Optical patching scheme for optical convolutional neural networks based on wavelength-division multiplexing and...

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Computer Engineering

Background:

• Optical neural networks (ONNs) are emerging as efficient deep learning accelerators.
• Optical convolutional neural networks (CNNs) offer novel architectures for convolutional models.
• Current optical CNNs rely heavily on electronics for data patching, requiring large input modulator arrays.

Purpose of the Study:

• To introduce and experimentally demonstrate an optical patching scheme for optical CNNs.
• To reduce the reliance on electronic data processing in optical CNNs.
• To decrease the scale of the input modulator array required for optical CNNs.

Main Methods:

• Implemented an optical patching scheme using optical delay lines to replace electronic processing.
• Utilized wavelength-division multiplexing (WDM) to enable simultaneous processing of multiple input data streams.
• Experimentally validated the proposed optical patching scheme in an optical CNN context.

Main Results:

• Successfully demonstrated an optical patching scheme, reducing the need for electronic data processing.
• Significantly decreased the scale of the input modulator array for optical CNNs.
• Showcased the efficiency and reduced system complexity through the use of optical delay lines and WDM.

Conclusions:

• The developed optical patching scheme offers a novel solution for data input challenges in ONNs.
• This approach alleviates the burden of electronic data processing, paving the way for more efficient optical deep learning accelerators.
• The integration of optical delay lines and WDM presents a promising direction for future optical CNN architectures.