Quantum Networks Demonstrate Improved Accuracy in Classification Tasks by 4%

A new photonic quantum neural network at University of Cambridge, in collaboration with Leonardo UK Ltd, achieves 100% classification accuracy on both online and offline learning tasks. Solomon McKiernan and Luca Sapienza report the successful implementation and training of variational quantum classifiers using single photons and probabilistic gates. The photonic quantum neural networks outperformed equivalent classical artificial neural networks, solving problems that required at least four times more parameters in the classical models. These findings demonstrate the potential of current photonic hardware to realise and effectively train gate-based quantum neural networks, marking a key step towards practical quantum machine learning applications.

Photonic quantum neural network surpasses classical performance on nonlinear classification tasks

A photonic quantum neural network (QNN) achieved 100% accuracy on a nonlinearly separable task, while a comparable artificial neural network (ANN) failed to learn the same problem. Classical models require at least quadruple the trainable parameters to solve similarly complex problems, yet the demonstrated QNN accomplished this with just two. This success demonstrates a proof-of-principle for realising and training gate-based QNNs on photonic hardware, suggesting an algorithmic advantage in specific machine learning applications.

Calculating the “effective dimension” of these networks, a measure of their capacity to learn complex patterns, revealed that the photonic QNN outperformed classical counterparts with an equivalent number of adjustable parameters. Further benchmarking using the Iris dataset showed the photonic QNN achieving up to 100% accuracy in both online and offline learning scenarios, demonstrating adaptability beyond simple, synthetic datasets. Successfully deploying circuits boasting comparatively strong effective dimension on a six-qubit photonic quantum processor validated the scalability of the approach, despite current limitations to small datasets and limited circuit depth.

Photonic quantum neural network outperforms classical models on benchmark datasets

A gate-based photonic quantum neural network achieved 100% accuracy on a nonlinearly separable task, exceeding the performance of a comparable classical artificial neural network which failed to converge. This success was observed using the XOR problem and a subset of the Iris dataset. Assessing durability to realistic noise processes, specifically photon loss and imperfections in phase-shifters, proved critical for practical implementation.

Photonic quantum neural network surpasses classical performance on nonlinear classification

A gate-based photonic quantum neural network (QNN) achieved 100% accuracy on a nonlinearly separable task, a feat that eluded a comparable classical artificial neural network (ANN). Single photons and probabilistic gates were utilised to emulate the standard quantum circuit model framework, paving the way for more efficient machine learning. Evaluating the expressive power of their QNNs by calculating their effective dimension, a measure linked to generalisation error, allowed for contrast with classical ANNs possessing an equivalent number of trainable parameters. Supervised binary classification tasks were employed to benchmark performance between photonic and superconducting QNNs, revealing that both quantum implementations exhibited lower cross-entropy loss and higher prediction accuracy than their classical counterparts.

The photonic QNN, trained using gradient-free optimisation, successfully converged on the XOR problem and a subset of the Iris dataset, while maintaining durability against realistic noise sources like photon loss and phase-shifter imperfections. Deploying circuits on a six-qubit photonic quantum processor achieved high classification accuracies in both online and offline learning scenarios. The simplest deployed QNN, containing only two trainable parameters, outperformed ANNs requiring at least four times more parameters to solve the same problems, suggesting an algorithmic advantage.

Previous work by Abbas et al. demonstrated enhanced effective dimension and faster training in numerical studies of similar QNN models, providing a foundation for this research. Future work will focus on exploring the limits of this approach and expanding the range of problems solvable with gate-based photonic QNNs, alongside investigating methods for mitigating the effects of noise and improving the overall performance of these quantum classifiers. Validating the potential of photonic quantum computation for specific tasks, the achievement of 100% accuracy on a challenging, nonlinearly separable problem where traditional computers failed is significant. The concept of “effective dimension”, quantifying a network’s learning capacity, proved key in demonstrating this advantage over classical artificial neural networks. Consequently, this work shifts the focus from theoretical possibilities to practical realisation, prompting investigation into scaling these networks and exploring their application to more complex, real-world datasets.

The researchers demonstrated superior performance from quantum neural networks when classifying data, achieving 100% accuracy on a challenging problem where equivalent classical networks failed to learn. This matters because it suggests a potential algorithmic advantage for quantum computation in machine learning tasks, even with a small number of trainable parameters. Using single photons and a six-qubit processor, they showed these networks were robust to realistic noise and maintained lower error rates than classical artificial neural networks. The authors intend to expand this approach to solve a wider range of problems and further improve performance.