It is expected to significantly improve the computational performance of integrated photonic circuits. The Tsinghua team proposed a diffraction pattern neural network framework

Photonic neural networks use photons rather than electrons to perform brain inspired calculations , To achieve significantly improved computing performance . However , The existing architecture can only handle data with regular structure , However, it cannot be generalized to graph structure data outside Euclidean space .

The researchers of Tsinghua University put forward the diffraction pattern neural network (DGNN), This is a calculation unit based on diffractive photons (DPU) And all-optical diagrams of on-chip optical devices to represent the learning architecture , To address this limitation . say concretely , The graph node attributes are encoded into the banded optical waveguide , from DPU transformation , And the feature representation is extracted by optical coupler polymerization .

DGNN During the light speed optical message passing through the graph structure , Capture complex dependencies between nodes' neighborhoods . They demonstrated... Through the benchmark database DGNN Application in node and graph level classification tasks , And achieved excellent performance . The work of the research team has opened up a new direction for the use of deep learning to design application specific integrated photonic circuits for efficient processing of large-scale graphics data structures .

The research 「 All-optical graph representation learning using integrated diffractive photonic computing units」 entitled , On 2022 year 6 month 15 Published on 《 Science Advances》.

Thesis link ： https://www.science.org/doi/10.1126/sciadv.abn7630