Centrala begrepp
Regularizing model parameters with prior knowledge of training data and Gaussian noise diffusion can improve the trainability of variational quantum circuits against barren plateaus and saddle points.
Sammanfattning
The authors propose a regularization strategy to improve the trainability of variational quantum circuits (VQCs) in the era of noisy intermediate-scale quantum (NISQ) devices. The strategy integrates two key mechanisms:
-
Leveraging prior knowledge of the training data to regularize the initial distribution of model parameters. This helps mitigate barren plateau issues, where the gradient variance exponentially decreases as the model size increases.
-
Diffusing Gaussian noise on the model parameters during training. This increases the volatility of the optimization process, helping the model avoid being trapped in saddle points.
The authors conduct extensive ablation studies across four public datasets - Iris, Wine, Titanic, and MNIST. The results demonstrate that:
- Incorporating prior knowledge of the training data in the initialization can effectively regularize various initial distributions and yield superior mitigation of barren plateau issues.
- Diffusing Gaussian noise during training can efficiently increase the volatility to avoid saddle points, while adequately alleviating the degradation of gradient variance.
- The authors also analyze the sensitivity of the key hyperparameter, max diffusion rate (drmax), and report the optimal values for each dataset and scenario.
Overall, the proposed regularization strategy, which combines prior knowledge and Gaussian noise diffusion, can significantly improve the trainability of VQCs compared to baseline methods.
Statistik
The number of qubits and layers in the variational quantum circuits are used as key metrics to analyze the effectiveness of the proposed regularization strategy.
Citat
"Regularizing model parameters with prior knowledge of the train data can effectively mitigate barren plateau issues."
"Diffusing Gaussian noise on model parameters during training can efficiently increase volatility to avoid being trapped in saddle points."