Quantum Recurrent Unit (QRU): An Efficient Parameterized Quantum Neural Network Architecture for NISQ Devices

A research team from National Taiwan University proposed the Quantum Recurrent Unit (QRU) architecture, optimized for the constraints of Noisy Intermediate-Scale Quantum (NISQ) devices. It implements an information selection mechanism via quantum C-SWAP gates, achieving a parameter reduction of 63.5% to 99.5% while maintaining a constant quantum circuit depth, and its effectiveness has been verified in oscillation prediction, breast cancer classification, and MNIST recognition tasks.

Modern machine learning models face the challenge of parameter explosion; the number of parameters in classical RNN variants grows linearly with sequence length. Current NISQ devices have limited qubits and short coherence times, making them unable to support deep quantum circuits. Traditional methods struggle to balance circuit depth and expressive power, so designing efficient quantum neural networks under NISQ constraints has become a core problem.

1. Quantum Domain Information Selection Mechanism

Drawing on the idea of classical GRU, it implements selective information transmission and forgetting at the quantum state level via C-SWAP gates, leveraging quantum parallelism and entanglement properties to avoid the complex implementation of classical activation functions.

2. Measurement Result Feedforward State Propagation

Partial measurement of the quantum state is performed at each time step, and the results are fed back to parameterize the next time step, maintaining a constant quantum circuit depth and solving the parameter explosion problem of classical RNNs.

3. Cross-Time-Step Parameter Sharing

The same set of trainable parameters is shared across all time steps, reducing training difficulty and improving generalization ability.

Oscillation Prediction Task

QRU achieves accuracy comparable to classical GRU (197 parameters) with only 72 parameters, a parameter reduction rate of 63.5%.

Breast Cancer Classification Task

QRU achieves an accuracy of 96.13% with 35 parameters, which is only 21% of the parameter count of classical ANN (a reduction of 78.7%).

MNIST Recognition Task

QRU achieves an accuracy of 98.05% with 132 parameters (surpassing classical CNN), a parameter reduction rate of 99.5%.

Validated on IBM Quantum's ibm_marrakesh device, using Observable Target Calibration (OTC) noise injection training and applying QESEM error mitigation technology, the deviation of measurement results is controlled within 0.3% of the ideal state vector.

Software Stack: PennyLane quantum simulation framework, JAX automatic differentiation, Qiskit hardware interface. Code Structure: Includes experimental directories for oscillation prediction, breast cancer classification, MNIST recognition, hardware validation, and supplementary experiments. Ablation Study: Validates the necessity of C-SWAP gates, the advantages of dual ground state measurement, and the impact of feature ordering.

• Proves the feasibility of practical quantum neural networks under NISQ constraints, providing a paradigm for architecture design.
• Demonstrates the advantages of quantum computing in parameter efficiency, offering potential for edge deployment, federated learning, and other scenarios.
• Opens up new directions for hybrid quantum-classical architectures; future extensions can include complex architectures like Transformers.