Version 2 2024-12-11, 17:00Version 2 2024-12-11, 17:00
Version 1 2024-10-18, 16:00Version 1 2024-10-18, 16:00
preprint
posted on 2024-12-11, 17:00authored byRuitao Wu, Juncheng Fang, Rui Pan, Rongyi Lin, Kaiyuan Li, Ting Lei, Luping Du, Xiaocong Yuan
Inspired by neural network algorithms in deep learning, diffractive optical networks have arisen as new platforms for manipulating light-matter interactions. Inherited from the deep learning black box nature, clear physical meanings have never been given for these complex diffractive networks at the layer level, even though the systems are visible in physical space. Using exemplified mode conversion systems, we show how various physical transformation rules within diffractive networks can be unveiled given properly defined input/output mode relations. Surprising physical transformation division phenomenon and an optical analogy of gradientvanishing-effect have been observed and discussed for high-dimensional mode sorting tasks. The use of physical interpretation for efficiently designing a parameter-varying network has also been demonstrated. These physically interpretable optical networks resolve the contradiction between rigorous physical theorem and operationally vague network structure, and pave the way for further interpret advanced deep learning tasks and other physical neural networks.
History
Disclaimer
This arXiv metadata record was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.