'Quantum AI' algorithms already outpace the fastest supercomputers, study says

Scientists say they have made a breakthrough after developing a quantum computing technique to run machine learning algorithms that outperform state-of-the-art classical computers.

The researchers revealed their findings in a study published June 2 in the journal Nature Photonics.

The scientists used a method that relies on a quantum photonic circuit and a bespoke machine learning algorithm.

Using only two photons, the team's technique successfully demonstrated increased speed, accuracy and efficiency over standard classical computing methods for running machine learning algorithms.

The scientists say this is one of the first times quantum machine learning has been used for real-world problems and provides benefits that cannot be simulated using binary computers. Furthermore, due to its novel architecture, it could be applied to quantum computing systems featuring only a single qubit, they said.

Unlike many existing methods for achieving speedup through hybrid quantum-classical computing techniques, this new method doesn't require entangled gates. Instead, it relies on photon injection.

Related: 'The science is solved': IBM to build monster 10,000-qubit quantum computer by 2029

Essentially, the team used a femtosecond laser — a laser that emits light in extremely short pulses measured in femtoseconds (10⁻¹⁵ seconds) to write on a borosilicate glass substrate to classify data points from a dataset. The photons were then injected in six distinct configurations, which were processed by a hybrid quantum-binary system.

The scientists determined where the photonic measurements outperformed those conducted via classical computing by measuring how long it took the photons to complete the quantum circuit. They then isolated the processes where quantum processing provided benefit and compared the results to the classical outputs.

The researchers found that experiments run using the photonic quantum circuit were faster, more accurate and more energy-efficient than those conducted using only classical computing techniques. This boosted performance applies to a special class of machine learning called "kernel-based machine learning" that can have myriad applications across data sorting.

While deep neural networks have become an increasingly popular alternative to kernel methods for machine learning over the past decade, kernel-based systems have seen a resurgence in the past few years due to their relative simplicity and advantages when working with small datasets.

The team's experiment could lead to more efficient algorithms in the fields of natural language processing and other supervised learning models.

Perhaps most importantly, the study showcases a novel method for identifying tasks that quantum computers excel at in hybrid computer systems.

The researchers say the techniques used are scalable, meaning they could lead to even better performance as the number of photons or qubits increases. This could, in turn, make it possible to develop machine learning systems capable of exceeding the limits of today's models, which increasingly face power consumption limitations due to the massive energy requirements needed to process data via electronics.

The researchers claim their techniques will "open the door to hybrid methods in which photonic processors are used to enhance the performance of standard machine learning methods."