Researchers Miras Seilkhan and Adilbek Taizhanov worked independently to explore the capabilities of quantum machine learning by comparing classical and quantum approaches to solving exclusive-or (XOR) problems. This study is important because it directly assesses the potential of variational quantum classifiers, machine learning models that leverage principles of superposition and quantum computation, against established classical algorithms. They evaluated logistic regression, multilayer perceptron, and two-qubit variational classifiers at different circuit depths, using synthetic XOR datasets with different levels of noise and sample sizes, measuring performance through accuracy and binary cross entropy. Their findings show that on this benchmark, deeper quantum circuits can achieve comparable accuracy to classical neural networks, but do not emerge with clear advantages in robustness or computational efficiency within the parameters tested.
The average absolute deviation of the hardware-implemented decision function is approximately 0.118, indicating a structured deviation despite preserving the global XOR structure. A two-qubit VQC with two layers of quantum gates can achieve comparable accuracy to a classical multilayer perceptron in this nonlinear task.
However, simpler, shallower quantum circuits and logistic regression struggle to reliably represent the XOR problem, highlighting the importance of model expressiveness. When a problem requires a nonlinear decision boundary, as with XOR, linear classifiers become inherently limited and cannot effectively separate data points. Quantum circuits, if deep enough, can represent these nonlinear functions, providing a potential route for quantum machine learning algorithms.
At the same time, the multilayer perceptron consistently achieved lower binary cross-entropy and significantly shorter training times, even when matching the accuracy of VQC. Currently, the team evaluated performance using a synthetic XOR dataset and varied parameters such as Gaussian noise and sample size to assess robustness. By systematically increasing the complexity of a quantum circuit, specifically its depth, they observed a clear correlation between depth and performance.
For example, a circuit with a depth of 1 could not reliably represent an XOR. Meanwhile, the depth 2 circuit achieved perfect test accuracy under representative conditions. Unlike shallow circuits, deeper VQCs demonstrated the ability to learn the nonlinear transformations required to correctly classify XOR data. However, no clear empirical advantage in robustness or efficiency of VQC was observed in the investigated settings.
Beyond accuracy, the hardware implementation of quantum circuits introduced structural deviations in the decision function, suggesting challenges in maintaining signal integrity during quantum computation. Although these quantum circuits can, in principle, solve problems that would be considered difficult with classical models, they currently offer no clear advantage in terms of efficiency or robustness.
The depth of quantum circuits allows for XOR function representation, but it lacks classical efficiency
Performance metrics revealed that a two-layer deep variational quantum classifier (VQC) achieved an accuracy comparable to that of a classical multilayer perceptron in solving the XOR problem. However, simpler, shallower circuits and logistic regression struggle to reliably represent this nonlinear function, showing that circuit depth is a critical factor in VQC performance.
Yet, despite achieving comparable accuracy, multilayer perceptrons exhibited lower binary cross entropy and significantly shorter training times. This shows that while quantum circuits can represent XOR functions of sufficient depth, they currently do not offer any practical advantage in terms of computational efficiency. The average absolute deviation of the decision function is approximately 0.118, indicating that although the global XOR structure is maintained, there are structural deviations in the output, suggesting that noise is introducing systematic errors.
By investigating its robustness to different noise levels, dataset sizes, and random seeds, the researchers confirmed that circuit depth remains important for successful performance on this task. For example, a depth 1 circuit consistently failed to learn the XOR function, whereas a depth 2 circuit achieved perfect test accuracy under representative conditions in parallel with a multilayer perceptron.
Increasing circuit depth poses optimization challenges and can lead to a barren plateau that makes gradient-based training difficult. This effort was limited to low-dimensional XOR benchmarks. We did not observe any clear empirical advantage in robustness or efficiency of VQC compared to multilayer perceptrons. Unlike more complex datasets, the XOR problem is relatively simple. Also, the observed performance may not generalize to more realistic scenarios, and this effort builds on existing research in quantum machine learning, specifically variational quantum classifiers. It provides a valuable comparison with established classical methods such as logistic regression and multilayer perceptron.
Demonstration of nonlinear function expression using quantum circuit depth
Yet, the persistent challenge of representing nonlinear relationships within machine learning models has long required exploration beyond traditional architectures. It is important to recognize that this demonstration does not immediately translate into real quantum benefits. Two-layer circuits showed comparable performance, but shallower circuits and even simple logistic regression struggled with this task, highlighting the importance of circuit depth. Yet, the multilayer perceptron achieved this performance with significantly faster training times.
Currently, the computational cost of operating these quantum circuits exceeds the potential benefits, a common limitation of many early quantum machine learning experiments. The project builds on a growing body of research focused on variational quantum classifiers, a hybrid quantum-classical approach favored by groups at Google Quantum AI and IBM Quantum.
Unlike previous efforts that focused on theoretical expressivity, this effort employed an actual two-qubit superconducting processor, introducing the complexity of hardware noise and limitations. To that extent, the hardware implementation revealed structural deviations in the decision function, even though the overall XOR structure was maintained. This is a reminder that real-world quantum devices are not perfect simulators.
Future research should move beyond low-dimensional benchmarks like XOR and address more complex high-dimensional datasets. Critical next steps include demonstrating true improvements in robustness or efficiency. We show that these quantum models can outperform classical quantum models in meaningful ways. Until then, this effort remains a confirmation of possibility rather than a harbinger of a quantum revolution in machine learning.
