This is the presentation of results from the team cirKITers during QHACK2021.
You can access the dynamic presentation at https://cirKITers.github.io/masKIT-presentation.
Nowadays training parameterized quantum circuits is very popular to explore the space of quantum states. This field heavily borrows from existing machine learning approaches. Similar to the effect of vanishing gradients in machine learning, we explore different plateaus such as local minima as well as Barren Plateaus in gradients making the training very hard or even unfeasible for various use cases.
Inspired by the classical dropouts in machine learning we explore the impacts of (temporarily) removing randomly selected gates from the circuit. Our experiments show that ensemble-based dropouts can speed up training or even enable training in presence of plateaus for parameterized quantum circuits.