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

AI summary

Overview

Research area: Neuromorphic computing and quantum machine learning, specifically hybrid spiking neural network (SNN) and parameterized quantum circuit (PQC) architectures for image classification.

Technical level: Advanced. The paper combines surrogate-gradient training of spiking convolutional networks, quantum data re-uploading circuits, variational quantum optimization, and analytical gradient-variance proofs.

Scope: The paper proposes SQDR-CNN, a hybrid spiking-quantum convolutional architecture that trains a spiking encoder and a quantum data-reuploading classifier jointly in one backpropagation pass, and benchmarks it against state-of-the-art SNNs on three grayscale image datasets under heavy parameter constraints.

What This Paper Is About

Hybrid spiking-quantum neural networks (SQNNs) have been limited because spiking activity is non-differentiable and current SNN encoders scale poorly, so most prior work had to freeze a pretrained spiking encoder or train on subsets of data rather than on full datasets. This paper introduces SQDR-CNN, which instead trains a convolutional SNN encoder and a quantum data-reuploading circuit end-to-end with surrogate gradients, so no pretrained encoder and no dataset subsetting are required. The goal is to show that a hybrid model can reach meaningful accuracy on standard vision benchmarks while using a tiny fraction of the parameters of conventional SNN baselines.

Key Contributions

1. SQDR-CNN architecture. The authors introduce a theoretically motivated model that fuses spiking convolutional layers, trained with surrogate gradients, with a quantum classifier built on data re-uploading—a combination they state was unexplored in prior hybrid spiking-quantum model design. The model reaches 86% of the mean top-performing accuracy of the SOTA SNN baselines while using only 0.5% of the smallest spiking model's parameters.

2. Benchmarking against established SNN architectures. SQDR-CNN is compared with ST-RSBP, SEW-ResNet18, LISNN, PLIF, Spiking-ResNet, ANN-ResNet18 and ANN-ResNet34, showing comparable accuracy at a far smaller parameter count.

3. Theoretical basis for existing practice. The paper supplies analysis for design choices that prior work used without clear explanation, notably the classical MLP placed after the PQC and how it can rescale quantum gradient variance.

4. Robustness and generalization evaluation. The authors report evaluating the model under varying training regimes, including different optimizers, noisy circuit components, and initialization schemes, and compare their coverage against prior works in a contribution table (Table 1).

Main Findings

• Quantum data re-uploading outperforms alternative PQCs. On MNIST, Fashion-MNIST and KMNIST, the data-reupload circuit [2b-9q] achieved 84.03%, 76.18% and 80.04% (average 80.08), ahead of SQNN's PQC (82.45, 74.50, 77.55; average 78.17) and HEA (80.30, 72.85, 76.26; average 76.47). Average gaps ranged from 1.7% (SQNN's PQC vs. HEA) to 3.61% (data re-upload vs. HEA).

• Extreme parameter efficiency. SQDR-CNN variants use 803, 857, 1,190 and 1,298 optimizable parameters. The paper states the best model, SQDR-CNN[4b-18q], reaches approximately 86% of the mean top-performing accuracy of the SOTA baselines (PLIF with 99.72%) while using only 0.5% of the smallest SOTA model's parameters — a 99.5% size reduction compared to LISNN.

• Accuracy results across datasets. SQDR-CNN[2b-9q]: 84.03 / 76.18 / 80.04; SQDR-CNN[4b-9q]: 88.32 / 79.99 / 84.36; SQDR-CNN[2b-18q]: 86.55 / 77.80 / 82.01; SQDR-CNN[4b-18q]: 88.32 / 82.18 / 83.93 (MNIST / Fashion-MNIST / KMNIST).

• Outperformance of Spiking-ResNet variants. All SQDR-CNN variants beat the Spiking-ResNet models, with gaps as low as 6.59% (SQDR-CNN[2b-9q] vs. Spiking-ResNet34 on Fashion-MNIST) and as high as 73.92% (SQDR-CNN[4b-9q] vs. Spiking-ResNet50 on KMNIST).

• Fashion-MNIST is the weakest dataset. The lowest relative performance appears there, where SQDR-CNN[2b-9q] attains roughly 78% of PLIF's accuracy while using merely 0.3% of LISNN's parameter count (99.7% reduction).

• Post-processing MLP rescales quantum gradients. Theorem 2 (concentration reduction over shared variance) shows that if Var(∂y/∂f) + E[∂y/∂f]² ≥ 1 and the two gradient terms are independent, then the variance of the PQC gradient multiplied by the MLP gradient is greater than or equal to the variance of the PQC gradient alone. Empirically, models including the MLP showed both higher classification accuracy and consistently larger PQC-layer gradient variance on MNIST, which the authors link to reduced barren-plateau effects.

• Convergence without pretraining or subsetting. Unlike its predecessor SQNN, SQDR-CNN converges to reasonable performance without relying on a pretrained spiking encoder or on subsets of the datasets.

• Noise, optimizer and initialization results are not reported in the available content. Section 5.2 opens with a definition of a NISQ algorithm and is truncated mid-sentence; the paper's stated intent is to benchmark under noisy simulated quantum environments, but no numerical results for noisy circuits, optimizer comparisons, or initialization schemes appear in the content provided.

Methodology in Plain English

Images first pass through a spiking convolutional encoder built from convolution, batch normalization and max pooling layers, ending in adaptive pooling that compresses the feature maps into a compact vector. Because spiking neurons emit discrete, non-differentiable spikes, the encoder is trained with a surrogate gradient—specifically an arctan surrogate function with α = 2 applied to the Heaviside step of the integrate-and-fire neurons—so backpropagation can flow through the spiking layers.

That feature vector is then encoded into a quantum circuit. The chosen circuit design repeats a pattern of single-qubit rotation gates followed by circularly arranged controlled-Z entanglers; the first rotation layer encodes classical data and the second acts as trainable parameters, and multiple such blocks are stacked to increase expressive power. Qubits are read out with Pauli-Z measurements in the computational basis.

The measurement results go to a classical multilayer perceptron that produces the final prediction, and the whole stack—spiking encoder, quantum circuit, and MLP—is optimized together with one backpropagation pass. Training used PyTorch with SpikingJelly for the spiking encoder and PennyLane for quantum simulation, phase coding with T = 10 time steps, the Adam optimizer (lr = 10⁻³, weight decay 10⁻³, β = (0.9, 0.999)), cross-entropy loss, 100 epochs, fixed random seeds, and an NVIDIA RTX 3060 GPU with 12 GB. The authors also derive a variance argument for why the trailing classical MLP helps the quantum circuit train.

Why This Matters

The work argues that spiking and quantum components can be optimized jointly rather than in a frozen two-stage pipeline, which removes the dependency on pretrained spiking encoders and dataset subsetting that constrained earlier SQNN work. If the reported parameter efficiency holds in broader settings, hybrid quantum-spiking models become plausible candidates for environments where model size, energy, and circuit width are all tightly constrained.

Real-world applications implied by the paper's focus on compact grayscale image classifiers:

• Resource-constrained edge and embedded vision, where models of a few hundred to roughly 1,300 parameters are attractive.
• Neuromorphic hardware deployments that already exploit sparse spiking activity for energy efficiency.
• Digit and character recognition pipelines of the kind represented by MNIST and KMNIST.
• Early-stage quantum-assisted classification on NISQ-era devices, where circuit width and depth are sharply limited.

Industry relevance centers on the two cost drivers the paper targets: model parameter count and quantum circuit size. The demonstration that a 9- or 18-qubit circuit paired with a small spiking encoder can approach standard SNN accuracy positions this line of work for teams exploring hybrid quantum-classical accelerators, low-power vision sensors, and near-term quantum cloud services.

Future Directions

• Noise robustness remains unquantified in the available content. The paper states an intent to benchmark SQDR-CNN under noisy simulated quantum environments and NISQ noise models, but the numerical outcome of that study is not present in the truncated text.
• Optimizer and initialization studies. The contribution list promises benchmarking across different optimizers and initialization schemes for the PQC; these comparisons are not reported in the content provided.
• Scaling the architecture. Since performance improved when moving from 2 to 4 data-reupload blocks and from 9 to 18 qubits on most datasets, the question of how far this scaling continues—and where barren plateaus or hardware limits intervene—remains open.
• Closing the accuracy gap. The best SQDR-CNN result reaches about 86% of the mean top-performing SOTA accuracy; narrowing that gap without giving up the 0.5% parameter footprint is the central open challenge the paper leaves.

Target Audience

This paper is most useful to researchers and graduate students working on neuromorphic computing, spiking neural networks, and quantum machine learning, particularly those interested in hybrid quantum-classical training pipelines and surrogate-gradient methods. It also suits engineers evaluating whether NISQ-era circuits can contribute to practical vision tasks under severe parameter and qubit budgets, and readers who want a worked example of joint backpropagation across spiking and quantum layers.

Authors’ abstract

The rapid advancement of artificial intelligence (AI) and deep learning (DL) has catalyzed the emergence of several optimization-driven subfields, notably neuromorphic computing and quantum machine learning. Leveraging the differentiable nature of hybrid models, researchers have explored their potential to address complex problems through unified optimization strategies. One such development is the Spiking Quantum Neural Network (SQNN), which combines principles from spiking neural networks (SNNs) and quantum computing. However, existing SQNN implementations often depend on pretrained SNNs due to the non-differentiable nature of spiking activity and the limited scalability of current SNN encoders. In this work, we propose a novel architecture, Spiking-Quantum Data Re-upload Convolutional Neural Network (SQDR-CNN), that enables joint training of convolutional SNNs and quantum circuits within a single backpropagation framework. Unlike its predecessor, SQDR-CNN allow convergence to reasonable performance without the reliance of pretrained spiking encoder and subsetting datasets. We also clarified some theoretical foundations, testing new design using quantum data-reupload with different training algorithm-initialization and evaluate the performance of the proposed model under noisy simulated quantum environments. As a result, we were able to achieve 86% of the mean top-performing accuracy of the SOTA SNN baselines, yet uses only 0.5% of the smallest spiking model's parameters. Through this integration of neuromorphic and quantum paradigms, we aim to open new research directions and foster technological progress in multi-modal, learnable systems.