masK'IT

Team: cirKITers
GitHub: https://github.com/cirKITers/masKIT

Motivation

• It is not about the difficulty of climbing plateaus, but leaving them during optimization processes.

Motivation

• Barren plateaus (but also local minima) can have a big impact on trainability for parameterized quantum circuits

Idea

• Dropouts are being used in Machine Learning
• Preventing overfitting
• Helping to train all parts of the Neural Network

Questions

Do dropouts have a similar effect to quantum circuits as in Machine Learning?

How do they influence circuits of varying depth and height?

To which ansätze are dropouts applicable if at all?

Experimental Setup

• Pennylane to implement hybrid quantum-classical use case
• Gradient Descent implementation for parameter optimization
• Layered circuit ansatz as a template for scaling

Circuit Ansatz

• Circuit based on ansatz from McClean, J.R., Boixo, S., Smelyanskiy, V.N. et al. Barren plateaus in quantum neural network training landscapes. Nat Commun 9, 4812 (2018). https://doi.org/10.1038/s41467-018-07090-4

Algorithm

						if cost does not change for several steps:
							create variations of current circuit
							perform another training step
							measure costs for variations
							pick circuit with lowest cost
						else:
							perform another training step
							measure current cost
					

Ensemble-based Dropout

Results

• Improved learning rate
• Plot even includes penalty in number of steps for calculated ensembles for dropout

Results

• Fast learning possible with pretrained parameters when dropped gates are added back into the circuit

Results

• Good learning also for higher numbers of layers
• Better understanding needed to improve performance

Conclusions so far

• Dropouts can be applied for QML but differ from ML
• Ensemble-based dropouts for parameterized quantum circuits
• Rapid learning rates can be achieved
• Method is able to escape plateaus where gradient descent converges either very slow or not at all
• Allows for great tuning potential

Future Work

• Exploration of varying sizes of circuits
• Exploration of varying circuit ansätze
• Exploration of varying learning methods
• Exploration of further dropout mechanism
• Developing better understanding of interdependencies
• Targeted dropout of defined parts of a circuit/layer