From quantum physics to practical machine learning

For decades, quantum computers represented a fascinating but practically unattainable technology. While theoretically capable of solving problems inaccessible to classical computers — such as factoring large numbers or simulating molecular interactions — their use in artificial intelligence remained more of a dream than a reality. Now, however, the situation may change.

A team led by Hsin-Yuan Huang from the quantum company Oratomic and Haimeng Zhao from the California Institute of Technology (Caltech) published a study that, according to experts, could lay the foundations for the real-world use of quantum computers in machine learning. Among the co-authors are prominent figures in quantum research — Hartmut Neven from Google Quantum AI, John Preskill from Caltech, and Jarrod McClean from Google.

Their work, published on the arXiv server under identifier 2604.07639, proves that even a relatively small quantum computer with several tens of logical qubits can process enormous datasets with performance that a classical machine requiring every atom in the observable universe could not match.

Why haven't quantum computers been used for AI until now?

The main obstacle was the so-called data loading bottleneck — the problem of loading classical data into a quantum system. Machine learning requires enormous amounts of data: restaurant reviews, genetic sequences, images, texts. For a quantum computer to work with this data effectively, it must be stored in quantum superposition — a specific state that enables parallel processing.

Until now, it was assumed that all data had to be pre-stored in specialized quantum memory. The problem is that such memory would have to be physically unrealizably large. "Researchers thought that storing data in superposition required huge memory units," explains Haimeng Zhao.

The new approach completely changes this assumption. Instead of the previous "downloading the entire movie before playing," the team proposed streaming data in small batches directly into the quantum processor. Data is not stored in advance but processed "on the fly" — similar to watching a video online.

Quantum oracle sketching and exponential advantage

The technique called quantum oracle sketching combines random sampling of classical data with quantum operations. Together with the classical shadows method, it enables creating concise classical models from enormous datasets — a task that, according to the team's proofs, is feasible for classical computers only with an exponentially larger system.

The specific numbers are remarkable. According to the authors, a quantum computer with approximately 300 logical qubits would surpass a classical computer built from every atom in the observable universe. And a computer with 60 logical qubits — which, according to Huang's estimates, is a realistic goal by the end of this decade — would achieve significant quantum advantage in machine learning tasks.

The researchers verified their theory on two real-world applications: sentiment analysis of movie reviews and processing single-cell RNA sequencing data. In both cases, they achieved a reduction in computational requirements by four to six orders of magnitude — that is, ten-thousandfold to millionfold acceleration or reduction in memory demands.

What do experts say?

Adrián Pérez-Salinas from ETH Zurich comments on the research with an analogy to feeding a powerful animal: "A quantum machine is a very powerful device, but you must first feed it. This study shows that it is enough to load data in parts, without overfeeding."

At the same time, he warns of the risk of dequantization — a phenomenon where algorithms originally designed for quantum computers were later modified to work on classical hardware with the same performance. "It will be important to examine how critical quantumness is for this new algorithm," he adds.

Vedran Dunjko from Leiden University in the Netherlands sees direct application in large scientific experiments, such as the Large Hadron Collider (LHC) at CERN, where petabytes of data are generated, most of which end up in the trash due to a lack of computational capacity. "This is not the majority of what GPUs are heating the planet for, but it can still be important," says Dunjko, with a nod to current AI-powered data centers.

European and Czech context

This development has special significance for Europe. The EU has long invested in quantum technologies through the Quantum Flagship program with a budget exceeding 1 billion euros and the EuroHPC initiative, which aims to build European quantum supercomputers. In 2023, the IBM Quantum Hub at CTU in Prague was opened in the Czech Republic, allowing Czech scientists and companies access to IBM quantum computers via the cloud.

The Czech Academy of Sciences, Charles University, and CTU are actively involved in European quantum research networks. Although the Czech Republic is not yet developing its own full-fledged quantum processor, its institutions play a significant role in theoretical research and the development of quantum algorithms — precisely the area that this new study advances.

From a regulatory perspective, the EU is preparing the AI Act, which defines rules for AI safety. Quantum computing technology is not directly regulated yet, but its potential to fundamentally accelerate AI training could raise questions about the transparency and safety of more powerful models in the future — an area where the EU wants to be a global pioneer.

When can we expect this?

The realistic horizon is the end of this decade. While 60 logical qubits sounds like a small number, it is actually an extremely demanding technical goal. Today's quantum computers have hundreds of physical qubits, but logical qubits — the ones truly resilient to errors — are still counted in single digits. Google announced the development of logical qubits in 2024, but the path to 60 functional ones is still long.

The study's authors are now working on extending their method to other types of algorithms and on designs for quantum computer configurations that could process data not only with minimal memory but also in practical time. If they succeed, machine learning could enter an era where quantum processors complement — and later perhaps even replace — today's GPU farms.