Spiking-quantum Convolutional Neural Network Achieves 86% Accuracy with 0.5% Data Re-upload and Joint Training

The pursuit of more efficient artificial intelligence drives innovation in areas like neuromorphic computing and quantum computation, and researchers are increasingly exploring hybrid models that combine the strengths of both. Luu Trong Nhan, Luu Trung Duong, and Pham Ngoc Nam, from Can Tho University and VinUniversity, alongside Truong Cong Thang from The University of Aizu, present a new architecture, the Spiking-Data Re-upload Convolutional Neural Network (SQDR-CNN), that overcomes limitations in existing hybrid spiking-quantum neural networks. This novel approach enables the joint training of both convolutional spiking neural networks and quantum circuits within a single optimisation process, crucially removing the need for pre-trained spiking components and extensive dataset reduction. The team demonstrates that SQDR-CNN achieves 86% of the accuracy of state-of-the-art spiking neural networks while utilising only 0. 5% of the parameters of the smallest comparable model, representing a significant step towards more efficient and scalable learnable systems.

Research focuses on developing, training, and applying spiking neural networks, which more closely mimic biological neurons than traditional artificial neural networks. Key advancements involve innovative training methods, such as surrogate gradient learning, which overcomes the challenges of non-differentiable spiking activity. Scientists are also investigating recurrent spiking neural networks to enhance temporal processing capabilities and applying these networks to tasks like object tracking and image classification.

Alongside spiking neural networks, quantum machine learning represents a significant area of investigation. Further research focuses on techniques like quantum tomography and noise-adaptive quantum circuits to improve the robustness and efficiency of quantum machine learning models. A growing trend involves combining spiking and quantum paradigms, creating hybrid models that capitalize on the advantages of both. This innovative system addresses limitations in both fields, specifically overcoming the challenges of non-differentiable spiking activity and the constraints of current quantum hardware. The core of SQDR-CNN lies in its fusion of spiking convolutional layers, trained using surrogate gradients, with a quantum data re-uploading classifier, which functions as a universal function approximator within the quantum domain, enabling end-to-end training. Rigorous benchmarking demonstrates that SQDR-CNN achieves 86% of the accuracy of state-of-the-art spiking neural network baselines, yet utilizes only 0.

5% of the smallest spiking model’s parameters, representing a significant reduction in computational cost. A systematic investigation of training regimes, including variations in optimizers, noisy circuit components, and initialization schemes, confirms the robustness and generalization capacity of SQDR-CNN. The work clarifies theoretical foundations underpinning practices previously used without clear explanation, establishing a more solid basis for future research. This work addresses limitations in existing hybrid quantum-spiking models, which often rely on pre-trained spiking encoders and struggle with scalability. The SQDR-CNN achieves 86% of the accuracy of state-of-the-art spiking neural network baselines, yet utilizes only 0. 5% of the smallest spiking model’s parameters, demonstrating a significant reduction in computational complexity.

Benchmarking against established spiking neural network architectures confirms comparable accuracy with a substantially smaller parameter set. Experiments reveal that the model performs robustly under varying training conditions, including different optimizers, noisy circuit components, and initialization schemes. This robustness is achieved through a theoretically grounded approach, clarifying practices previously used without clear explanation. Researchers systematically evaluated the model’s performance, demonstrating its ability to handle diverse training regimes and maintain accuracy even with simulated noise. The SQDR-CNN incorporates a quantum data re-uploading classifier, acting as a universal function approximator within the quantum domain, and utilizes spiking convolutional layers trained via surrogate gradients. This integration of spiking and quantum paradigms opens new research directions and fosters technological progress in multi-modal, learnable systems, paving the way for more efficient and robust artificial intelligence.

Spiking Quantum CNNs Achieve High Fidelity

This work presents SQDR-CNN, a novel hybrid model integrating spiking convolutional neural networks with quantum data re-upload techniques. Researchers successfully demonstrated end-to-end optimization within this framework, overcoming limitations found in previous approaches that relied on pre-trained spiking encoders. Experimental results indicate that SQDR-CNN variants achieve competitive performance against state-of-the-art benchmarks while utilizing significantly fewer parameters than traditional spiking neural networks. The team rigorously evaluated the model’s performance under various conditions, including simulated noisy quantum circuits and different optimization settings, establishing a valuable reference point for future investigations. While acknowledging the challenges inherent in noisy intermediate-scale quantum (NISQ) devices, the authors suggest that future work should focus on strategies to mitigate quantum errors and barren plateaus, potentially through the use of noise-resilient circuit designs or adaptive parameter initialization. The source code and datasets used in this research are publicly available, encouraging further exploration and development within the field of hybrid quantum machine learning.