Models Represent XOR with 100% Accuracy

Researchers Miras Seilkhan and Adilbek Taizhanov, working independently, investigated the capabilities of quantum machine learning by comparing classical and quantum approaches to solving the exclusive OR (XOR) problem.

This research is significant because it directly assesses the potential of variational quantum classifiers, machine learning models leveraging the principles of superposition and quantum computation, against established classical algorithms. They evaluated logistic regression, a multilayer perceptron, and a two-qubit variational classifier with varying circuit depths, using synthetic XOR datasets with differing levels of noise and sample sizes, measuring performance through accuracy and binary cross-entropy. Their findings demonstrate that while deeper quantum circuits can achieve comparable accuracy to classical neural networks on this benchmark, no clear advantage in robustness or computational efficiency emerges within the parameters tested. The mean absolute deviation of the hardware execution’s decision function was approximately 0.118, indicating structured deviations despite preserving the global XOR structure. A two-qubit VQC with two layers of quantum gates can achieve accuracy comparable to that of a classical multilayer perceptron on this non-linear task. Yet, simpler, shallower quantum circuits and logistic regression struggle to reliably represent the XOR problem, highlighting the importance of model expressivity. Once a problem requires nonlinear decision boundaries, as with XOR, linear classifiers are inherently limited, unable to separate data points effectively. Quantum circuits, when sufficiently deep, can represent these nonlinear functions, offering a potential pathway for quantum machine learning algorithms. At the same time, the multilayer perceptron consistently achieved lower binary cross-entropy and substantially shorter training times, even when matching the VQC’s accuracy. Now, the team evaluated performance using synthetic XOR datasets, varying parameters like Gaussian noise and sample size to assess robustness. By systematically increasing the complexity of the quantum circuit, specifically, its depth, they observed a clear correlation between depth and performance. For instance, a depth-1 circuit failed to represent XOR reliably. Meanwhile, the depth-2 circuit attained perfect test accuracy under representative conditions. Unlike the shallower circuits, the deeper VQC demonstrated an ability to learn the necessary nonlinear transformation to correctly classify the XOR data. However, no clear empirical advantage in robustness or efficiency was observed for the VQC in the examined settings. Beyond accuracy, the hardware execution of the quantum circuit introduced structured deviations in the decision function, suggesting challenges in maintaining signal integrity during quantum computation. These quantum circuits can, in principle, solve problems that classical models find challenging, but currently do not offer a clear advantage in terms of efficiency or robustness. Quantum circuit depth enables XOR function representation but lacks classical efficiency Performance metrics revealed that a two-layer deep variational quantum classifier (VQC) attained accuracy comparable to that of classical multilayer perceptrons when solving the XOR problem. Yet, simpler, shallower circuits, and logistic regression struggled to reliably represent this non-linear function, demonstrating that circuit depth is a decisive factor in the performance of VQCs. Still, despite achieving comparable accuracy, the multilayer perceptron exhibited lower binary cross-entropy and substantially shorter training times, indicating that while quantum circuits can represent the XOR function with sufficient depth, they do not currently offer a practical advantage in terms of computational efficiency. The mean absolute deviation of the decision function was approximately 0.118, signifying structured deviations in the output despite the preservation of the global XOR structure, suggesting that noise introduces systematic errors. By examining robustness across varying noise levels, dataset sizes, and random seeds, researchers confirmed that circuit depth remains critical for successful performance on this task. For instance, a depth-1 circuit consistently failed to learn the XOR function, while the depth-2 circuit, alongside the multilayer perceptron, achieved perfect test accuracy under representative conditions. Through increasing the depth of the circuit introduces optimisation challenges, potentially leading to barren plateaus where gradient-based training becomes difficult. This effort was limited to a low-dimensional XOR benchmark. No clear empirical advantage in robustness or efficiency was observed for the VQC compared to the multilayer perceptron — unlike more complex datasets, the XOR problem is relatively simple. In turn, the observed performance may not generalise to more realistic scenarios, and this effort builds upon existing research in quantum machine learning, specifically variational quantum classifiers. Provides a valuable comparison against established classical methods like logistic regression and multilayer perceptrons. Demonstrating non-linear function representation via quantum circuit depth Still, the persistent challenge of representing non-linear relationships within machine learning models has long demanded exploration beyond conventional architectures. It is important to acknowledge that this demonstration does not immediately translate into a practical quantum advantage. While the two-layer circuit performed comparably, shallower circuits and even simple logistic regression struggled with the task, highlighting the importance of circuit depth. Even so, the multilayer perceptron achieved this performance with substantially faster training times. At present, the computational cost of operating these quantum circuits outweighs any potential benefit, a limitation shared by many early quantum machine learning experiments. This project builds upon a growing body of work focused on variational quantum classifiers, a hybrid quantum-classical approach favoured by groups at Google Quantum AI and IBM Quantum. Unlike earlier attempts that focused on theoretical expressivity, this effort employed a real two-qubit superconducting processor, introducing the complexities of hardware noise and limitations. On that front, the hardware execution did reveal structured deviations in the decision function, despite preserving the overall XOR structure, a reminder that real-world quantum devices are not perfect simulators. Future investigations must move beyond low-dimensional benchmarks like XOR and address more complex, high-dimensional datasets. A critical next step involves demonstrating genuine improvements in either robustness or efficiency. Showing that these quantum models can outperform their classical counterparts in a meaningful way. Until then, this effort remains a confirmation of potential rather than a harbinger of a quantum revolution in machine learning. 👉 More information 🗞 Comparing Classical and Quantum Variational Classifiers on the XOR Problem 🧠 ArXiv: https://arxiv.org/abs/2602.24220 Tags: