Computers with increasingly efficient neural networks facilitated recent progress in machine learning and artificial intelligence (AI). For example, these computers are able to identify patterns in large and unstructured amounts of data, translate texts automatically or recognise handwriting. Furthermore, the system is capable of learning - the more data is entered, the more reliable the result is likely to be.
Quantum computers could increase computing power even further. In order to boost machine learning, researchers intend to develop a quantum mechanical version of neural networks. A team at Leibniz University Hannover has proposed a promising structure for an extremely robust and particularly flexible network, which is highly capable of learning. Their findings have now been published in the scientific journal "Nature Communications".
Their quantum neural network comprises several layers. Each layer consists of a number of quantum neurons formed by single qubits. Qubits are quantum states that can be modified and correspond to bits in conventional computers. The first layer is used to enter quantum data by manipulating qubits, while the last layer generates output. The number of intermediate layer varies. Depending on the requirements, this enables researches to generate appropriate networks: networks consisting of few intermediate layers work faster, while networks with numerous layers can handle complex tasks.
In contrast to existing quantum neural networks, which are usually limited to solving specific quantum tasks, the structure developed at the Institute of Theoretical Physics of Leibniz University Hannover is able to carry out a wide variety of tasks. Furthermore, the structure is efficient, capable of learning and tolerates internal noise that could distort the data.
In order to test the efficiency of the structure, the researchers fed training data into their quantum neural network. The outcome was remarkable: only a small amount of data was needed to achieve optimal results. Moreover, the system is able to differentiate between relevant and irrelevant data in an efficient manner. Since experiments often generate "waste data", this is a vital property. Even in cases where more than half of the data did not correspond to a pattern, the quantum neural network was still capable of learning.
However, much work still needs to be done prior to implementing the system. Within the scope of research in the field of gravitational waves, where experiments generate numerous quantum states, quantum machine learning could be used to identify relevant states.
The research was conducted within the framework of SFB 1227 "DQ-mat", which is in receipt of funding from the German Research Foundation (DFG). The collaborative research centre focuses on controlling complex quantum mechanical systems and consists of researchers in experimental and theoretical physics from Leibniz University Hannover, Center of Applied Space Technology and Microgravity (ZARM) in Bremen, and PTB Braunschweig.
Original article
Training deep quantum neural networks Kerstin Beer, Dmytro Bondarenko, Terry Farrelly, Tobias J. Osborne, Robert Salzmann, Daniel Scheiermann & Ramona Wolf Nature Communications 11: 808 (2020) DOI: doi.org/10.1038/s41467-020-14454-2
Note to editors:
For further information, please contact Kerstin Beer, Institute of Theoretical Physics, Leibniz University Hannover (Tel. +49 511 762 17505, Email kerstin.beer@itp.uni-hannover.de).
