Quantum-classical hybrid neural networks are a potentially transformative application for emerging quantum hardware, but important questions remain regarding the true benefits of incorporating quantum processing. Dominik Freinberger and Philipp Moser from the Medical Informatics Research Unit at RISC Software GmbH, together with their colleagues, are tackling this challenge by rigorously investigating the performance of these complex models. Their work systematically evaluates common hybrid architectures using medical signal data and different types of images, carefully investigating how quantum factors such as encoding and entanglement affect the overall outcome. This study reveals that, while hybrid models match classical performance in limited cases, they frequently exhibit degraded metrics when quantum components are introduced, providing a realistic assessment that is essential to guide future development. This multimodal analysis promotes a careful approach to the design and implementation of quantum classical neural networks in real-world short-term applications.
Their work systematically evaluates common hybrid architectures using medical signal data and different types of images, carefully investigating how quantum factors such as encoding and entanglement affect the overall outcome. This study reveals that, while hybrid models match classical performance in limited cases, they frequently exhibit degraded metrics when quantum components are introduced, providing a realistic assessment that is essential to guide future development. This multimodal analysis promotes a careful approach to the design and implementation of quantum classical neural networks in real-world short-term applications.
Quantum machine learning is emerging as a promising application domain for near-term quantum hardware, especially through quantum-classical hybrid models. These approaches aim to exploit the strengths of both quantum and classical computation and overcome the limitations of the individual paradigms. In this study, we present a realistic assessment of the role of quantum computation within hybrid quantum-classical neural networks, highlighting the potential benefits and current limitations. The authors explore various quantum neural network architectures and analyze their performance compared to their classical counterparts, considering factors such as expressiveness, trainability, and computational complexity. Through detailed investigation, this study reveals the conditions under which quantum enhancement can realistically be expected in real-world machine learning tasks.
This study highlights that while quantum neural networks offer theoretical advantages in certain scenarios, achieving significant speedups and accuracy remains a major challenge. An important consideration is the noise and decoherence effects inherent in current quantum hardware, which can significantly degrade performance. Additionally, the overhead associated with quantum-classical communication and data encoding often negates potential benefits. The authors demonstrate that careful consideration of these practical constraints is important for designing effective hybrid models. Specifically, this work considers the use of variational quantum circuits as trainable parameters within classical neural networks, allowing gradient descent to rely on classical optimizers while leveraging the ability of quantum computers to explore high-dimensional parameter spaces.
However, the optimization environment for these hybrid models is complex and prone to sterile stagnation, which prevents efficient training. Addressing these challenges requires new optimization strategies and customized circuit designs. The analysis also includes a comparison of different quantum encoding schemes, such as amplitude encoding and angular encoding, to assess their impact on model expressiveness and resource requirements. The results show that the encoding choice significantly influences the performance and scalability of hybrid networks, highlighting the need to develop efficient and noise-resistant encoding strategies. Ultimately, this study provides a nuanced perspective on the potential of quantum machine learning and advocates a pragmatic approach that recognizes both the opportunities and limitations of near-term quantum technologies.
The research team designed a rigorous statistical study to analyze the contribution of quantum processing within a hybrid quantum-classical neural network (HQNN). Breaking away from previous approaches that often relied on pre-trained components, this work focused on a complete hybrid training scheme that jointly optimizes both classical and quantum elements. Scientists systematically varied the complexity of the classical preprocessing, the dimensionality of the latent space, the quantum encoding method, and the measurement strategy to accurately compare the performance of quantum and pure classical components. In our experiments, we used three different medical data modalities that represent common data types in the medical field: a 1-dimensional ECG signal, a 2-dimensional chest ultrasound image, and a 3-dimensional chest CT scan.
The research team obtained a 1D MIT-BIH arrhythmia dataset containing 105,026 annotated ECG recordings sampled at 360Hz and formulated a binary classification task to distinguish between normal and arrhythmic beats. For 2-dimensional imaging, the 2D BreastMNIST dataset of 780 grayscale images was utilized, and the images were resized to 224×224 pixels and normalized to the following range: [-1, 1]. This 3D data was obtained from the NoduleMNIST3D dataset, which consists of 1,633 CT scans of 64x64x64 voxels and classifies pulmonary nodules as benign or malignant.
To ensure a robust evaluation, this study pioneered a cross-validation approach, sampling 7,064 instances from the MIT-BIH dataset to create five folds, and carefully avoiding duplicate subjects to prevent data leakage. Similarly, five training and validation folds were created on the MedMNIST dataset, maintaining a balanced class distribution across each fold. The core of the methodology included a prototype HQNN architecture consisting of a classical neural network layer L(x̃, w) that preprocesses the raw input into a latent feature representation x. This latent representation was processed by a quantum neural network U(x, θ) before the final linear layer produced the output logit.
The researchers implemented three variations of the classic preprocessing layer L: 3conv (three convolutional layers), 1conv (one convolutional layer), and 0conv (one fully connected layer) to evaluate its impact. They investigated 16 and 256 latent dimensions and compared angle-based encoding with no activation using a scaled activation function π · Tanh. QNN U utilized 4 or 8 qubits corresponding to 16 or 256 latent dimensions, respectively, allowing detailed examination of the impact of quantum components on overall model performance. This careful setup allowed this study to provide realistic insights into the contribution of quantum components and advocate careful design choices in short-term applications.
Scientists conducted a rigorous statistical study to assess the contribution of quantum components within a hybrid quantum-classical neural network (HQNN) architecture. The research team systematically evaluated a common hybrid model using medical signal data, specifically one-dimensional ECG signals, two-dimensional chest ultrasound images, and three-dimensional chest CT scans, examining the influence of both classical and quantum components. Experiments revealed that in the best-case scenario, the hybrid model achieved performance comparable to a fully classical model, but in many cases performance metrics degraded when quantum components were included. In this study, we focused on a binary classification task and utilized the MIT-BIH arrhythmia dataset containing 105,026 annotated ECG recordings sampled at 360 Hz, each representing a 1-second cardiac cycle with 360 features, to formulate a task to distinguish between normal and arrhythmic heart beats.
For two-dimensional data, the team used the BreastMNIST dataset, which consists of 780 grayscale breast ultrasound images resized to 224 x 224 pixels and normalized to a range of -1 to 1. Three-dimensional data analysis included the NoduleMNIST3D dataset, which consists of 1,633 CT scans of 64 x 64 x 64 voxels and classifies pulmonary nodules as either benign or benign. malignant. Testing demonstrated that the team sampled 7,064 instances from the MIT-BIH dataset and created five cross-validation folds to ensure there was no subject overlap to prevent data leakage. Similarly, five separate training and validation folds were created on the MedMNIST dataset to maintain a balanced class distribution.
The adopted HQNN architecture consists of a classical neural network layer L(x̃, w) that processes the raw input into a latent feature representation x, which is encoded into a quantum circuit U(x, θ). Confirmed by measurements.
