A Lightweight Hybrid Quantum–Classical Neural Architecturewith Structured Representation Adaptation

Authors:
  • WNWan Nursyafeeza Binti Wan Mohd Nasir
    Wan Nursyafeeza Binti Wan Mohd Nasir
  • SRSheikh Sayed Bin Rahman
    Sheikh Sayed Bin Rahman
  • CAChimeremma Sandra Amadi
    Chimeremma Sandra Amadi

Abstract
Hybrid quantum-classical neural networks aim to integrate quantum feature transformations with classical deep learning; however their efficacy is limited by the mismatch between high-dimensional classical features and low-dimensional quantum inputs. This study proposes a lightweight hybrid quantum-classical architecture for image classification that addresses the mismatch between high-dimensional classical features and low-dimensional quantum inputs through a structured representation adapter. Within a single framework, the model integrates a variational quantum circuit, nonlinear dimensionality compression and a compact convolutional backbone. Experiments on the Fashion-MNIST dataset show that the proposed method reduces the number of trainable parameters by approximately 69% while achieving 85.68% accuracy, which closely matches the classical baseline accuracy of 85.95%, highlighting a significant trade-off between accuracy and efficiency. Ablation studies show that the adapter is important for maintaining the performance, while the quantum layer contributes only slight improvement. These results imply that effective hybrid learning depends more on architectural design and representation alignment than increasing quantum complexity. The proposed approach provides a fair trade-off between training stability, accuracy and parameter efficiency. These findings challenge the assumption that increased quantum complexity directly leads to better performance in hybrid learning systems.