Abstract:

The data reuploading trick, originally proposed for universal quantum computing, enables the universal approximation property. We recently extended this concept to achieve universal non-quantum photonic computing using practical photonic integrated circuits (PICs) composed solely of 50:50 beam splitters and phase shifters, eliminating the need for nonlinear photonic devices. In this approach, input data are repeatedly embedded as rotation angles. In this presentation, we explore strategies to enhance the performance of this method by increasing the number of layers and optical channels (lanes) in various configurations. For a classical two-dimensional, four- class classification problem of wavy lines, we evaluated a two-mode class-embedding circuit with output, a four- mode configuration, and four stacking methods of two-mode circuits with average pooling, alongside a baseline configuration using projection in a complex domain. The first three configurations demonstrated excellent accuracy. Additionally, we investigated the effect of shifting the order of input data layer by layer, which significantly improved performance in certain applications. These findings highlight a novel architectural approach to photonic neural networks enabled by data reuploading in PICs. This approach offers unique features that set it apart from traditional photonic neural network architectures, providing a promising direction for future advancements in photonic computing