⚛️ quantum physics

Pulsed learning for quantum data re-uploading models

This paper proposes a pulse-based variant of data re-uploading for quantum machine learning that embeds trainable parameters directly into system dynamics, demonstrating superior accuracy, generalization, and noise resilience compared to traditional gate-based models on simulated superconducting hardware.

Original authors: Ignacio B. Acedo, Pablo Rodriguez-Grasa, Pablo Garcia-Azorin, Javier Gonzalez-Conde

Published 2026-07-23

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Ignacio B. Acedo, Pablo Rodriguez-Grasa, Pablo Garcia-Azorin, Javier Gonzalez-Conde

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to recognize patterns, like telling the difference between a cat and a dog. In the world of quantum computing, this is called Quantum Machine Learning (QML). For a long time, scientists have tried to do this by building circuits out of "gates," which are like rigid, pre-made Lego bricks. You snap them together to build a machine that learns. But there's a problem: the quantum computers we have right now are noisy and fragile, like a house of cards in a windstorm. When you stack too many of these rigid Lego bricks together, the noise messes everything up, and the robot forgets what it was supposed to learn.

Recently, researchers have realized that maybe we shouldn't be building with Lego bricks at all. Instead of snapping together pre-made blocks, what if we could speak the computer's native language directly? Think of it like this: instead of telling a musician to play a specific note from a sheet of music (a gate), you give them a continuous stream of instructions on exactly how to move their fingers, how hard to press the keys, and how long to hold the note (a pulse). This approach, called "pulse control," lets you sculpt the sound of the music in real-time, rather than just stacking pre-recorded clips. The big question is: if we stop using the rigid Lego bricks and start sculpting with the raw sound waves, can we build a smarter, tougher robot that doesn't fall apart when the wind blows?

This paper explores exactly that idea. The authors, a team of researchers from Spain, decided to take a popular quantum learning method called "data re-uploading" and rebuild it from the ground up using these raw pulses instead of standard gates. In their "data re-uploading" model, the computer is fed the same piece of information over and over again, layer by layer, to help it understand the pattern better. Usually, this is done by stacking layers of gates. But here, the authors replaced every single gate with a custom-designed microwave pulse, the actual signal that talks to the quantum computer's hardware.

They tested this new "pulse-native" model on a simulated version of a real quantum computer (specifically, one that uses superconducting transmon qubits, which are like tiny, super-cooled circuits). They pitted their new pulse-based robot against the old gate-based robot using four different puzzles: distinguishing between the numbers 0 and 8, sorting types of flowers, and solving some tricky geometric shapes.

The results were quite promising. In these simulations, the pulse-based model turned out to be a much better student. First, it learned faster and needed fewer layers to get the job done. While the old gate-based model started to "overthink" and memorize the training data (a problem called overfitting) when the puzzles got deep, the pulse model stayed focused and generalized better to new data. Second, and perhaps most importantly, the pulse model was incredibly tough. When the researchers cranked up the "noise" to simulate a very messy, error-prone environment, the gate-based robot quickly gave up and started guessing randomly. The pulse-based robot, however, kept its cool, maintaining useful accuracy even when the noise was strong enough to break the other model.

The authors suggest that this happens because speaking the computer's native language allows for a more efficient and robust way of processing information. By avoiding the rigid steps of standard gates, the pulse model can navigate the noisy quantum world more smoothly. While these results are currently based on simulations and limited to two-qubit systems (due to the heavy computing power needed to simulate these complex pulses), the findings point toward a future where quantum machine learning might work much better if we stop treating quantum computers like standard computers and start treating them like the unique, wave-based machines they actually are.

Technical Summary: Pulsed Learning for Quantum Data Re-uploading Models

Problem Statement
Quantum Machine Learning (QML) on Noisy Intermediate-Scale Quantum (NISQ) hardware faces significant hurdles, primarily stemming from the limitations of Variational Quantum Circuits (VQCs). While data re-uploading has emerged as a promising strategy to enhance model expressivity with shallow circuits, standard gate-based implementations suffer from trainability issues, noise sensitivity, and the risk of overfitting. Furthermore, recent analyses suggest that certain gate-based architectures offer no inherent advantage over classical methods once their structure is scrutinized. A critical gap exists between theoretical pulse-based advantages and scalable, hardware-compliant QML implementations, as many current approaches merely reparameterize fixed gate templates rather than designing native control strategies.

Methodology
The authors propose a "pulse-native" framework that translates the data re-uploading paradigm directly to the control pulse level, bypassing the abstraction of discrete quantum gates.

Hardware Model: The study focuses on superconducting transmon qubits, utilizing the native language of the hardware: microwave pulses. The system is modeled using a driven effective Hamiltonian where control parameters include pulse amplitude ( $\Omega$ ), phase ( $\gamma$ ), and frequency detuning ( $\Delta\omega$ ).
Pulse-Based Ansatz:
- Single-Qubit Operations: Instead of decomposing operations into $R_X, R_Y, R_Z$ gates, the model uses resonant microwave pulses for $XY$-plane rotations and "Virtual Z" (VZ) gates for $Z$ -rotations. The trainable parameters for a single-qubit block are the pulse amplitude, phase, and the angles of the VZ rotations.
- Two-Qubit Interactions: Rather than compiling complex entangling gates (like CNOT) from multiple pulses, the authors employ a single microwave pulse applied to one qubit at the frequency of another, inspired by the Cross-Resonance (CR) gate. This interaction is parameterized by the pulse's amplitude, phase, and frequency detuning, allowing the model to dynamically modulate entanglement strength.
- Architecture: The model implements a data re-uploading Quantum Neural Network (QNN) where input data is repeatedly embedded. The variational layers are replaced by these parameterized pulse sequences.
Training Strategy: The study utilizes a "warm-start" initialization strategy. A single-qubit pulse model is trained first; its optimized parameters are then used to initialize the single-qubit blocks of a two-qubit model, while entangling parameters start at zero. This ensures the cost function begins at the optimal single-qubit value.
Simulation Setup: The models were benchmarked on a simulated superconducting transmon processor using realistic noise profiles extracted from the IBM Brisbane device. The simulations were restricted to two-qubit systems due to computational costs. Four binary classification tasks were used: MNIST (digits 0 vs. 8), Iris (versicolor vs. virginica), and two synthetic datasets (Corners and Helix).

Key Contributions

General Methodology: The paper establishes a framework for converting any gate-based variational QML architecture into a pulse-level implementation, treating physical pulse attributes as trainable parameters rather than reparameterizing fixed gate templates.
Hardware Alignment: By operating directly at the pulse-control level, the approach avoids compilation overhead and leverages the native dynamics of transmon qubits, including frequency detuning and anharmonicity.
Empirical Benchmarking: The work provides a rigorous comparison between pulse-based and gate-based data re-uploading models under realistic noise conditions, demonstrating that pulse-native architectures can outperform their gate-based counterparts.

Results
The numerical simulations yielded three consistent findings across the four classification tasks:

Improved Generalization: The pulse-based model achieved comparable or superior test accuracy with fewer re-uploading layers compared to the gate-based model. Crucially, the pulse model exhibited a delayed onset of overfitting; while gate-based models showed a widening train-test gap at higher depths (particularly on the Helix and MNIST datasets), the pulse model maintained a closer correspondence between training and testing performance.
Enhanced Noise Resilience: Under increasing depolarizing noise, the pulse-based model retained useful classification accuracy at noise strengths where the gate-based model degraded to near-random performance. This advantage was observed on both training and test curves, suggesting a structural robustness to noise rather than merely a regularization effect.
Consistency: These advantages were observed consistently across diverse datasets with different geometric decision boundaries, indicating the findings are not artifacts of a specific dataset.

Significance and Claims
The authors position pulse-native architectures as a viable and hardware-aligned path toward practical QML on current NISQ devices. They argue that by compressing multi-gate sequences into shorter pulse schedules and optimizing directly within the device's native dynamics, the approach offers a more favorable trade-off between expressivity and robustness. The paper claims that this methodology allows for the design of problem-specific ansätze that make more efficient use of limited coherence times and reduce calibration overhead.

The authors remain modest regarding the scope of their work, acknowledging that their simulations were limited to two-qubit systems due to computational constraints and that the encoding stage remained gate-based. They suggest that future work should focus on scaling to larger systems, developing specialized optimizers for the potentially rugged loss landscapes of pulse-parameterized models, and extending the pulse-level adaptation to the data encoding stage.

Technical Summary: Pulsed Learning for Quantum Data Re-uploading Models

More like this