Recurrent diffractive deep neural networks

What we believe is a novel recurrent diffractive deep neural network (RD2NN) is proposed for image time division multiplexing and frequency division multiplexing. The RD2NN is formed by a diffractive deep neural network (D2NN) with its output connected backward to the input. Therefore, it enables the signals to be generated sequentially in the time domain. By precoding the images via the inverse Fourier transform (IFT), one may also realize frequency division multiplexed images. To train the RD2NN, a modified optical real-time back-propagation (BP) algorithm is proposed, which expands the RD2NN into sequential D2NNs with identical phase configurations. The temporal output images of the previous stage D2NN are used as the input image for the next stage D2NN during the training. Five consecutive images are generated either in the time domain or in the frequency domain with a 7-layer trained RD2NN.