Visual explanations of machine learning model estimating charge states in quantum dots
Authors:
Yui Muto,
Takumi Nakaso,
Motoya Shinozaki,
Takumi Aizawa,
Takahito Kitada,
Takashi Nakajima,
Matthieu R. Delbecq,
Jun Yoneda,
Kenta Takeda,
Akito Noiri,
Arne Ludwig,
Andreas D. Wieck,
Seigo Tarucha,
Atsunori Kanemura,
Motoki Shiga,
Tomohiro Otsuka
Abstract:
Charge state recognition in quantum dot devices is important in the preparation of quantum bits for quantum information processing. Toward auto-tuning of larger-scale quantum devices, automatic charge state recognition by machine learning has been demonstrated. For further development of this technology, an understanding of the operation of the machine learning model, which is usually a black box,…
▽ More
Charge state recognition in quantum dot devices is important in the preparation of quantum bits for quantum information processing. Toward auto-tuning of larger-scale quantum devices, automatic charge state recognition by machine learning has been demonstrated. For further development of this technology, an understanding of the operation of the machine learning model, which is usually a black box, will be useful. In this study, we analyze the explainability of the machine learning model estimating charge states in quantum dots by gradient-weighted class activation map**, which identified class-discriminative regions for the predictions. The model predicts the state based on the change transition lines, indicating that human-like recognition is realized. We also demonstrate improvements of the model by utilizing feedback from the map** results. Due to the simplicity of our simulation and pre-processing methods, our approach offers scalability without significant additional simulation costs, demonstrating its suitability for future quantum dot system expansions.
△ Less
Submitted 27 December, 2023; v1 submitted 26 October, 2022;
originally announced October 2022.
