For more than a decade, scientists have described neural networks that could run on quantum computers. The promise was clear: quantum bits can exist in combinations of states and can share linked relationships that ordinary computers cannot copy easily. Those features might let a quantum network solve certain pattern problems more effectively than any classical system. Until recently almost every discussion of the idea stayed on paper or on ordinary computers that only simulated quantum behavior. In the paper “Benchmarking a Tunable Quantum Neural Network on Trapped-Ion and Superconducting Hardware,” Djamil Lakhdar-Hamina, Xingxin Liu, Richard Barney, Sarah H. Miller, Alaina M. Green, Norbert M. Linke, and Victor Galitski report that they have moved the idea into the laboratory. They built a neural network, trained it in the ordinary way, and then successfully ran the decision-making stage on two different kinds of quantum hardware: a trapped-ion system and IBM’s superconducting processors.[1]
A neural network is a computer method that learns to recognize patterns by dividing a problem into many small decisions. Picture a large office building with clerks on every floor. On the ground floor the clerks examine the raw material, for example the dark and light spots that form a handwritten number. Each clerk notices one small detail and sends a simple yes-or-no message to the floor above. On the next floor other clerks collect those messages, decide which ones matter most, and send their own yes-or-no messages higher still. At the top floor a final clerk announces the answer; this scribble is most likely a 4, or a 9, and so on. The strength of the connections between the clerks is what the network learns through experience. Networks of this kind already let computers recognize faces in photographs, translate spoken language, recommend products, and predict the folded shapes of proteins from their chemical sequences. All that progress, however, has taken place on ordinary classical computers.
Teaching the network is a process of repeated practice and correction. The system shows thousands of examples whose correct answers are already known. After each example it compares its own answer with the true one and slightly strengthens or weakens every connection so that the next similar example will be closer to correct. The same collection of practice examples is shown again. After every complete pass through the practice collection the network is checked on a fresh batch of examples it has never seen. At first its score on the fresh batch rises. After enough passes the score stops rising and simply stays roughly the same. Training then ends, because further practice brings little additional improvement.
Quantum computers replace ordinary yes-or-no messages with qubits. A qubit can sit in a mixture of yes and no at the same time until it is measured. When the measurement occurs the qubit lands on one answer or the other by chance, according to probabilities set by the operations that came before. Qubits can also become linked so that the state of one instantly affects the others. Researchers have hoped that this built-in chance and these linked relationships might let a quantum network explore several possible answers at once or handle messy data more gracefully than a classical network. Turning that hope into working systems has been difficult. Present-day quantum machines still make small errors on almost every operation, measurements are never perfect, and many proposed designs require the results of early measurements to control later steps. That kind of mid-stream feedback remains hard to perform reliably. Because of these obstacles, genuine tests of quantum neural networks on actual quantum hardware have been rare.
The researchers designed a network that can be adjusted continuously between ordinary and quantum behavior. Each yes-or-no decision is carried out by a qubit that is tilted by a controlled angle and then measured. A single setting determines how large those tilts are. When the setting is at its lowest value the tilts are forced to extremes, the measurements become completely predictable, and the network behaves exactly like a classical network. When the setting is raised the tilts become intermediate, the measurements grow uncertain, and the same input can produce different answers on successive runs. All the connection strengths were trained on an ordinary computer using the standard collection of handwritten digits known as MNIST. Only the final step of reading new digits was performed on the quantum machines. Because the network can be executed one qubit at a time, the entire system could be run on as little as a single physical qubit. That low demand made the experiment possible on quantum computers available today.
The researchers carried out the reading step on three platforms: a trapped-ion computer controlled by microwaves, the same ion computer controlled by lasers, and IBM’s superconducting circuits. On a random sample of fifty-five test digits the network’s accuracy rose above the pure classical level once the quantum setting reached moderate values, then declined when the setting was raised so high that the answers became nearly random. The improvement matched earlier computer simulations, but on the real machines the accuracy at low settings was sometimes a little higher still. The extra improvement came from the machines’ own small physical errors.
To understand that phenomenon the team examined digits that the classical network always misread. For one such borderline digit a perfect simulation of the quantum network also failed completely when the quantum setting was at zero. On the real superconducting machine, the same digit came out correct half the time; on the microwave ion machine it came out correct one time in ten. The researchers picture the trained network as having two nearby possible resting places for its final answer, one correct and one incorrect. The machines’ natural errors occasionally push the system from the wrong resting place into the right one. Clear digits that both the classical and quantum networks handle correctly showed no such sensitivity, as if those digits rested in a single stable place.
The team further tested the effect of extra error by deliberately inserting pairs of operations that cancel each other in a perfect machine. In a real machine those pairs simply inject additional noise. On the trapped-ion platform even a modest number of extra single-qubit pairs produced a clear rise in accuracy for the borderline digit, while the same pairs left clear digits largely unchanged. Extra pairs of two-qubit operations first rose then lowered performance. The superconducting machines responded differently. Their native errors already appeared sufficient to help the borderline cases, so further added error produced little gain and eventually reduced accuracy. The number of extra pairs needed to reduce the network to pure guessing thereby gave a practical way to compare the quality of the two hardware platforms on the identical task.
Because the network was executed on real quantum machines, these effects of physical error became visible for the first time. The experiment therefore accomplishes more than a simple demonstration that a quantum neural network can function on present-day hardware. It also supplies the first direct side-by-side comparison of such a network across trapped-ion and superconducting technologies for a standard image-classification task. That comparison is valuable because the two platforms have different sources of error and different strengths. Seeing how the same network behaves on both platforms reveals which features of the hardware matter most for this kind of learning system.
The authors point out that the network they tested is still simple enough for ordinary computers to simulate completely, because every qubit is measured before the next layer begins. To reach problems that ordinary computers cannot handle efficiently, future networks will need to measure only some of the qubits in each layer while leaving others linked together, then use those partial measurement results to control the operations of the next layer. Such networks would combine quantum learning with the study of how measurement itself can change the behavior of large collections of qubits. Training them may require new methods that search for good designs by trial and reward. The researchers also suggest that keeping a small, controlled amount of the machines’ natural error during both training and reading could systematically improve performance on noisy data that appear in real applications.
The work shows that quantum neural networks can be run on the quantum computers that exist today. That demonstration gives engineers and scientists a concrete foundation for deciding which hardware platforms and which network designs are worth developing further as the machines grow larger and more stable. Fields that routinely encounter ambiguous data, such as medical imaging, satellite reconnaissance, and the reading of handwritten forms, could benefit from a second opinion guided by the controlled chance that quantum hardware provides.
The network tested in this study remains modest in size, but the fact that it ran successfully on two different quantum platforms marks the moment when quantum neural networks moved from theoretical discussion into experimental reality. The next steps will be to enlarge the networks, to incorporate partial measurements and feedback, and to test whether the same principles continue to hold as the systems become more complex. These steps are now possible because the first working examples have been shown to function on real quantum hardware.
This article is shared at no charge for educational and informational purposes only.
Red Sky Alliance is a Cyber Threat Analysis and Intelligence Service organization. We provide indicators of compromise information (CTI) via a notification/Tier I analysis service (RedXray) or an analysis service (CTAC). For questions, comments or assistance, please contact the office directly at 1-844-492-7225, or feedback@redskyalliance.com
Weekly Cyber Intelligence Briefings:
- Reporting: https://www.redskyalliance.org/
- Website: https://www.redskyalliance.com/
- LinkedIn: https://www.linkedin.com/company/64265941
Weekly Cyber Intelligence Briefings:
REDSHORTS - Weekly Cyber Intelligence Briefings
https://attendee.gotowebinar.com/register/7855487668891299929
[1] https://six3ro.substack.com/p/quantum-neural-networks-finally-run
Comments