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

Convolutional neural network (CNN) is currently one of the most important artificial neural networks. However, traditional CNN hardware architectures suffer from significant increases in energy consumption and processing time as the demand for artificial intelligence tasks grows. Here, a novel optical convolution computing strategy is proposed that leverages a continuously adjustable photoluminescent device (CA-PLD) as the optical convolution kernel, enabling parallel, all-optical convolution computing and greatly simplifying the traditional convolution process. Under ultraviolet illumination, the CA-PLD exhibits visible long-afterglow emission characteristics due to the charge trapping and retention effects. This allows for continuously adjustable light weights, facilitating arbitrary convolution operations. Building on this, parallel and efficient multiply-accumulate operations have been successfully demonstrated using CA-PLD arrays with different weight combinations. Moreover, space-transformable CA-PLD units enable applications in dilated convolution. In a semantic segmentation task with 20 categories, the CA-PLD units achieve higher Intersection over Union (IoU) values and accuracy. Therefore, the weight-adjustable and spatial transformable CA-PLD proposed in this work holds promise for future applications in intelligent optical computing systems and optical implementations of non-von Neumann architectures.