Year: 2025
Type: article
Abstract: A scheme of a photonic convolution processor based on a tunable electro-optic frequency comb is proposed. The optical frequency comb (OFC) is generated using a dual-parallel Mach–Zehnder modulator (DPMZM) driven by an RF signal. By adjusting the working parameters of the DPMZM, OFCs with different line number and frequency spacing can be produced to reconfigure the convolution kernel dimensions. A linearly chirped fiber Bragg grating (LCFBG) is employed to implement interleaving of temporal and spectral dimensions. The interleaved signals are sampled at specific time and summed after optoelectronic conversion, and the convolution operation is completed. In this work, using a 10 GHz RF signal, a 4-line OFC with a frequency spacing of 20 GHz and a 2-line OFC with a spacing of 40 GHz are generated to obtain a 2 × 2 and 1 × 2 convolution kernels, respectively. The convolution results are fed into an electronic pooling layer and a fully connected layer for classifying the images of the MNIST handwritten digit dataset. The results demonstrate that a classification accuracy of 95.7% is achieved using the 2 × 2 convolution kernel, and a higher classification accuracy of 96.5% is obtained with the 1 × 2 convolution kernel.
Source: Photonics
Authors: Jiaming Wang, Juanjuan Yan
Institution: Beihang University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
Awards: G2160058833, 宽带、低相噪光学频率梳的产生机理及实验研究
OpenAlex ID: W7117109140