WiMi Hologram Cloud, Inc., the world’s leading hologram augmented reality (“AR”) technology provider, announced that it is researching the use of neural networks for machine learning to optimize the parameters of dual-field quantum key distribution (TF-QKD) systems. This innovative approach aims to leverage the strong fitting ability and generalization performance of neural networks to directly predict the optimal parameter configuration of TF-QKD systems, significantly reducing computational time and resource consumption.
In this study, WiMi trained and evaluated three different types of neural network models.
Backpropagation Neural Network (BPNN): Based on the error backpropagation algorithm, BPNN minimizes prediction errors by continuously adjusting the network’s weights and biases. BPNN has become a preferred model in many fields due to its flexibility and wide applicability.
Radial basis function neural networks (RBFNN): Using radial basis functions as activation functions for hidden layer neurons, RBFNNs efficiently handle nonlinear problems and are particularly suited for high-dimensional data and scenarios that require high accuracy.
Generalized Regression Neural Network (GRNN): Based on probability density estimation, GRNN uses kernel function techniques to achieve nonlinear regression and is good at handling problems with small sample data and uncertainty.
Through training and testing these three neural network models, WiMi found that all models were able to predict the optimal parameters of the TF-QKD system with some accuracy. Among them, RBFNN and GRNN showed particularly good performance in high-dimensional parameter space and showed higher prediction accuracy. Compared to LSA, the neural network-based prediction method achieved a significant reduction in computation time, which was reduced by several orders of magnitude. BPNN had the fastest calculation speed due to its relatively simple structure. On the other hand, although RBFNN and GRNN are slightly more complex in terms of computational cost, they are still within an acceptable range, and the improved prediction accuracy has often resulted in more practical application value.
Also read: AiThority interview with Matej Bukovinski, Chief Technology Officer at Nutrient
Considering the different optimization needs of different TF-QKD systems, such as real-time and accuracy requirements, WiMi also conducted a comprehensive comparison of prediction accuracy and time consumption. This result shows that BPNN is an ideal choice in scenarios where fast response is required even with low request accuracy. On the other hand, RBFNN or GRNN are better suited for applications that prioritize high accuracy and can tolerate a certain amount of computation time.
The main technical advantages of using neural networks for TF-QKD system parameter optimization are to significantly reduce the computational complexity of parameter optimization, accelerate the key generation speed, and improve the real-time responsiveness of the system. Neural networks can automatically learn and adapt to changes in the quantum communication environment, providing the potential to dynamically adjust system parameters. As quantum communication technology develops, neural network models will be further upgraded and optimized to support more complex quantum key distribution protocols and higher security requirements.
WiMi will continue to deepen its research on neural networks for TF-QKD parameter optimization and explore more advanced neural network architectures and training strategies such as deep learning and reinforcement learning, aiming to realize a more efficient and intelligent quantum key distribution system. At the same time, it will strengthen integration with quantum communication hardware platforms, promote the practical application and commercialization of quantum communication technology, and contribute to the development of secure and efficient quantum communication networks.
