The Role of Quantum in Hybrid Quantum-Classical Neural Networks: A Realistic Assessment

TL;DR

This study tackles the question of what quantum components actually contribute to the performance of hybrid quantum-classical neural networks (HQNNs) on real-world data. Using a rigorous statistical framework across 1D ECG, 2D ultrasound, and 3D chest CT tasks, it systematically varies classical preprocessing, latent space size, quantum encoding, entanglement, and observables, comparing against fully classical baselines. The key finding is that hybrid models often underperform compared with classical references, though amplitude encoding and QCNN variants can match or exceed classical performance in some cases; smaller latent spaces and specific encoding choices drive these outcomes. The work provides practical guidance for near-term HQNN design, emphasizing cautious claims about quantum advantages and highlighting directions for objective evaluation as hardware and algorithms mature.

Abstract

Quantum machine learning has emerged as a promising application domain for near-term quantum hardware, particularly through hybrid quantum-classical models that leverage both classical and quantum processing. Although numerous hybrid architectures have been proposed and demonstrated successfully on benchmark tasks, a significant open question remains regarding the specific contribution of quantum components to the overall performance of these models. In this work, we aim to shed light on the impact of quantum processing within hybrid quantum-classical neural network architectures through a rigorous statistical study. We systematically assess common hybrid models on medical signal data as well as planar and volumetric images, examining the influence attributable to classical and quantum aspects such as encoding schemes, entanglement, and circuit size. We find that in best-case scenarios, hybrid models show performance comparable to their classical counterparts, however, in most cases, performance metrics deteriorate under the influence of quantum components. Our multi-modal analysis provides realistic insights into the contributions of quantum components and advocates for cautious claims and design choices for hybrid models in near-term applications.

Figures (6)

• Figure 1: Examples from the three datasets used in this study: 1D ECG (a), 2D breast ultrasound (b), and 3D chest CT (c).
• Figure 2: Schematic illustration of the general HQNN architecture. A classical network $L(\boldsymbol{\Tilde{x}}, \boldsymbol{w})$ (with 3, 1 or no convolutional layers) and a fully-connected layer (with or without $\pi \cdot \tanh$ activation) first extracts a latent feature vector $\boldsymbol{x}$ (16 or 256 features). $\boldsymbol{x}$ is then encoded into a quantum circuit $U(\boldsymbol{x}, \boldsymbol{\theta})$ using either angle or amplitude encoding (with optional entanglement). The expectation value of either a local or global Pauli‑Z measurement is finally linearly mapped to the output logit.
• Figure 3: Hybrid and classical model performances with different levels of classical pre-processing across data modalities. The influence of the pre-processing depth (3conv, 1conv, 0conv) is statistically compared between hybrid models and classical reference models (*: $p<0.05$, **: $p<0.01$, ***: $p<0.001$).
• Figure 4: Hybrid and classical model performances with different latent space dimensions across data modalities. The influence of the latent space dimension (16 vs. 256 features) is statistically compared between hybrid models and classical reference models (*: $p<0.05$, **: $p<0.01$, ***: $p<0.001$).
• Figure 5: ROC-AUC scores achieved by hybrid models, grouped by QNN architecture. Amplitude encoding models (Amp-Gen, QCNN) generally outperformed angle encoding models (Ang-RY, Ang-Arb).