Quantum Recurrent Unit: A Parameter-Efficient Quantum Neural Network Architecture for NISQ Devices

The rapid growth of modern machine learning (ML) models presents fundamental challenges in parameter efficiency and computational resource requirements. This study introduces the Quantum Recurrent Unit (QRU), a novel quantum neural network (NN) architecture specifically designed to address these challenges while remaining compatible with Noisy Intermediate-Scale Quantum (NISQ) devices. QRU leverages quantum controlled-SWAP (C-SWAP; Fredkin) gates to implement an information selection mechanism inspired by classical Gated Recurrent Units (GRUs), enabling selective processing of temporal information via quantum operations. Through its innovative recurrent architecture featuring measurement results feedforward state propagation and shared parameters across time steps, QRU achieves constant circuit depth and constant parameter count regardless of input sequence length, effectively circumventing stringent NISQ hardware constraints. We systematically validate QRU through three progressive experiments: (1) oscillatory behavior prediction, where 72-parameter QRU matches 197-parameter classical GRU performance; (2) Wisconsin Diagnostic Breast Cancer classification, where 35 parameters achieve 96.13% accuracy comparable to 167-parameter artificial NNs; and (3) MNIST handwritten digit recognition, where 132 parameters reach 98.05% accuracy, outperforming a 27,265-parameter convolutional NN. These results demonstrate that QRU consistently achieves comparable or superior performance with significantly fewer parameters than classical NNs while maintaining constant quantum circuit depth. The architecture’s quantum-native design, combining C-SWAP-based information selection with novel recurrent processing, suggests QRU’s potential as a fundamental building block for next-generation ML systems, offering a promising pathway toward more efficient and scalable quantum ML architectures.

💡 Research Summary

The paper introduces the Quantum Recurrent Unit (QRU), a novel quantum neural network architecture tailored for Noisy Intermediate‑Scale Quantum (NISQ) devices. QRU addresses two major challenges in modern machine learning: the explosive growth of model parameters with increasing task complexity, and the stringent hardware constraints of near‑term quantum computers (limited qubit counts, shallow coherence times, and gate noise). Inspired by the gated recurrent unit (GRU), QRU replaces the classical update and reset gates with quantum controlled‑SWAP (C‑SWAP, also known as the Fredkin gate). The C‑SWAP acts as a quantum information‑selection mechanism: a control qubit, whose state encodes a learned mixing ratio, determines whether two target qubits (representing the previous hidden state and the new candidate state) are swapped. This operation simultaneously performs the update and reset functions within a single quantum operation, leveraging superposition to blend information without the need for separate classical calculations.

The architecture consists of four main components: (1) Input and hidden‑state encoding via angle encoding, where classical data vectors are mapped to rotation angles on data and hidden qubits; (2) the C‑SWAP‑based update/reset block that implements the quantum gating; (3) a variational layer composed of parameterized single‑qubit rotations and entangling CNOTs, providing universal function approximation under NISQ depth limits; and (4) measurement in both Z and X bases at the end of each time step. Measurement outcomes are split into two streams: one supplies the model’s output for the current step, the other is fed forward as the hidden state for the next QRU. Crucially, the same set of trainable parameters (weights and rotation angles) is reused at every time step, guaranteeing that circuit depth and parameter count remain constant regardless of sequence length.

Three experiments demonstrate QRU’s parameter efficiency and performance. In a synthetic oscillatory time‑series prediction task, a 72‑parameter QRU matches the mean‑squared error of a 197‑parameter classical GRU. For the Wisconsin Diagnostic Breast Cancer (WDBC) dataset, a QRU with only 35 trainable parameters achieves 96.13 % classification accuracy, comparable to a 167‑parameter artificial neural network. Finally, on the MNIST handwritten digit benchmark, a 132‑parameter QRU reaches 98.05 % accuracy, surpassing a conventional convolutional neural network that uses 27,265 parameters. Across all tasks, QRU maintains a fixed, shallow circuit depth, making it amenable to execution on current NISQ hardware.

The authors discuss limitations: the current results are obtained via quantum simulators; real‑world implementation would need efficient, low‑error C‑SWAP gates, which are presently costly on hardware. They also note that measurement noise and decoherence could affect performance, suggesting future work on error mitigation, hardware‑friendly gate decompositions, and integration of QRU modules into larger architectures such as Transformers or hybrid quantum‑classical pipelines.

In summary, QRU offers a quantum‑native recurrent processing unit that combines information‑selective gating with parameter sharing, achieving constant-depth, highly parameter‑efficient quantum neural networks. The work provides strong evidence that carefully designed quantum architectures can rival or exceed classical counterparts while respecting the practical constraints of NISQ devices, opening a pathway toward scalable quantum machine‑learning systems.