Quantum machine learning promises revolutionary advances, but current methods face challenges in working effectively with today's limited quantum hardware and are susceptible to training and noise. Ignacio B. Acedo, Pablo Rodriguez-Grasa, Pablo Garcia-Azorin, and Javier Gonzalez-Conde from institutions such as Quantum Mads and the Basque University UPV/EHU are investigating a fundamentally different approach by moving learning directly to the level of the control pulses that operate the qubits. Their work introduced a pulse-based method for re-uploading data, embedded trainable parameters into the dynamics of the quantum system itself, and benchmarked the technique on a real superconducting processor with realistic noise. The results show that this pulse-based model consistently achieves higher accuracy and better generalization than traditional gate-based methods, while being significantly more resilient to noise and error effects, suggesting a promising path towards practical quantum machine learning in the near term.
Pulse level control for quantum machine learning
Scientists are moving from designing quantum circuits using predefined gates to direct control of the pulses that operate the qubits, which provides greater flexibility and optimization possibilities. Pulse-based circuits are more expressive and easier to train than traditional gate-based circuits, allowing them to address issues that hinder quantum machine learning (QML) training. This approach allows us to tune our circuits to the specific characteristics of quantum hardware, maximizing performance and minimizing errors. This study recognizes the limitations of current noisy intermediate-scale quantum (NISQ) devices, such as limited qubit numbers and coherence times, and highlights the importance of accurate noise modeling both in simulation and in the development of noise mitigation strategies.
Pulse-level control allows you to design circuits that are more robust to certain hardware defects, optimize features such as cross-resonant gates, and take advantage of virtual Z-gates. Thorough characterization of quantum hardware is essential for accurate pulse design and calibration. This research focuses on using pulse-level control to build and train quantum neural networks (QNNs), often using the MNIST handwritten digits dataset as a benchmark. Techniques such as warm start, which initializes optimization from a suitable starting point, are being investigated to improve training efficiency and navigate complex optimization environments.
Pulse-level control is seen as a potential solution to the barren plateau problem that can make training deep QNNs difficult. Key technologies include Trotter-Suzuki decomposition, used to simulate quantum circuits, and cross-resonance (CR) gates, a two-qubit gate implementation common in superconducting qubits. Experiments frequently reference the IBM and Brisbane quantum computers, which contain 127-qubit processors. Essentially, this study describes a rapidly evolving field in which scientists are focused on developing practical and robust QML algorithms for NISQ devices, leveraging the full potential of quantum hardware through precise pulse level control.
Re-uploading pulse-level data for quantum machine learning
Scientists have developed a new approach to quantum machine learning by going beyond traditional gate-based circuits and implementing data re-upload directly at the pulse-controlled level. This work pioneers a way to incorporate trainable parameters into the fundamental dynamics of quantum systems, rather than applying them as discrete gate operations. The team designed a superconducting transmon processor that leverages realistic noise profiles to simulate real-world conditions and serves as a platform for benchmarking this pulse-based model against gate-based models. This work involved formulating a pulse-based variant of data reupload, where input data is iteratively embedded into the system dynamics through precisely tailored control signals.
The researchers systematically increased the noise intensity to assess the resilience of their pulse-level implementation and carefully measured fidelity and performance under varying levels of decoherence and control errors. The pulse-based model consistently outperformed the gate-based approach, showing 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, increasing the efficiency of quantum computation. This work demonstrates that pulse-level implementations maintain high fidelity over longer periods of time and exhibits enhanced robustness to noise and errors, suggesting a viable path for practical quantum machine learning in the noisy mesoscale quantum era. This innovative methodology provides further freedom in information encoding and model optimization, potentially reducing the severity of barren plateaus and improving the structure of the optimization environment.
Pulse-based data reupload outperforms gate-based models
Scientists have achieved a major breakthrough in quantum machine learning by developing a pulse-based data re-upload model that outperforms traditional gate-based approaches on noisy intermediate-scale quantum (NISQ) hardware. This work introduces a method to embed trainable parameters directly into the dynamics of quantum systems, operating natively at the pulse-controlled level and bypassing the limitations of traditional gate-based circuits. The team formulated a pulse-based variant of data re-upload, enabling a hardware-aware approach to variational quantum machine learning. Experiments demonstrate that the pulse-based model exhibits consistently higher test accuracy and improved generalization under equivalent noise conditions compared to the gate-based model.
A systematic investigation of noise immunity revealed that pulse-level implementations maintain high fidelity over longer periods of time and have enhanced robustness to decoherence and control errors. Specifically, this study shows that the pulse-based approach maintains signal integrity and accuracy even as the noise intensity increases systematically, an important advantage for practical quantum computation. The team's methodology provides a general framework for converting the training of variational quantum machine learning models into pulse-level implementations, allowing the design of parameterized pulses for specific tasks. This approach avoids the rigid structure often imposed on pulse-level models and provides greater flexibility and expressiveness. By compressing the multi-gate sequence into a shorter pulse schedule, the team reduced the circuit depth and execution time, further improving the efficiency of quantum computation.
Pulse-based quantum machine learning outperforms gates
This work demonstrates a pulse-based implementation of a data re-upload model and provides a new approach to quantum machine learning in near-term hardware. By embedding trainable parameters directly into the control pulses that operate the quantum system, the team achieved a more efficient and hardware-tailored representation of the learning model, avoiding the limitations associated with traditional gate-based circuits. Benchmarks of superconducting processors with realistic noise profiles show that this pulse-based model consistently outperforms gate-based models, exhibiting both higher test accuracy and improved generalization capabilities under comparable noise conditions. The findings show that the pulse-level implementation maintains high fidelity over longer periods of time and is more resilient to decoherence and control errors, which are critical challenges in current quantum computing technologies. Although the simulations were limited by computational resources, the observed trends suggest that pulse-based models can maintain good test accuracy even with increased complexity, suggesting the potential for extending these architectures. This work establishes direct pulse-level optimization as a promising pathway toward scalable and experimentally executable quantum machine learning that integrates expressiveness, generalization, and robustness within a hardware-native framework.
