A Simple Optical Convolution Strategy Based on Versatile Adjustable Optical Convolution Kernel for All-Optical...

Area of Science:

• Optoelectronics
• Artificial Intelligence Hardware
• Optical Computing

Background:

• Traditional Convolutional Neural Network (CNN) hardware faces challenges with high energy consumption and processing time.
• Growing demand for artificial intelligence tasks exacerbates limitations of current CNN architectures.

Purpose of the Study:

• To propose a novel optical convolution computing strategy to address energy and speed limitations in CNNs.
• To introduce continuously adjustable photoluminescent devices (CA-PLDs) as a solution for efficient optical convolution.

Main Methods:

• Leveraging CA-PLDs as optical convolution kernels for parallel, all-optical convolution.
• Utilizing the long-afterglow emission characteristics of CA-PLDs for continuously adjustable light weights.
• Demonstrating parallel multiply-accumulate operations using CA-PLD arrays and space-transformable units for dilated convolution.

Main Results:

• Successfully demonstrated parallel and efficient multiply-accumulate operations with CA-PLD arrays.
• Achieved higher Intersection over Union (IoU) values and accuracy in a 20-category semantic segmentation task.
• Showcased the potential of space-transformable CA-PLD units for dilated convolution applications.

Conclusions:

• The proposed weight-adjustable and spatially transformable CA-PLD offers a promising approach for intelligent optical computing.
• CA-PLDs can significantly simplify the traditional convolution process, enabling efficient all-optical computation.
• This technology holds potential for future non-von Neumann optical computing architectures.