Metasurface-Generated\nLarge and Arbitrary Analog Convolution\nKernels for Accelerated Machine Vision

In the rapidly evolving field of artificial intelligence,\nconvolutional\nneural networks are essential for tackling complex challenges, such\nas machine vision and medical diagnosis. Recently, to address the\nchallenges in processing speed and power consumption of conventional\ndigital convolution operations, many optical components have been\nsuggested to replace the digital convolution layer in the neural network,\naccelerating various machine vision tasks. Nonetheless, the analogous\nnature of the optical convolution kernel has not been fully explored.\nHere, we develop a spatial frequency domain training method to create\narbitrarily shaped analog convolution kernels using an optical metasurface\nas the convolution layer, with its receptive field largely surpassing\ndigital convolution kernels. By employing spatial multiplexing, the\nmultiple parallel convolution kernels with both positive and negative\nweights are generated under the incoherent illumination condition.\nWe experimentally demonstrate a 98.59% classification accuracy on\nthe MNIST data set, with simulations showing 92.63% and 68.67% accuracy\non the Fashion-MNIST and CIFAR-10 data sets with additional digital\nlayers. This work underscores the unique advantage of analogue optical\nconvolution, offering a promising avenue to accelerate machine vision\ntasks, especially in edge devices.