Comparison of hybrid quantum neural networks for MNIST classification task

Source abstract

This study presents a comparative analysis of two parameterized quantum circuit architectures in hybrid quantum-classical neural networks for the MNIST handwritten digit classification task. A classical convolutional block was used for feature extraction and dimensionality reduction, with features then fed to a quantum layer. Schemes using Pauli operation chains and Bell states for entanglement creation were compared. Experimental results revealed a significant impact of gate selection: the Bell state-based model, despite having fewer operations, demonstrated statistically significantly higher accuracy and convergence speed. This indicates that entanglement depth on key qubits is a more significant factor than the total number of parameterized operations. Thus, targeted application of deep entanglement is more effective for training than superficial transformation of all qubits, highlighting the importance of optimizing quantum circuit topology.