paves the way for large-scale, realistic photonic hardware implementations.