Impact of quantum noise on the training of quantum Generative Adversarial Networks

K. Borras, Su Yeon Chang, Lena Funcke, Michele Grossi, Tobias Hartung, Karl Jansen, D. Kruecker, Stefan Kühn, Florian Rehm, Cenk Tüysüz, S. VallecorsaJournal of Physics Conference Series 2021

TL;DR

This paper investigates how quantum noise, specifically readout and gate errors, affects the training of quantum Generative Adversarial Networks (qGANs) for a simplified high-energy physics application. They find that while qGANs can be trained with moderate readout noise, two-qubit gate errors significantly hinder convergence, highlighting the critical need for error mitigation techniques on current noisy quantum hardware.

What Problem Does This Paper Address?

The advent of Noisy Intermediate-Scale Quantum (NISQ) devices presents a significant challenge for quantum machine learning (QML). These devices are inherently susceptible to various sources of quantum noise, such as errors in measurements (readout errors) and imperfect quantum operations (gate errors). This noise can degrade the performance of quantum algorithms, making it difficult to achieve reliable results. This paper specifically addresses how these intrinsic quantum noise sources impact the training process of quantum Generative Adversarial Networks (qGANs), a promising QML model for generating data distributions. Understanding these effects is crucial for determining the feasibility of using NISQ devices for QML tasks and for developing effective strategies to overcome noise.

Methods

The researchers employ a hybrid quantum-classical qGAN model, consisting of a classical discriminator and a quantum generator implemented as a variational quantum circuit. They simulate this model on a classical computer using IBM's Qiskit framework, introducing controlled levels of quantum noise.

Key methods include:

  • Noise Simulation: They simulate two primary types of noise:
  • Readout Errors: Modeled as uncorrelated bit-flip probabilities for each qubit, where a measured '0' is flipped to '1' and vice versa with a probability 'p'.
  • Two-Qubit Gate Errors: These are also incorporated, though the paper focuses more heavily on readout errors initially.
  • Hyperparameter Optimization: They investigate the impact of various hyperparameters (generator learning rate, discriminator learning rate, and exponential decay rate) on training performance under different noise levels.
  • Error Mitigation Techniques: They evaluate two methods for mitigating readout errors:
  • Conventional Bit-Flip (BF) Mitigation: A method that requires significant computational resources for larger qubit systems.
  • Independent Bit-Flip (IBF) Mitigation: A more resource-efficient method that assumes uncorrelated multi-qubit readout errors.
  • Performance Metric: The Kullback-Leibler (KL) divergence is used to quantify the difference between the target data distribution and the distribution generated by the qGAN. Lower KL divergence indicates better performance.

Datasets

The paper uses a simplified high-energy physics (HEP) use case: reproducing electron energy profiles in HEP calorimeters. The "dataset" here refers to the target probability distribution representing these energy profiles, which is binned into $2^n$ pixels, where $n$ is the number of qubits in the generator. No external datasets are explicitly mentioned or used.

Key Insights and Ideas

  • Noise as a Double-Edged Sword (Initially): For very low levels of readout noise (p < 0.06), the qGAN performance was surprisingly comparable to, or even slightly better than, the noise-free case. This hints at a potential regularization effect, similar to artificial noise injection in classical deep learning.
  • The Criticality of Gate Errors: While qGANs can tolerate some readout noise, the introduction of even small two-qubit gate errors (around 1.5%) severely disrupts training, preventing convergence. This underscores that gate fidelity is a major bottleneck for NISQ-era QML.
  • Hyperparameter Sensitivity: The generator's learning rate was found to have the most significant impact on training stability and performance, especially as readout noise levels increased. This highlights the importance of careful hyperparameter tuning in noisy quantum environments.
  • Error Mitigation is Essential: For higher readout error rates (p ≥ 0.06), error mitigation techniques (both BF and IBF) are crucial for achieving performance comparable to the noise-free case.
  • Training Instabilities Persist: The inherent instability of GAN training (fluctuations in performance) was not significantly alleviated by increasing the number of training repetitions or by the presence of readout noise.

Key Results

  • Readout Noise Impact: qGANs can be trained with readout error probabilities up to approximately 6% with performance comparable to noise-free training. Beyond this threshold, performance degrades significantly without mitigation.
  • Error Mitigation Effectiveness: Both BF and IBF mitigation methods successfully improved qGAN performance in the presence of readout noise, bringing it closer to the noise-free baseline. The IBF method is noted for its computational efficiency.
  • Gate Error Impact: With a readout error rate of 2.5% and a two-qubit gate error rate of 1.5%, qGAN training failed to converge. This demonstrates the severe impact of gate errors.
  • Hyperparameter Importance: The generator learning rate was identified as the most influential hyperparameter for qGAN training under noise.

Limitations and Future Work

  • Limitations:
    • The study is based on classical simulations of noisy quantum devices, not actual hardware.
  • The use case is a simplified HEP problem, and results may not generalize directly to more complex tasks.
  • The paper acknowledges that training instabilities are a persistent challenge that was not fully overcome.
  • Future Work:
    • Extending the study to train qGANs on real quantum hardware.
  • Investigating more advanced mitigation techniques to address common training instabilities.
  • Exploring the impact of quantum noise and error mitigation on other QML models beyond qGANs.

Summary updated

Abstract

Abstract Current noisy intermediate-scale quantum devices suffer from various sources of intrinsic quantum noise. Overcoming the effects of noise is a major challenge, for which different error mitigation and error correction techniques have been proposed. In this paper, we conduct a first study of the performance of quantum Generative Adversarial Networks (qGANs) in the presence of different types of quantum noise, focusing on a simplified use case in high-energy physics. In particular, we explore the effects of readout and two-qubit gate errors on the qGAN training process. Simulating a noisy quantum device classically with IBM’s Qiskit framework, we examine the threshold of error rates up to which a reliable training is possible. In addition, we investigate the importance of various hyperparameters for the training process in the presence of different error rates, and we explore the impact of readout error mitigation on the results.
BibTeX
@article{borras2021impact,
  author = {K. Borras and Su Yeon Chang and Lena Funcke and Michele Grossi and Tobias Hartung and Karl Jansen and D. Kruecker and Stefan Kühn and Florian Rehm and Cenk Tüysüz and S. Vallecorsa},
  title = {{Impact of quantum noise on the training of quantum Generative Adversarial Networks}},
  journal = {Journal of Physics Conference Series},
  year = {2021},
}
Figure