Quantum computers could run artificially intelligent algorithms that initially appear to require too much computational power thanks to a new method for writing quantum programs.
While quantum computers have the potential to outperform conventional computers for many different types of calculations, so far they have only done so for a few specialised computations. That is because those that currently exist are relatively small and not very powerful.
at Leiden University in the Netherlands and his colleagues wanted to see whether even these imperfect computers could be used for quantum machine learning, the quantum version of the process commonly used for training AIs.
Advertisement
Researchers program quantum computers by writing lists of operations called quantum circuits. Complicated programs, like those used for machine learning, require large quantum circuits, but many of them are too large for existing quantum computers. So Marshall and his colleagues developed a method for making these big circuits smaller.
Their approach was to divide large quantum circuits into smaller sub-circuits, then determine, through mathematical calculations, which of these could be ignored without incurring too many errors. The researchers performed tests of this approach by running two quantum machine learning algorithms on a simulation of a quantum computer.
In the first, they developed an AI that could recognise handwritten digits using just eight simulated qubits. Normally, a comparable program would need a 64-qubit machine.
And in the second test, the researchers used a five-qubit AI to predict the outcome of a quantum process, which would normally take 10 qubits to perform. Though the difference in size was less stark here, this second task was specifically designed to break the new method yet didn鈥檛, says also at Leiden University.
The team presented the work on 7 February at the聽in Ghent, Belgium.
Though promising, this method may not work for every complex quantum circuit, says at Terra Quantum, a quantum computing company in Switzerland. However, he says that even without access to quantum hardware, the new method could improve how quantum algorithms are run on conventional supercomputers, which are already used for commercial purposes like planning navigation routes for vehicles.
鈥淯ltimately, we don鈥檛 yet know where combining quantum computers and machine learning will be most useful, but I think we can鈥檛 just wait for bigger quantum computers. Once they do get larger there will still be this question of how far we can push them with algorithms,鈥� says Dunjko. The team is also planning on testing the method on a physical quantum computer.
