[Paper] Parameter efficient hybrid spiking-quantum convolutional neural network with surrogate gradient and quantum data-reupload

Source: arXiv - 2512.03895v1

Overview

The paper introduces SQDR‑CNN, a hybrid spiking‑quantum convolutional neural network that can be trained end‑to‑end with ordinary back‑propagation. By marrying spiking neural networks (SNNs) with quantum data‑reupload circuits, the authors demonstrate that a tiny, fully trainable model can reach near‑state‑of‑the‑art (SOTA) accuracy on vision benchmarks while using only a fraction of the parameters of conventional SNNs.

Key Contributions

• Jointly trainable hybrid architecture – First design that integrates a convolutional SNN encoder and a variational quantum circuit (VQC) within a single differentiable pipeline, eliminating the need for a pretrained spiking encoder.
• Quantum data‑reupload layer – Introduces a reusable quantum feature‑encoding scheme that repeatedly injects classical activations into a shallow VQC, boosting expressivity without deepening the circuit.
• Surrogate‑gradient back‑propagation – Adapts surrogate gradient methods to propagate errors through the spiking non‑linearity and the quantum layer simultaneously.
• Parameter efficiency – Achieves ~86 % of the best SNN accuracy while using only 0.5 % of the parameters of the smallest competing spiking model.
• Robustness to quantum noise – Evaluates the model under realistic noisy quantum simulators, showing graceful degradation and confirming feasibility on near‑term quantum hardware.

Methodology

1. Convolutional Spiking Front‑end – Input images are processed by a few 2‑D convolutional layers followed by leaky‑integrate‑and‑fire (LIF) neurons. The spiking activity is approximated with a smooth surrogate function so gradients can flow.
2. Quantum Data‑Reupload Block – The spike‑based feature maps are flattened and encoded into rotation angles of qubits. A shallow VQC (typically 2–3 layers of parameterized single‑qubit rotations and entangling CNOTs) is executed repeatedly, each round “re‑uploading” the same data with updated parameters. This mimics a deeper network while keeping circuit depth low.
3. Hybrid Loss & Optimization – The output probabilities from the quantum measurement are fed to a cross‑entropy loss. The whole system—CNN, spiking dynamics, and VQC—is optimized with Adam, using the surrogate gradient for spikes and the parameter‑shift rule for quantum gates.
4. Noise Modeling – Depolarizing and read‑out errors are injected into the quantum simulator to emulate NISQ‑era hardware, allowing the authors to assess stability under realistic conditions.

Results & Findings

• Parameter reduction: SQDR‑CNN uses ≈0.5 % of the parameters of the smallest SNN baseline yet still reaches a competitive accuracy.
• Training stability: The joint back‑propagation converges without any pre‑training of the spiking encoder, unlike prior SQNN works that relied on frozen SNNs.
• Noise resilience: Accuracy drops by < 3 % when depolarizing noise of 1 % is added, indicating the shallow quantum circuit can tolerate near‑term hardware imperfections.

Practical Implications

• Edge AI with ultra‑low memory – The extreme parameter efficiency makes the model attractive for ultra‑resource‑constrained devices (e.g., wearables, IoT sensors) that can offload the quantum block to a cloud‑based quantum processor.
• Hybrid inference pipelines – Developers can keep the bulk of the computation on classical GPUs (the spiking CNN) and invoke a lightweight quantum service only for the final classification boost, reducing quantum runtime and cost.
• Energy‑aware neuromorphic systems – Spiking encoders are inherently event‑driven and power‑sparse; coupling them with a small quantum circuit could further lower the energy per inference compared to dense CNNs.
• Rapid prototyping of quantum‑enhanced models – The surrogate‑gradient + parameter‑shift training loop can be plugged into existing PyTorch/TensorFlow workflows, letting ML engineers experiment with quantum layers without deep expertise in quantum programming.

Limitations & Future Work

• Scalability to larger datasets – Experiments were limited to relatively small vision benchmarks; extending to ImageNet‑scale tasks may require deeper classical back‑bones or more sophisticated quantum encodings.
• Hardware access – The study relies on simulated noisy quantum devices; real‑world performance on actual NISQ hardware could differ due to calibration errors and limited qubit counts.
• Circuit depth vs. expressivity trade‑off – While data‑reupload mitigates depth, the expressive power of shallow VQCs remains bounded; future work could explore adaptive circuit architectures or hybrid quantum‑classical attention mechanisms.
• Energy‑budget analysis – A detailed comparison of total energy consumption (including quantum communication overhead) is still missing and would be essential for edge‑deployment scenarios.

Bottom line: SQDR‑CNN shows that a carefully designed spiking‑quantum hybrid can deliver SOTA‑level performance with a fraction of the parameters, opening a practical pathway for developers to experiment with quantum‑enhanced AI in low‑resource environments.

Authors

• Luu Trong Nhan
• Luu Trung Duong
• Pham Ngoc Nam
• Truong Cong Thang

Paper Information

• arXiv ID: 2512.03895v1
• Categories: cs.NE
• Published: December 3, 2025
• PDF: Download PDF