Pulsed Learning Enables Quantum Data Re-uploading Models on Noisy Intermediate-Scale Hardware, Addressing Variational Circuit Limitations
Summarize this article with:
Quantum machine learning promises revolutionary advances, but current methods struggle to function effectively on today’s limited quantum hardware, facing challenges with both training and susceptibility to noise. Ignacio B. Acedo, Pablo Rodriguez-Grasa, Pablo Garcia-Azorin, and Javier Gonzalez-Conde, from institutions including Quantum Mads and the University of the Basque Country UPV/EHU, investigate a fundamentally different approach by shifting learning directly to the level of the control pulses that manipulate quantum bits. Their work introduces a pulse-based method for data re-uploading, embedding trainable parameters into the very dynamics of the quantum system, and benchmarks this technique on a real superconducting processor with realistic noise. The results demonstrate that this pulse-based model consistently achieves higher accuracy and better generalisation than traditional gate-based methods, while also proving significantly more resilient to the effects of noise and errors, suggesting a promising pathway towards practical quantum machine learning in the near term. Pulse-Level Control for Quantum Machine Learning Scientists are shifting from designing quantum circuits using pre-defined gates to directly controlling the pulses that manipulate qubits, offering greater flexibility and potential for optimization. Pulse-based circuits can be more expressive and trainable than traditional gate-based circuits, addressing issues that hinder quantum machine learning (QML) training. This approach allows tailoring circuits to the specific characteristics of the quantum hardware, maximizing performance and minimizing errors. The research acknowledges the limitations of current noisy intermediate-scale quantum (NISQ) devices, including limited qubit count and coherence times, and emphasizes the importance of accurate noise modeling for both simulation and developing noise mitigation strategies. Pulse-level control allows designing circuits that are more robust to specific hardware imperfections, optimizing for features like cross-resonance gates and utilizing virtual Z gates. Thorough characterization of the quantum hardware is essential for accurate pulse design and calibration. The work heavily emphasizes the use of pulse-level control for building and training Quantum Neural Networks (QNNs), frequently using the MNIST handwritten digit dataset as a benchmark. Techniques like warm starts, which initialize optimization from a good starting point, are explored to improve training efficiency and navigate complex optimization landscapes. Pulse-level control is seen as a potential solution to the barren plateau problem, which can make training deep QNNs difficult. Key technologies include Trotter-Suzuki decomposition, used for simulating quantum circuits, and Cross-Resonance (CR) gates, a common two-qubit gate implementation in superconducting qubits. Experiments frequently reference IBM Quantum computers, including their 127-qubit processor, and the Brisbane Quantum Computer. In essence, the research describes a rapidly evolving field where scientists are harnessing the full potential of quantum hardware through precise pulse-level control, with a strong focus on developing practical and robust QML algorithms for NISQ devices.,. Pulse-Level Data Re-uploading for Quantum Machine Learning Scientists developed a novel approach to quantum machine learning by implementing data re-uploading directly at the pulse-control level, moving beyond traditional gate-based circuits. This work pioneers a method where trainable parameters are embedded into the fundamental dynamics of the quantum system, rather than applied as discrete gate operations.
The team engineered a superconducting transmon processor to serve as the platform for benchmarking this pulse-based model against its gate-based counterpart, utilizing realistic noise profiles to simulate practical conditions. The study involved formulating a pulse-based variant of data re-uploading, where input data is repeatedly embedded into the system’s dynamics through precisely tailored control signals. Researchers systematically increased noise strength to assess the resilience of the pulse-level implementation, carefully measuring fidelity and performance under varying levels of decoherence and control errors. The pulse-based model consistently outperformed the gate-based approach, exhibiting higher test accuracy and improved generalization capabilities under equivalent noise conditions. By compressing multi-gate sequences into shorter pulse schedules, the team reduced circuit depth and execution time, enhancing the efficiency of the quantum computation. The study demonstrates that pulse-level implementations retain higher fidelity for longer durations, indicating enhanced robustness against noise and errors, suggesting a viable path forward for practical quantum machine learning in the noisy intermediate-scale quantum era. This innovative methodology offers additional degrees of freedom for encoding information and optimizing the model, potentially reducing the severity of barren plateaus and improving the structure of the optimization landscape.,. Pulse-Based Data Re-uploading Beats Gate-Based Models Scientists have achieved a significant breakthrough in quantum machine learning by developing a pulse-based data re-uploading model that outperforms conventional gate-based approaches on noisy intermediate-scale quantum (NISQ) hardware. This work introduces a method for embedding trainable parameters directly into the dynamics of the quantum system, operating natively at the pulse-control level and bypassing the limitations of traditional gate-based circuits.
The team formulated a pulse-based variant of data re-uploading, allowing for a hardware-aware approach to variational quantum machine learning. Experiments demonstrate that the pulse-based model consistently exhibits higher test accuracy and improved generalization under equivalent noise conditions when compared to its gate-based counterpart. Systematic investigations into noise resilience reveal that pulse-level implementations retain higher fidelity for longer durations, demonstrating enhanced robustness to decoherence and control errors. Specifically, the research shows that the pulse-based approach maintains signal integrity and accuracy even as noise strength is systematically increased, a critical advantage for practical quantum computation.
The team’s methodology provides a general framework for translating the training of variational quantum machine learning models into pulse-level implementations, allowing the design of parametrized pulses for a given task. This approach avoids rigid structures often imposed on pulse-level models, offering greater flexibility and expressiveness. By compressing multi-gate sequences into shorter pulse schedules, the team also reduced circuit depth and execution time, further enhancing the efficiency of the quantum computation.,. Pulse-Based Quantum Machine Learning Outperforms Gates This research demonstrates a pulse-based implementation of a data re-uploading model, offering a new approach to quantum machine learning on near-term hardware. By directly embedding trainable parameters into the control pulses that manipulate the quantum system, the team achieved a more efficient and hardware-aligned representation of the learning model, circumventing limitations associated with traditional gate-based circuits. Benchmarking on a superconducting processor with realistic noise profiles reveals that this pulse-based model consistently outperforms its gate-based counterpart, exhibiting both higher test accuracy and improved generalization capabilities under comparable noise conditions. The findings indicate that pulse-level implementations maintain higher fidelity for longer durations, demonstrating enhanced resilience to decoherence and control errors, which are significant challenges in current quantum computing technology. While simulations were limited by computational resources, observed trends suggest that pulse-based models can sustain superior test accuracy even with increased complexity, implying a potential for scaling these architectures. This work establishes direct pulse-level optimization as a promising pathway toward scalable and experimentally viable quantum machine learning, uniting expressivity, generalization, and robustness within a hardware-native framework. 👉 More information 🗞 Pulsed learning for quantum data re-uploading models 🧠 ArXiv: https://arxiv.org/abs/2512.10670 Tags: