Machine learning extends quantum data processing beyond current limits

A novel quantum error mitigation (QEM) framework addresses the key challenge of correcting noise in continuous variable (CV) quantum systems and degrading quantum states essential for quantum information processing. Jingpeng Zhang from the School of Physics at Sun Yat-sen University and colleagues developed the framework in collaboration with Qianchuan Zhao from the School of Automation at Tsinghua University and Jie Han from Sun Yat-sen University and the National Key Laboratory of Optoelectronic Materials and Technology. Continuously variable quantum systems utilize an infinite-dimensional Hilbert space, which offers advantages in encoding and manipulating quantum information compared to qubit-based systems, but they are particularly susceptible to environmental noise that limits coherence and fidelity. This noise arises from interactions with the surrounding environment, causing photon loss and dephasing, gradually destroying the quantum state.

Their work introduces an extrapolation approach that uses a time-conditional Swin Transformer to compensate for noise accumulation without the need for extensive training data covering the entire quantum evolution. This advance represents a sharp step toward practical noise mitigation in CV quantum systems and demonstrates accurate state recovery even in the long-time regime where traditional methods fail. Traditional QEM techniques often rely on noise characterization and application of complex error correction codes, which require extensive computational resources and precise knowledge of the noise process. Therefore, the ability to reduce errors with limited training data is critical to scaling up CV quantum systems and realizing their full potential.

Time-conditional machine learning modifies quantum states beyond training data

The error rate was reduced to 0.6%, a significant improvement over existing methods that typically fail when initial training parameters are exceeded. This breakthrough enables accurate quantum state recovery even when extrapolating beyond the data used to train the system. This is something that has not been previously achievable with continuous variable (CV) quantum computing. Naren Manjunath of the Perimeter Institute and colleagues at Sun Yat-sen University and Tsinghua University have developed the time-conditional Swin Transformer, a machine learning model that explicitly models how noise accumulates over time. Swin Transformers were originally developed for computer vision and are excellent at capturing long-range dependencies in data, making them suitable for modeling complex correlations that exist in quantum phase space. The “time-conditional” aspect allows the temporal evolution of quantum states to be incorporated into the model, which is important for accurately predicting and compensating for noise accumulation.

This model captures subtle long-range correlations in the phase space of quantum systems and enables more reliable error correction. Adaptive layer normalization proved to be key to modeling this accumulation and increasing reliability. Layer normalization is a technique used to stabilize the training process of deep neural networks, and its adaptive implementation allows the model to dynamically adjust internal parameters based on the time evolution of the quantum state. The quantum state was successfully recovered even in the presence of non-Markov noise, a complex form of environmental destruction. Markov noise assumes that future states depend only on the current state, whereas non-Markov noise has built-in memory effects that make it much more difficult to mitigate. Simulations reveal the framework’s ability to accurately reconstruct states on timescales that exceed those used for initial training. This has been a major limitation of previous machine learning-based quantum error mitigation techniques. This extrapolation capability is essential for performing longer and more complex quantum calculations. Although the framework achieved an impressive error rate of 0.6% in simulations, these results currently rely on idealized conditions and performance on real imperfect quantum hardware has not yet been demonstrated.

Reducing time-dependent errors with adaptive normalization in Swin Transformer

The time-conditional Swin Transformer forms the core of this advancement and is a machine learning model used to decipher patterns in complex data. Unlike previous methods, this model does not require exhaustive training data covering the entire quantum process. Instead, it learns how to correct for errors by explicitly considering how noise accumulates over time and embedding evolutionary time using adaptive layer normalization. The Swin Transformer architecture utilizes a hierarchical structure with shifted windows to enable efficient processing of high-dimensional data while capturing both local and global features. By tuning the model in time, the framework can effectively learn the dynamics of the noise and predict its impact on the quantum state at different points in time.

This allows the model to build “correction maps” that predict and reduce the effects of noise even beyond the initial training period. Numerical simulations were performed in a continuous variable quantum system, utilizing a number of qubits corresponding to a 10-dimensional phase space. In these systems, both photon loss and phase shift, common sources of environmental noise, occurred at varying intensities. Photon loss refers to the irreversible loss of photons from a quantum system, and phase shift refers to the loss of phase coherence, both of which contribute to the degradation of quantum information. The 10-dimensional phase space represents the range of possible quantum states and allows a comprehensive evaluation of the framework’s performance.

In our simulations, we used both Markov and non-Markov noise models to evaluate the strong resistance of our framework to different types of environmental damage. Previous methods required complete evolutionary data, a major experimental hurdle. The team chose a machine learning approach to reduce quantum errors, avoiding exhaustive data collection by focusing on time-dependent error correction. This alleviates the burden of extensive calibration required by existing methods, which is a major obstacle to scaling up quantum computers. Traditional calibration methods require characterizing the noise at every point in time, which is computationally expensive and time-consuming. By learning the dynamics of the noise, the framework can significantly reduce the amount of required calibration data.

Reducing data demands increases the durability of continuous variable quantum computations

A novel quantum error mitigation framework was designed to avoid the need for exhaustive data collection in continuous variable (CV) quantum systems. This is an important step toward building more powerful and reliable quantum computers. However, current approaches rely on simulation, raising questions about their performance on real-world quantum hardware, which is susceptible to unpredictable defects. Transitioning from simulation to real hardware presents significant challenges because real devices have manufacturing, control, and measurement imperfections that simulation cannot fully capture.

It is important to note that this framework currently relies on simulation rather than direct testing using physical quantum hardware. Despite this limitation, the development of quantum error mitigation techniques that require significantly less data represents a substantial advance. The team’s simulations investigated the framework’s behavior at different noise intensities and provided insight into its durability under different conditions. To optimize performance in real-world scenarios, it is important to understand the robustness of the framework to different noise levels.

This detailed analysis is critical to adapting the framework to the complexity of real quantum devices. By employing a time-conditional Swin Transformer, the framework accurately compensates for noise that accumulates during quantum computations and captures subtle correlations within quantum systems. There was no traditional method that matched this. A new approach has been established to reduce quantum errors in continuous variable systems that avoids the need for large amounts of training data. Future research will focus on implementing this framework in physical CV quantum systems and evaluating its performance in realistic experimental environments, paving the way towards more robust and scalable quantum computation.