forming diverse structured light arrays with general redefinable neural network

We propose a redefinable neural network (RediNet), realizing general modulation on diverse structured light arrays through a single approach. Exploiting the information sparsity of the array distribution, a redefinable dimension designation is used in RediNet, removing the burden of processing pixel-wise distributions. The prowess of originally generating arbitraryresolution holographs with fixed network is firstly demonstrated. The versatility is showcased in the generation of 2D/3D foci arrays, Bessel and Airy beams arrays, (perfect) vortex beam arrays, multi-channel compound vortex arrays and even snowflake-intensity arrays with arbitrarily-built phase functions. Considering the fine resolution, high speed, and unprecedented universality, RediNet can serve extensive applications such as next-generation optical communication, parallel laser direct writing, optical traps, and so on.