Nonlinear processing with linear optics.

Area of Science:

• Photonics
• Optical Computing
• Artificial Intelligence Hardware

Background:

• Deep neural networks (DNNs) offer breakthroughs but require significant electronic computing power.
• Optical implementations promise enhanced energy efficiency and speed by utilizing optical bandwidth and interconnections.
• A key challenge is implementing multilayer optical networks without electronics due to the lack of low-power optical nonlinearities.

Purpose of the Study:

• To present a novel optical framework for realizing programmable linear and nonlinear transformations in neural networks.
• To overcome the limitations of electronic components in multilayer optical network implementation.
• To achieve low-power, high-speed optical computing.

Main Methods:

• Utilizing multiple scattering of light through a scattering medium.
• Leveraging the nonlinear relationship between scattering potential (data) and the scattered optical field.
• Employing low-power continuous-wave light for optical nonlinear computing.

Main Results:

• Demonstrated a novel framework capable of synthesizing programmable linear and nonlinear transformations concurrently.
• Showcased nonlinear optical computing with low-power continuous-wave light through repeated data scattering.
• Empirically found that the scaling of this optical framework follows a power law.

Conclusions:

• Multiple scattering offers a viable pathway for low-power, high-speed optical neural network implementation.
• This approach bypasses the need for electronic components in multilayer optical networks.
• The power-law scaling suggests potential for efficient scalability of optical computing architectures.