Sign in Subscribe
Artificial intelligence

Optimized Dimension-Reduction Algorithm for Diffractive Neural Networks

Strategic Insights

— 1 min read

Researchers from Southeast University have made a significant breakthrough in the field of photonics, proposing an optimized dimension-reduction algorithm for diffractive neural networks. This innovative approach enables diffractive neural networks to obtain inputs with stronger resolving capabilities, leading to improved accuracy in classification tasks. The researchers claim that their method achieves accuracy improvements of up to 7.3% on the Fashion-MNIST dataset compared to the traditional uniform grid sampling method.

Key Takeaways:

  • The optimized dimension-reduction algorithm, based on an improved Fisher score, enables diffractive neural networks to directly filter diffractive information injected into the network.
  • Numerical results demonstrate that the approach achieves accuracy improvements of 5% and 7.3% on the MNIST and Fashion-MNIST classification datasets, respectively.
  • The method holds an accuracy advantage of 4.8% in the pincushion distorted MNIST dataset and expands it to 6.7% in the tangential distortion scenario.
  • The research aims to miniaturize diffractive neural networks for wearable devices, industrial inspection, and autonomous driving systems.
  • The project was funded by the National Key Research & Development Program of China, State Key Laboratory of Millimeter Waves, Southeast University, China, Fundamental Research Funds for the Central Universities, Ministry of Education, China - 111 Project, China Postdoctoral Science Foundation, and Natural Science Foundation of Jiangsu Province.

Statistics:

  • 5% accuracy improvement on the MNIST classification dataset.
  • 7.3% accuracy improvement on the Fashion-MNIST classification dataset.
  • 4.8% accuracy advantage in the pincushion distorted MNIST dataset.
  • 6.7% accuracy advantage in the tangential distortion scenario.
  • The research was conducted by Qian Ma, Ruisi Li, Ze Gu, Yu Ming Ning, Xinxin Gao, and Tie Jun Cui from Southeast University.

Sources:

  • Optimized Dimension-reduction Algorithm On Input Layer for Diffractive Neural Networks. APL Photonics, 2025;10(7).
  • Qian Ma, Ruisi Li, Ze Gu, Yu Ming Ning, Xinxin Gao, and Tie Jun Cui. Optimized Dimension-reduction Algorithm On Input Layer for Diffractive Neural Networks. APL Photonics, 2025;10(7).

Read more