Onur Danaci and colleagues at Leiden University demonstrated performance decoupling in quantum machine learning using several qubits. Simulations and analyzes reveal that coherent quantum processing of quantum data outperforms classical processing following fixed measurements, even when considering the noise of current quantum devices. Clear advantages were demonstrated at scales of 30–40 noise qubits, revealing that the main limitation was the time required to acquire sufficient data. Their thorough evaluation of hardware constraints suggests that important learning benefits are achievable with near-term quantum technologies.
Quantum benefits emerge in machine learning with 30-40 noisy qubits
Performance separation of 30 to 40 noisy qubits is demonstrated between coherent quantum processing and fixed measurement schemes in learning problems. Unlike previous demonstrations that required noise-free conditions or focused on benchmark tasks such as quantum memory or transmitted quantum information, this result takes advantage of realistic and imperfect quantum data. This scale represents a critical threshold, as the data acquisition required for coherent processing currently exceeds the limits of classical computation, and comparable measurement-first strategies would take months or even years. The importance of this discovery lies in demonstrating quantum benefits using currently available hardware and moving beyond purely theoretical possibilities to address the practical challenges of noisy intermediate-scale quantum (NISQ) devices.
The team systematically modeled realistic hardware constraints, including errors in state preparation and gate operations, and confirmed that this advantage can be exploited in near-term quantum devices. The simulations incorporate realistic hardware limitations that model errors in state preparation, gating, and readout, factors that commonly degrade performance in real quantum devices. We used a thermal relaxation channel with amplitude 0.1 to model data degradation that represents a common noise source in qubit systems. This modeling is critical because it recognizes that current quantum computers are not perfect and practical algorithms must be robust to these imperfections. The simulations also took into account limitations in qubit connectivity, which limit interactions between qubits, and finite coherence times, which is the period during which qubits remain in a quantum state. However, significant improvements in qubit coherence and error correction are still required to extend these algorithms beyond the simulation stage, and their practicality has not yet been demonstrated. Currently, classical calculations take months or even years to achieve the same results as coherent processing, significantly increasing the time required for data acquisition. This difference highlights the potential for quantum computers to speed up machine learning tasks where data acquisition is a critical bottleneck.
Coherent quantum processing outperforms fixed measurement approaches through persistent states
Just like carefully adjusting the balance of a spinning top, maintaining quantum information over long periods of time has proven to be central to achieving this performance improvement. Coherent processing allows quantum computers to manipulate data while maintaining delicate quantum states. This is a feature that doesn’t exist in the obvious classic bits. This conservation of quantum states is achieved through the principles of superposition and entanglement, allowing qubits to represent and process much more information than classical bits. The team intentionally employed this technique to build quantum circuits that evolve quantum data without immediately breaking it down into measurable values. This is in contrast to fixed measurement schemes that take a single snapshot of a complex scene and lose information about the relationships between elements. Although fixed measurement schemes are easy to implement, they discard the rich quantum information encoded in the superposition of states, which limits the ability to extract complex patterns from the data.
The team carefully modeled realistic imperfections in quantum hardware, such as errors in state preparation and gate operations, to ensure their simulations reflected real-world limitations. This approach was advantageous for coherent processing as it allowed for more complex data evolution than fixed measurement schemes relying on a single snapshot. The ability to perform multiple operations on quantum data before measurement allows you to explore a larger solution space and identify more nuanced relationships. The simulations took into account factors such as limited qubit connectivity and coherence time, which are essential to accurately reflect real-world limitations. These limitations are not just theoretical concerns. These directly impact the fidelity of quantum computations and the accuracy of machine learning algorithms.
Advantages and limitations of scaling in short-term quantum machine learning
Quantum machine learning is poised to unlock insights from data inaccessible to classical computers, especially in fields like materials science where quantum states are fundamental. The ability to efficiently simulate and analyze quantum systems could revolutionize the discovery of new materials with tailored properties. However, this latest work reveals a subtle tension. Although coherent processing clearly outperforms fixed measurement schemes, the simulations relied on specific learning problems that exhibit an “asymptotic advantage.” This means that while the benefits of quantum grow infinitely with scale, success in all machine learning tasks is not guaranteed. Proving universal benefit remains an open question and requires exploration beyond this initial promising scenario. The specific learning problems used in this study were chosen to highlight the potential of coherent processing, and further research is needed to determine whether these results generalize to other machine learning algorithms.
Recognizing that this quantum advantage now depends on learning problems that can be advantageously extended does not diminish its importance. Demonstrating performance separation in just 30-40 qubits, even for noisy qubits, is a major step forward and shifts the bottleneck from computational power to the speed of data acquisition. This is a practical concern for any machine learning application. This discovery suggests that quantum devices of the near future will be able to tackle complex tasks that were previously out of reach. Its impact extends beyond the specific learning problems used in the simulation. It suggests that quantum computers could be used to accelerate data analysis in a wide range of fields. The challenge now lies in developing algorithms and hardware that can overcome the limitations of current quantum devices and unleash the full potential of quantum machine learning.
Coherent processing outperforms traditional fixed measurement methods in learning tasks, even when using noisy quantum data. Achieving this performance separation on a scale of 30–40 qubits indicates that an important threshold has been crossed. Classical calculations are no longer the main limitation, but rather the time required to obtain sufficient data. This discovery suggests that near-future quantum devices have viable advantages for certain machine learning applications, shifting the focus from theoretical possibilities to practical implementation. Future research will extend this advantage to a broader range of machine learning tasks and focus on developing techniques to reduce the effects of noise and improve qubit coherence, paving the way for practical quantum machine learning applications.
Researchers demonstrated the performance benefits of coherent quantum processing in learning problems using 30 to 40 noisy qubits. This result is important because it shows that at this scale, the main limitation is no longer the speed of classical calculations, but the time required to collect sufficient data. This research suggests that in the near future, quantum devices will be able to outperform classical approaches for certain machine learning tasks. The authors plan to extend this advantage to other algorithms in future work to improve qubit stability.
