Universiteit Leiden

nl en

PNAS Paper Prize for quantum machine learning

‘We hope our paper highlights the possibilities and benefits of including artificial intelligence in quantum physics to do new discoveries.’ Vedran Dunjko of the Leiden Institute of Advanced Computer Science contributed to a paper that was published in PNAS last year and now received a Cozzarelli Prize for outstanding papers.

The paper of Dunjko and his colleagues forms a rather unique bridge between artificial intelligence and quantum physics. They showed that a reinforcement learning system can be used to design new quantum experiments.

Looking for new quantum states

In reinforcement learning, a system teaches itself by trial and error. It performs actions with the aim of maximizing rewards: learning by doing in order to achieve the best outcomes. ‘In this case, we tasked the system to design quantum optical experiments which would generate complex quantum states’, Dunjko tells. ‘Designing such experiments seems difficult for experimentalists, in part because our intuition about quantum systems is often wrong’.

Therefore he and his colleagues delegated this task to an automated system. ‘We were, in particular, looking for states with high degrees of entanglement across multiple systems and dimensions’, he tells. This entanglement occurs when particles interact in such a way that they cannot be described independently of the state of the others anymore. It is a quintessential feature of quantum mechanics. ‘It’s at the heart of quantum computing and quantum information processing, where systems are correlated in a stronger way than is classically possible.’

Quantum states

The state of a quantum system literally is the situation it is in at any given moment. A physical description of that state provides as complete information as possible about the system at that moment.

Applied entanglement

Entanglement makes it possible to correlate systems in a much stronger way than was classically possible. This is done by creating entanglement between different stations, which are set increasingly further away from each other. Because entangled particles interact so strongly, one could ultimately teleport the quantum information from one end of the chain to the other, without any disturbance and without the signal fading.

Unraveling quantum physics

When they run the system, something interesting happened. ‘In the process of learning how to design various types of quantum states, we noticed that our system started preferentially using certain fixed blocks of basic experimental elements. In other words, it identified simple “gadgets” which were useful in a spectrum of differing experiments. We analyzed these gadgets and found that they correspond to different, and interesting implementations of a previously known quantum device – a “parity sorter”-- which had been designed by researchers in 2002. This device was shown to be useful to generate entangled states. We were surprised to notice that our system had, in a manner of speaking, reinvented it.’

Even though the system did not invent a completely new gadget, Vedran believes this finding highlights the substantial possibilities and benefits of including advanced artificial intelligence machinery to unravel new mysteries in quantum physics. And that it shows that such systems could play an integral role in genuine research.

Bridging the gap

The Cozzarelli prize is awarded each year to most outstanding and original publications in PNAS of the year before. Dunjko was delighted when he found out about the prize, which he believes was awarded to them because of the multidisciplinary perspective they provided to their findings. ‘We hope this will bring quantum physics, artificial intelligence and machine learning closer together.’ They already saw that their work inspired some follow-up ideas for bridging those fields. ‘Despite the fact that we used relatively simple techniques, we felt our paper was quite fun and innovative. I for one am very happy the Editor and the Associate Editors (who have selected this work) agree with us.’

This website uses cookies.  More information.