Quantum Neural Networks Evolve with Differentiable Architecture Search for Improved Performance.

Differentiable Architecture Search (DARTS) presents a novel quantum neural network architecture search algorithm. It models the search as the evolution of a mixed state, yielding circuits for tasks including state initialisation, Hamiltonian optimisation and image classification. DARTS demonstrates improved convergence and noise robustness compared with existing techniques.

The pursuit of quantum advantage in machine learning necessitates the development of effective quantum neural networks (QNNs). Variational quantum algorithms (VQAs) offer a pathway to constructing these networks, yet identifying optimal circuit architectures for general machine learning tasks remains a significant challenge. Researchers at INSAIT, Sofia University and Nottingham Trent University have addressed this by developing a novel, differentiable architecture search algorithm – RhoDARTS – which models the search process using the principles of density matrix simulations. This approach, detailed in their paper, RhoDARTS: Differentiable Quantum Architecture Search with Density Matrix Simulations, allows for the evolution of mixed states within the search space, yielding circuits validated across state initialisation, Hamiltonian optimisation and image classification. The work is led by Swagat Kumar, Jan-Nico Zaech, Colin M. Wilmott and Luc Van Gool.

Differentiable Architecture Search Optimises Quantum Neural Networks for Image Classification

Recent advances in machine learning are driving exploration of quantum neural networks (QNNs) as potential successors to classical models. Researchers are actively investigating architectures and encoding methods suitable for implementation on noisy intermediate-scale quantum (NISQ) devices – quantum computers with a limited number of qubits and susceptibility to errors. A new study details a differentiable quantum architecture search (QAS) algorithm, termed DARTS, which systematically discovers optimised QNN designs for image classification, demonstrating improved performance and resilience compared with existing methods.

DARTS models the search process as the evolution of a mixed quantum state, allowing for a more nuanced exploration of circuit topologies than traditional approaches. The research investigates the impact of different classical data encoding schemes on QNN performance, focusing on methods to translate data – exemplified by the MNIST handwritten digit dataset – into quantum states suitable for processing by a variational quantum circuit. Three primary encoding techniques were examined: amplitude encoding, angle encoding, and basis encoding.

Amplitude encoding maps data values to the amplitudes of a quantum state, while angle encoding utilises data to define rotation angles applied to qubits, establishing a direct relationship between classical input and quantum state manipulation. Basis encoding, conversely, selects computational basis states based on the classical data, offering a discrete representation of input values within the quantum system. The study details the construction of quantum circuits implementing these encoding schemes, employing sequences of single-qubit rotation gates – RY, RZ, and RX – applied to qubits, with the gate angles determined by the input data. Variations in circuit complexity and the specific combination of gate types were systematically explored to assess the impact of architectural choices on overall performance.

The encoded quantum states then serve as input to the QNN, where parameters within the circuit are optimised using classical algorithms to achieve a desired outcome, such as image classification. Results demonstrate that the choice of encoding scheme influences QNN performance, highlighting the importance of selecting an appropriate method for translating classical data into a quantum representation. The complexity of the encoding circuit, alongside the types of quantum gates employed, affects the overall efficiency and accuracy of the network, demanding careful consideration of architectural trade-offs.

Optimisation of circuit parameters proves crucial, with the selection of both the optimisation algorithm and its associated parameters significantly impacting the final results, emphasising the need for robust and efficient optimisation strategies.

Specific circuits, labelled W-4, W-5, W-6 and Z-4, Z-5, Z-6, represent variations in qubit number and circuit structure, all designed to represent pixel values through amplitude and angle encoding, allowing for a systematic exploration of architectural parameters. These circuits utilise rotations around the X, Y, and Z axes – implemented via RX, RY, and RZ gates – to map image data onto quantum states, providing a versatile toolkit for quantum image processing.

The improved resilience to noise demonstrated by DARTS is critical for practical implementation on NISQ devices, where quantum computations are susceptible to errors, demanding robust architectures and error mitigation strategies. This ability suggests a pathway towards more reliable quantum machine learning applications, paving the way for practical quantum image processing systems.

Future work should focus on scaling these encoding schemes to higher-resolution images and more complex datasets, expanding the applicability of these techniques to real-world image processing tasks. Investigating the transferability of learned architectures across different image classification tasks represents another important avenue for research, potentially enabling the development of general-purpose quantum image processing systems. Exploring alternative encoding strategies, beyond amplitude and angle encoding, could further enhance the performance and efficiency of QNNs. Finally, integrating these optimised circuits into complete quantum machine learning pipelines, including measurement and classification layers, will allow for a comprehensive evaluation of their performance on real-world image classification problems and provide a benchmark for assessing the potential advantages of quantum machine learning.