Realistic Assessment Enables Quantum Contribution to Hybrid Neural Network Architectures

Hybrid quantum-classical neural networks represent a potentially transformative application for emerging quantum hardware, yet a critical question persists regarding the true benefit of incorporating quantum processing. Dominik Freinberger and Philipp Moser, from the Research Unit Medical Informatics at RISC Software GmbH, alongside their colleagues, address this challenge with a rigorous investigation into the performance of these complex models. Their research systematically evaluates common hybrid architectures using medical signal data and various image types, carefully examining how quantum elements such as encoding and entanglement influence overall results. The study reveals that while hybrid models can match classical performance in limited instances, they frequently exhibit diminished metrics when quantum components are introduced, offering a realistic assessment vital for guiding future development. This multi-modal analysis encourages a measured approach to the design and implementation of hybrid quantum-classical neural networks in practical, near-term applications.

Their research systematically evaluates common hybrid architectures using medical signal data and various image types, carefully examining how quantum elements such as encoding and entanglement influence overall results. The study reveals that while hybrid models can match classical performance in limited instances, they frequently exhibit diminished metrics when quantum components are introduced, offering a realistic assessment vital for guiding future development. This multi-modal analysis encourages a measured approach to the design and implementation of hybrid quantum-classical neural networks in practical, near-term applications.

Quantum machine learning has emerged as a promising application domain for near-term quantum hardware, particularly through hybrid quantum-classical models. These approaches leverage the strengths of both quantum and classical computation, aiming to overcome the limitations of each individual paradigm. This work presents a realistic assessment of the role of quantum computation within hybrid quantum-classical neural networks, focusing on the potential benefits and current limitations. The authors examine various quantum neural network architectures and analyse their performance in relation to classical counterparts, considering factors such as expressivity, trainability and computational complexity. Through a detailed investigation, the research clarifies the conditions under which quantum enhancements can be realistically expected in practical machine learning tasks.

The study highlights that while quantum neural networks offer theoretical advantages in certain scenarios, achieving substantial speedups or improved accuracy remains a significant challenge. A key consideration is the impact of noise and decoherence inherent in current quantum hardware, which can severely degrade performance. Furthermore, the overhead associated with quantum-classical communication and data encoding often negates potential gains. The authors demonstrate that careful consideration of these practical constraints is crucial for designing effective hybrid models. Specifically, the research explores the use of variational quantum circuits as trainable parameters within classical neural networks, allowing for leveraging the quantum computer’s ability to explore high-dimensional parameter spaces while relying on classical optimisers for gradient descent.

However, the optimisation landscape of these hybrid models can be complex and prone to barren plateaus, hindering efficient training. Addressing these challenges requires novel optimisation strategies and tailored circuit designs. The analysis also encompasses a comparison of different quantum encoding schemes, such as amplitude encoding and angle encoding, evaluating their impact on model expressivity and resource requirements. Results indicate that the choice of encoding significantly affects the performance and scalability of the hybrid network, emphasising the need for developing encoding strategies that are both efficient and robust to noise. Ultimately, the work provides a nuanced perspective on the potential of quantum machine learning, advocating for a pragmatic approach that acknowledges both the opportunities and limitations of near-term quantum technology.

The research team engineered a rigorous statistical study to dissect the contribution of quantum processing within hybrid quantum-classical neural networks (HQNNs). Departing from prior approaches that often relied on pre-trained components, this work focused on fully hybrid training schemes, jointly optimising both classical and quantum elements. Scientists systematically varied classical pre-processing complexity, latent space dimensionality, quantum encoding methods, and measurement strategies to precisely compare the performance of quantum versus purely classical components. Experiments employed three distinct medical data modalities , one-dimensional ECG signals, two-dimensional breast ultrasound images, and three-dimensional chest CT scans , representing prevalent data types in healthcare.

The team sourced the 1D MIT-BIH Arrhythmia dataset, containing 105,026 annotated ECG recordings sampled at 360Hz, formulating a binary classification task to distinguish normal from arrhythmic beats. For two-dimensional imaging, the 2D BreastMNIST dataset of 780 grayscale images was utilised, with images resized to 224×224 pixels and normalised to a range of [-1, 1]. The three-dimensional data came from the NoduleMNIST3D dataset, comprising 1,633 CT scans of 64x64x64 voxels, classifying pulmonary nodules as benign or malignant.

To ensure robust evaluation, the study pioneered a cross-validation approach, sampling 7,064 instances from the MIT-BIH dataset and creating five folds, carefully avoiding subject overlap to prevent data leakage. Similarly, five training and validation folds were created for the MedMNIST datasets, maintaining balanced class distributions across each fold. The core of the methodology involved a prototypical HQNN architecture, comprising classical neural network layers, L(x̃, w), that pre-processed raw input into a latent feature representation, x. This latent representation was then processed by a quantum neural network, U(x, θ), before a final linear layer produced the output logits.

Researchers implemented three variants of the classical pre-processing layer, L, to assess its influence: 3conv (three convolutional layers), 1conv (one convolutional layer), and 0conv (a single fully-connected layer). They explored latent dimensions of 16 and 256, and for angle-based encoding, compared using a scaled activation function π · tanh with no activation. The QNN, U, utilised 4 or 8 qubits corresponding to the latent dimensions of 16 or 256, respectively, enabling a detailed examination of the quantum component’s impact on overall model performance. This meticulous setup allowed the study to provide realistic insights into the contributions of quantum components and advocate for cautious design choices in near-term applications.

Scientists have conducted a rigorous statistical study to assess the contribution of quantum components within hybrid quantum-classical neural network (HQNN) architectures. The research team systematically evaluated common hybrid models using medical signal data , specifically, one-dimensional ECG signals, two-dimensional breast ultrasound images, and three-dimensional chest CT scans , to examine the influence of both classical and quantum elements. Experiments revealed that, in best-case scenarios, hybrid models achieved performance comparable to their fully classical counterparts, but frequently, performance metrics deteriorated with the inclusion of quantum components. The study focused on binary classification tasks, utilizing the MIT-BIH Arrhythmia dataset containing 105,026 annotated ECG recordings sampled at 360Hz, each representing a one-second cardiac cycle with 360 features, and formulating a task to distinguish between normal and arrhythmic heartbeats.

For the two-dimensional data, the team employed the BreastMNIST dataset, comprising 780 grayscale breast ultrasound images resized to 224×224 pixels, normalized to a range of -1 to 1. Analysis of the three-dimensional data involved the NoduleMNIST3D dataset, consisting of 1,633 CT scans of 64x64x64 voxels, classifying pulmonary nodules as either benign or malignant. Tests demonstrated that the team sampled 7,064 instances from the MIT-BIH dataset and created five cross-validation folds, ensuring no subject overlap to prevent data leakage. Similarly, five separate training and validation folds were created for the MedMNIST datasets, maintaining balanced class distributions.

The HQNN architecture adopted comprised classical neural network layers, L(x̃, w), processing raw input into a latent feature representation, x, which was then encoded into a quantum circuit, U(x, θ). Measurements confirm.