High-Efficiency Reconfigurable Optical Convolutional Neural Network

As the intricacy of network models escalates and the volume of data swells exponentially, the demands on the processing capabilities of neural networks have become increasingly rigorous. Optical computing, with the excellent performance of extremely large bandwidth and ultra-low loss, is poised to be a disruptive technology that supports the needs of the next generation high performance artificial intelligence. In this paper, an optical convolutional neural network, combining a novel architectural design with a compatible data encoding technique, is introduced to achieve superior data efficiency by harnessing the dimensions of wavelength, temporal and parallel space. Utilizing 3×3 convolution kernel on 28×28-pixel images, the proportion of valid data in the output temporal waveform achieves as high as 92.86%. Furthermore, by employing multi-kernel parallel for processing larger size image, the data efficiency is ulteriorly enhanced to an astonishing 99%, and the potential computing power is increased to 10.51 TOPS, showcasing a remarkable leap in performance. During the experiment, five reconfigurable kernels including two positive and three real-value kernels, in conjunction with an electrical fully connected layer, successfully classified handwritten digital images from “0” to “9” in Modified National Institute of Standards and Technology (MNIST) database, achieving an accuracy of 92.81%. The defining feature of the proposed structure is that high-efficiency reconfigurable optical convolutional neural networks streamline the preprocessing and extraction of the data, markedly reducing the meaninglessdata ratio. This innovation, offering a new architecture for optical convolutional neural networks, not only conserves computing power but also improve processing speed, which is expected to alleviate the expanding demand for computing performance in deep learning.