Artificial intelligence may soon get a power boost from quantum computing. Researchers have shown that quantum computers could sort and classify complex data sets that algorithms running on classical computers struggle to handle.
“We’ve shown a particular way of moving data into higher dimensions that would be hard for a classical computer but manageable for a quantum computer,” says Antonio Córcoles at IBM in New York.
Classification problems are key to machine learning – for example, given sufficient data AI can learn on its own how to classify whether a pixel in a mammogram is indicative of a cancerous tumour or healthy tissue.
Advertisement
Category mapping
The simplest way to classify data is to map it and separate it into two categories. Then, based on that training, an algorithm can tell which category any new data point fits into. Optimising where that line between the two categories cuts through the data – making sure it is as far away from each of the two closest data points – reduces possible classification errors.
“You have to bring that data to a higher dimensional space until you find a dimension from which you can cleanly cut it. You can always do that, but for complicated data, that dimension is very high,” says Córcoles – so high, that a classical computer would take a very long time to handle the problem, and in some cases it would take so long that it would be practically impossible with the machines we use today.
That is where quantum computers comes in handy. Because qubits, their basic units of information, can inhabit superpositions of two states, they can do this classification process more efficiently for data sets with more variables.
This new work uses an algorithm running on a quantum circuit made of two superconducting qubits. Córcoles and his colleagues ran the algorithm on three data sets, each containing 20 data points to be classified into two categories. The first two sets were classified correctly 100 per cent of the time, while the third had a success rate of 94.75 per cent.
Córcoles says the lower success rate on the third data set is due to noise in the calculations created by small changes in the quantum circuit’s calibration. Qubits, or quantum bits, are famously fussier than their classical counterparts. To maintain their fragile superpositions – in which they don’t just read out 1 or 0 but can exist in infinite combinations of the two – quantum circuits must be shielded from noise in the form of heat and vibrations that can disrupt a qubit. But in practice, it is difficult at the moment to provide fully effective noise shielding.
“This work puts quantum and classical machine learning on the same playing field for the first time,” says Ciarán Lee at University College London. “Although these results are very promising, they are only a first step to machine learning on small quantum computers. They only tested their approach on hand-crafted data designed to be hard for an ordinary computer. Applying their methods to messy real-world data will be a significant challenge.”
Córcoles agrees, but says this proof of concept shows that even with small quantum circuits that are plagued by noise, we can still solve problems that are difficult for classical computers. “We are really far away from having noiseless quantum processors,” he says. But in the current era of where we have to live with that noise, it is still possible to classify data with little to no error, he says.
Nature
