摘要:Comparing the properties of qubits in different quantum computers is challenging, but AI can tell them apart even when the state of the qubits is 98 per cent similar
AI can tell how similar two quantum states are, even using imperfect data about their properties. This could make it a useful tool for assessing how different quantum computers stack up against one another in the future.
Quantum computers process information by manipulating the states of their quantum bits, or qubits. But comparing the properties of qubits in two different quantum computers – or comparing the output of one quantum device with what it should ideally be – is challenging because of small errors and “noise” in the data.
Yan Zhu at the University of Hong Kong and his colleagues trained a machine learning algorithm to compare two quantum states. They used experimental data gathered from quantum computers as well as data produced by simulating small quantum computers on conventional computers.
Some of this training data was purposefully imperfect, analogous to training facial recognition AI on blurry or overexposed images. Zhu says that all existing quantum computers are subject to noise and sometimes make errors, so for an AI to be useful in labs in the future, it has to learn how to deal with those imperfections.
Read more: Mysterious sounds in stratosphere can't be traced to any known source
The researchers then put the AI to the test. They had it compare two quantum states called “cat states”, in which special qubits are in a superposition of two states at once, similar to the cat in Erwin Schrödinger’s famous thought experiment. Quantum devices that use particles of light, or photons, as qubits often work best when photons are in their cat states. Given two complicated cat states, as large as 225 or 256 photons each, the AI could tell them apart even when they were 98 per cent the same.
Giulio Chiribella at the University of Hong Kong says that this success is striking because his team never provided the AI with built-in knowledge of quantum physics; they didn’t train it on equations it should use to represent each state. But in comparing the two states, the AI found its own way to represent the complex cat states. “The neural network makes its own sort of mini quantum theory of the states that it sees from experiments. It doesn’t represent the states with the same mathematics that we do,” he says.
An active effort in the field of quantum computing is to build a device that outperforms all conventional computers – and traditional mathematical methods won’t be sufficient to characterise the states of any future quantum computer large enough and powerful enough to do that, says Cristian Bonato at Heriot-Watt University in the UK. “Quantum states are hard to describe because their complexity grows exponentially with their size, but neural networks are good at finding patterns among huge numbers of parameters. They could become very valuable for assessing quantum computers in the future,” he says.
Journal reference: Physical Review Letters, forthcoming