Although in its preliminary stages, IBM's quantum computing team - in partnership with Oxford university and MIT - have developed and tested a quantum algorithm that they say could enable machine learning on quantum machines in the "near future".
The research paves the way for quantum machines to 'feature map' complex sets of data - that is, breaking down data into core identifiers, for example separating all of the pixels from a single image, and then placing them in a new grid based on characteristics such as by colour.
Where a quantum machine would differ from a 'regular' machine learning algorithm is that it would be able to take enormous data sets full of complex information and spot patterns that might be invisible to a classical computer.
The paper, titled: 'Supervised learning with quantum enhanced feature spaces', has been published in peer-reviewed journal Nature, and sets out how the researchers put together a "blueprint with new quantum data classification algorithms and feature maps".
IBM said of the research: "That's important for AI because the larger and more diverse a data set is, the more difficult it is to separate that data out into meaningful classes for training a machine learning algorithm. Bad classification results from the machine learning process could introduce undesirable results; for example, impairing a medical device's ability to identify cancer cells based on mammography data."
Antonio D. Córcoles, who co-authored the paper and works in the experimental team at IBM, told Computerworld UK that the main idea wasn't to simply copy the techniques well-established in classical machine learning and shift them over to quantum, but to find a distinction in the algorithm that provides the quantum method with an advantage.
Today, says Córcoles, quantum machines aren't large enough or noiseless enough to tackle the problems in a bigger way, but the paper, and the team's work, demonstrates that a quantum approach is possible, laying the foundations for future work.
"At this stage we are not going to use this algorithm to try to tackle real world data," says Córcoles, adding that quantum enthusiasts or researchers could try it if they liked, because their findings will be rolled in to open source quantum framework, Qiskit.
"But that was not our goal," he says. "Our goal was to say: we are going to prove that there are feature maps you can use that give you an advantage. Now that we have proven that, the next stage of research will be, can we understand these quantum feature maps better, to gain insight into how we can apply them to real-world problems?"
Quantum advantage is, loosely defined, when using a quantum methodology brings about better results than those with traditional computers.
IBM physicist Kristan Temme told The Next Web that this experiment was designed to work with the noisy, imperfect quantum systems of today.
There are clear implications for the experiment in encouraging further research in the field, says Córcoles, as well as providing educational use for other researchers and enthusiasts.
"I think it's very important to test these ideas in a real setting rather than just having them expressed theoretically," says Córcoles, adding that sometimes the theory translated into a lab setting brings up new, unexpected problems, complicating matters.
"So having systems that you can actually use with these ideas is very important for us," he says. "Now, for the rest of the world - we think this is interesting to the artificial intelligence community in general, because this gives a path that is very preliminary, that is very exploratory still, but is also very hopeful because as far as we know in complexity theory, this is an important result.
"For all we know this algorithm will scale well for quantum, but it will scale extremely poorly for classical [computing]. So now people will have something that they can try to map the problems into, and to see if they can get some advantage from using this."
"They can try to see or gain some insight for the level of noise that this algorithm will tolerate, by trying to run it in more than two qubits - which is what we did in the paper - and will see what kind of success they will get," says Córcoles.
Potential next steps are also in the early stages but Córcoles says that the feature map could become more generalised, and they are working on a way to find feature maps that are similar to the one demonstrated but can be tuned to specific data sets.
The team's work will be available in Qiskit Aqua, a library of cross-domain quantum algorithms, allowing researchers to run it in their simulators or in the actual quantum devices. Córcoles adds that he used parts of Qiskit to actually run the experiments.
"Qiskit is very powerful at compiling your circuits into something that is more efficient than you initially wrote," he says. "That was useful for me as a user in the lab."