Fig. 1. Model and training of a multitasking optoelectronic hybrid neural network based on a nonlinear metasurface. (a) Sketch of the network model with handwritten digital inputs for parallel classification and image reconstruction; (b) training convergence results of the classification and reconstruction network; (c) fundamental frequency phase distribution of the nonlinear metasurface obtained after training the network

Fig. 2. Schematic and transmission spectrum of nonlinear metasurface units. (a) U-shaped resonant unit with directional angle generating linear and nonlinear geometric phases under circular polarization; (b) transmission spectrum of the metasurface unit under circularly polarized incident light

Fig. 3. Simulation results of proposed network classification performance at fundamental frequency. (a) Inputs of 10 handwritten digits randomly selected from the correct results, optical output of the middle part of the network, and the output of the whole network; (b) some misclassified cases; (c) confusion matrix of the full test set

Fig. 4. Simulation results of proposed network encoding and reconstructing the input image under multiple frequencies. (a) Input images of 10 handwritten digits (same as the input of the classification network), optical output images of the middle part of the network, and output images of the whole network, i.e., the reconstructed image; (b) imaging results of the misrecognized cases of the classification network; (c) reconstruction quality statistics of the 10 handwritten digits in the full test set