Photonic diffractive generators through sampling noises from scattering media.

Area of Science:

• Photonic Computing
• Optical Neural Networks
• Generative Models

Background:

• Photonic computing offers high parallelism, low latency, and energy efficiency for neural network (NN) accelerators.
• Existing photonic accelerators primarily focus on discriminative NNs, leaving large-scale generative machines underexplored due to data accessibility and hardware challenges.

Purpose of the Study:

• To develop a large-scale generative photonic computing machine for image generation.
• To address data accessibility and hardware feasibility issues in photonic generative models.
• To enhance image generation performance using advanced photonic NN architectures.

Main Methods:

• Harnessing random light scattering in disordered media as a native noise source.
• Employing large-scale diffractive optical computing for image generation.
• Designing two encoding strategies for image-to-optical noise latent space mapping.
• Utilizing cascaded and parallel configurations of diffraction layers in photonic NN architectures.

Main Results:

• Successfully generated clear and meaningful synthesized images across standard public datasets.
• Demonstrated hardware consistency by leveraging the spatial parallelism of light.
• Overcame training problems through effective encoding strategies for experimental data accessibility.

Conclusions:

• This work presents a significant contribution to photonic computing by realizing a functional photonic generative machine.
• The developed photonic generator paves the way for sophisticated applications like real-world data augmentation and multimodal generation.
• The study highlights the potential of disordered media and diffractive optics for advanced generative tasks.