State preparation often suffers from various sources of device-dependant noise. It is hard to characterise and let alone mitigate the noise. Machine learning can be used to automate the denoising task. Quantum computers are the ultimate machines to process quantum signals. As such, we propose to use machine learning on quantum computers to denoise quantum data.
For this, we design a flexible family of trainable variational circuits that are capable of implementing arbitrary quantum channels. Due to the resemblance to their classical counterparts, we call these circuits quantum neutral networks (QNNs). We show that QNNs use the data efficiently and can be trained on imperfect devices even when part of the training data is corrupted.
We proceed to design QNN architectures that are tailored to cancel various types of noise. We show how quantum autoencoders can significantly reduce shot-to-shot noise. This is achieved without access to supervisory data and autoencoders can be trained once to denoise different highly entangled states. Then we turn our attention to extrapolation of parameters like temperature. We show how to train QNNs to produce states that are colder than any state seen during training. Finally, we develop recurrent QNNs for processing that requires memory. We use recurrent QNNs to design quantum bandwidth filters---a task greatly complicated by the no-cloning theorem. For recurrent networks, we further economize resources by combining quantum and classical processing.
Quantum denoisers offer a practical application of quantum computers to the experimental preparation of entangled states. This technology is particularly suited to an apparatus that already carries a highly-tunable quantum control system on board, such as quantum logic.
Based on: Nat. Commun. 11, 808; Phys. Rev. Lett. 124, 130502; PhD thesis: https://doi.org/10.15488/11050 and some work in progress.
|Presenter name||Dmytro Bondarenko|
|How will you attend ICAP-27?||I am planning on in-person attendance|