Abstract We report a self-supervised deep learning framework for Rydberg sensors that enables single-shot noise suppression matching the accuracy of multi-measurement averaging. The framework eliminates the need for clean reference signals (rarely available in quantum sensing) by training on two sets of noisy signals with identical statistical distributions. When evaluated on Rydberg sensing datasets, the framework outperforms wavelet transform and Kalman filtering, achieving a denoising effect equivalent to 10000-set averaging while reducing the computation time by three orders of magnitude. We further validate performance across diverse noise profiles and quantify the complexity-performance trade-off of U-Net and Transformer architectures, providing actionable guidance for optimizing deep learning-based denoising in Rydberg sensor systems.
Abstract We report a self-supervised deep learning framework for Rydberg sensors that enables single-shot noise suppression matching the accuracy of multi-measurement averaging. The framework eliminates the need for clean reference signals (rarely available in quantum sensing) by training on two sets of noisy signals with identical statistical distributions. When evaluated on Rydberg sensing datasets, the framework outperforms wavelet transform and Kalman filtering, achieving a denoising effect equivalent to 10000-set averaging while reducing the computation time by three orders of magnitude. We further validate performance across diverse noise profiles and quantify the complexity-performance trade-off of U-Net and Transformer architectures, providing actionable guidance for optimizing deep learning-based denoising in Rydberg sensor systems.
We report a self-supervised deep learning framework for Rydberg sensors that enables single-shot noise suppression matching the accuracy of multi-measurement averaging. The framework eliminates the need for clean reference signals (hardly required in quantum sensing) by training on two sets of noisy signals with identical statistical distributions. When evaluated on Rydberg sensing datasets, the framework outperforms wavelet transform and Kalman filtering, achieving a denoising effect equivalent to 10,000-set averaging while reducing computation time by three orders of magnitude. We further validate performance across diverse noise profiles and quantify the complexity-performance trade-off of U-Net and Transformer architectures, providing actionable guidance for optimizing deep learning-based denoising in Rydberg sensor systems.
Microwave Kinetic Inductance Detectors (MKIDs) are superconducting sensors primarily used in astrophysics and quantum measurements for their energy- and time-resolution of single-photon detection. However, their sensitivity and precision are limited by various noise sources. Conventional denoising approaches enhance the signal-to-noise ratio (SNR) but often blur signal features and are ineffective with non-stationary noise. This work explores the use of machine learning (ML) denoising techniques for MKID signals in steady state resonators and in photon-induced transients. Different denoising techniques are evaluated on a dataset of frequency-domain S21 measurements and time-domain phase pulses with illumination of 660 nm, 780 nm, and 1310 nm wavelengths. Traditional filtering approaches (moving average, median, Wiener, Savitzky-Golay, wavelet, Kalman, total variation denoising) were evaluated in comparison with ML models including convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and gated recurrent units (GRUs). Due to absence of clean reference signals in experimental settings, a Noise2Noise training approach was used for self-supervised learning from noisy data. The models were assessed using mean squared error (MSE), signal-to-noise ratio (SNR), variation in peak depth and optical resolution metrics. Findings of this thesis show that machine learning models achieve substantially better denoising performance compared to conventional methods across all metrics. GRU models outperformed other models when denoising steady-state resonator signals, as they captured local features of the spectral response, and they also were the most effective at denoising transient photon signals due to their capacity to capture temporal dynamics. Conventional filters, while being computationally efficient, were found to distort the signal and were not as effective in retaining low-amplitude photon-induced features. In summary, this study demonstrates that machine learning denoising is a robust and versatile method for improving MKID signals without prior knowledge of noise properties. The results also demonstrate the opportunity to enhance superconducting detector readouts by combining data-driven approaches, leading to improved sensitivity and resolution for applications such as astrophysical imaging and quantum sensing.
Abstract The resurgence of self-supervised learning, whereby a deep learning model generates its own supervisory signal from the data, promises a scalable way to tackle the dramatically increasing size of real-world data sets without human annotation. However, the staggering computational complexity of these methods is such that for state-of-the-art performance, classical hardware requirements represent a significant bottleneck to further progress. Here we take the first steps to understanding whether quantum neural networks (QNNs) could meet the demand for more powerful architectures and test its effectiveness in proof-of-principle hybrid experiments. Interestingly, we observe a numerical advantage for the learning of visual representations using small-scale QNN over equivalently structured classical networks, even when the quantum circuits are sampled with only 100 shots. Furthermore, we apply our best quantum model to classify unseen images on the ibmq_paris quantum computer and find that current noisy devices can already achieve equal accuracy to the equivalent classical model on downstream tasks.
We present a systematic investigation of deep learning methods applied to quantum error mitigation of noisy output probability distributions from measured quantum circuits. We compare different architectures, from fully connected neural networks to transformers, and we test different design/training modalities, identifying sequence-to-sequence, attention-based models as the most effective on our datasets. These models consistently produce mitigated distributions that are closer to the ideal outputs when tested on both simulated and real device data obtained from IBM superconducting quantum processing units (QPU) up to five qubits. Across several different circuit depths, our approach outperforms other baseline error mitigation techniques. We perform a series of ablation studies to examine: how different input features (circuit, device properties, noisy output statistics) affect performance; cross-dataset generalization across circuit families; and transfer learning to a different IBM QPU. We observe that generalization performance across similar devices with the same architecture works effectively, without needing to fully retrain models.
The resurgence of self-supervised learning, whereby a deep learning model generates its own supervisory signal from the data, promises a scalable way to tackle the dramatically increasing size of real-world data sets without human annotation. However, the staggering computational complexity of these methods is such that for state-of-the-art performance, classical hardware requirements represent a significant bottleneck to further progress. Here we take the first steps to understanding whether quantum neural networks could meet the demand for more powerful architectures and test its effectiveness in proof-of-principle hybrid experiments. Interestingly, we observe a numerical advantage for the learning of visual representations using small-scale quantum neural networks over equivalently structured classical networks, even when the quantum circuits are sampled with only 100 shots. Furthermore, we apply our best quantum model to classify unseen images on the ibmq\_paris quantum computer and find that current noisy devices can already achieve equal accuracy to the equivalent classical model on downstream tasks.
In recent years, the Rydberg-atom sensor has attracted significant attention due to its high sensitivity, wide dynamic range, and miniaturization capabilities. However, noise from laser power fluctuations, frequency drift, and environment is introduced alongside the intrinsic quantum noise of Rydberg atoms, significantly degrading system performance for microwave sensing. This study tries to reduce the noise impact by combining the Rydberg-atom sensor with the convolutional neural network (CNN)-transformer model. The CNN module captures local temporal features of the signals, while the transformer module extracts global correlations within these features. We apply it to the following classification tasks of atomic microwave sensing: multifrequency microwave recognition, broadband spectrum monitoring, and automatic modulation recognition. We demonstrate that the CNN-transformer model achieves high accuracy across the three tasks and can effectively reduce the noise impact, which holds significant potential for applications in radar detection and wireless communication.