Ghost Classification Using Meta‐Decoders and Optical‐Electronic Correlations

ABSTRACT Diffractive neural networks, as a representative approach to free‐space optical diffractive information processing, exploit the intrinsic advantages of light, including low power consumption and parallelism, to efficiently perform various visual tasks. For a specific visual task, such as optical classification, a physical decoder composed of cascaded diffractive surfaces must be carefully trained and subsequently fabricated with high precision. However, the precision manufacturing of diffractive processors typically involves substantial cost and produces devices that are not reprogrammable, thereby limiting the achievable parallelism for handling multiple targets. In this work, linear optical decoders in diffractive computing are virtualized as meta‐decoders without a physical embodiment. This approach enables a hybrid optical‐electronic classification framework that exploits correlations between optically inferred fields and computer‐generated virtual reference fields. The proposed scheme integrates computational ghost diffraction with diffractive computing, referred to as ghost classification. It provides several advantages, including single‐point detection, a lens‐free configuration, pattern‐independent flexibility, reprogrammability, and the ability to classify multi‐class targets in parallel. This work leverages the complementary strengths of hybrid optical‐electronic inference while incorporating lightweight electrical computations through multiplication‐only correlation operations. The resulting framework serves as a transitional architecture in which each processing unit remains physically interpretable rather than a black box.