A new photonic quantum neural network from the University of Cambridge, in collaboration with Leonardo UK Ltd, has achieved 100% classification accuracy in both online and offline learning tasks. Solomon McKiernan and Luca Sapienza report the successful implementation and training of a variational quantum classifier using single photons and stochastic gates. Photonic quantum neural networks outperformed comparable classical artificial neural networks, solving problems that required at least four times as many parameters in classical models. These findings demonstrate the potential of current photonic hardware to enable and effectively train gate-based quantum neural networks, an important step toward practical quantum machine learning applications.
Photonic quantum neural networks surpass classical performance on nonlinear classification tasks
A photonic quantum neural network (QNN) achieved 100% accuracy on a nonlinear separable task, while an equivalent artificial neural network (ANN) failed to learn the same problem. Classical models require at least four times as many trainable parameters to solve similarly complex problems, but a proven QNN achieved this with just two. This success represents a proof of principle for realizing and training gate-based QNNs on photonic hardware and suggests the benefits of the algorithm in certain machine learning applications.
Calculating the “effective dimensions” of these networks (a measure of their ability to learn complex patterns) reveals that photonic QNNs outperform their traditional counterparts with a comparable number of tunable parameters. Further benchmarks using the Iris dataset show that the photonic QNN achieves up to 100% accuracy in both online and offline learning scenarios, demonstrating its adaptability beyond simple synthetic datasets. Successful deployment of a circuit with relatively powerful effective dimensions on a 6-qubit photonic quantum processor validates the scalability of the approach despite current limitations on small datasets and limited circuit depth.
Photonic quantum neural networks outperform classical models on benchmark datasets
A gate-based photonic quantum neural network achieved 100% accuracy on a nonlinearly separable task, outperforming an equivalent classical artificial neural network that failed to converge. This success was observed using the XOR problem and a subset of the Iris dataset. For practical implementation, it turns out to be important to evaluate the robustness to realistic noise processes, especially photon losses and phase shifter imperfections.
Photonic quantum neural networks surpass classical performance in nonlinear classification
A gate-based photonic quantum neural network (QNN) achieved 100% accuracy on nonlinear separable tasks. This is a feat that comparable classical artificial neural networks (ANNs) have not been able to achieve. They utilized single photons and stochastic gates to emulate standard quantum circuit modeling frameworks, paving the way for more efficient machine learning. Assessing the expressive power of QNNs by calculating their effective dimensionality, a measure related to generalization error, enabled a contrast with classical ANNs with a comparable number of trainable parameters. We used a supervised binary classification task to benchmark the performance between photonic and superconducting QNNs and found that both quantum implementations exhibited lower cross-entropy loss and higher prediction accuracy than their classical counterparts.
A photonic QNN trained using gradient-free optimization successfully converged on the XOR problem and a subset of the Iris dataset while remaining robust to realistic noise sources such as photon loss and phase shifter imperfections. By implementing the circuit on a 6-qubit photonic quantum processor, high classification accuracy was achieved in both online and offline learning scenarios. The simplest unrolled QNN, containing only two trainable parameters, performs better than an ANN that requires at least four times as many parameters to solve the same problem, suggesting an advantage of the algorithm.
A previous study by Abbas et al. demonstrated enhanced effective dimensionality and faster training in a numerical study of a similar QNN model, providing the basis for this work. Future work will focus on investigating the limitations of this approach and expanding the range of problems that can be solved with gate-based photonic QNNs, in parallel with investigating how to reduce the effects of noise and improve the overall performance of these quantum classifiers. Significantly, we validated the potential of optical quantum computation for specific tasks and achieved 100% accuracy on difficult nonlinear separable problems where traditional computers have failed. The concept of “effective dimensionality,” which quantifies a network’s learning ability, turns out to be key to demonstrating this advantage over classical artificial neural networks. As a result, this study shifts the focus from theoretical possibilities to practical realizations and invites investigation into the extension of these networks and consideration of their application to more complex real-world datasets.
Researchers demonstrated the superior performance of quantum neural networks in classifying data, achieving 100% accuracy on difficult problems that comparable classical networks could not learn. This is important because it suggests the potential algorithmic benefits of quantum computation in machine learning tasks, even when the number of trainable parameters is small. Using a single photon and a six-qubit processor, they showed that these networks are robust to realistic noise and maintain lower error rates than classical artificial neural networks. The authors plan to extend this approach to solve a wider range of problems and further improve performance.
