Compact optical convolution processing unit based on multimode interference.

Area of Science:

• Photonics
• Deep Learning
• Optical Computing

Background:

• Convolutional neural networks (CNNs) face limitations in speed and energy efficiency for massive data processing due to electrical frequency and memory access.
• Current optical computing schemes struggle with scalability, as optical element count increases quadratically with computational matrix size.

Purpose of the Study:

• To demonstrate a compact, scalable on-chip optical convolutional processing unit for deep learning applications.
• To address the scalability challenges in optical computing for large-scale integration.

Main Methods:

• Fabrication of a compact on-chip optical convolutional processing unit on a silicon nitride platform.
• Utilizing multimode interference cells and phase shifters to create correlated real-valued kernels for parallel convolution.
• Experimental demonstration of ten-class classification of handwritten digits using the MNIST database.

Main Results:

• Successful experimental demonstration of handwritten digit classification using the developed optical unit.
• The proposed design exhibits linear scalability with respect to computational size.
• The optical unit performs parallel convolution operations using interrelated kernels.

Conclusions:

• The developed on-chip optical convolutional processing unit offers a promising solution for overcoming the limitations of electrical deep learning systems.
• The linear scalability of the design indicates significant potential for large-scale integration in future optical computing architectures.
• This work paves the way for more efficient and faster deep learning hardware.