What is my quantum computer good for? Quantum capability learning with physics-aware neural networks

Summary

The paper introduces a novel quantum-physics-aware neural network (qpa-NN) architecture for quantum capability learning. The model achieves error reduction in capability prediction on both experimental and simulated data.

Strengths

1. The qpa-NN architecture incorporates quantum physics principles, which provides a new perspective of designing models in the field of learning-based quantum capability learning. 2. The demonstrated reduction in mean absolute error over CNN-based models is commendable.

Weaknesses

1. The reviewer's main concern is regarding the qubit scale of the dataset. However, the dataset used in the experiments, which consists of maximum 5 qubits, appears too small both from the perspective of current quantum hardware and classical simulations. The reviewers are curious to know the reason behind the inability to collect data on systems with more qubits (either experimental data or simulated data). Is it due to the high difficulty of experimental deployment, or is it because the current models struggle to train on larger datasets? 2. The evaluation method mentioned in the paper, Process Fidelity (Eq. 3), may have scalability issues. Although the authors attempted to explain in Section 3.2 how some approximations can be used to calculate Eq. 3 relatively efficiently, this description is hard to follow. For instance, can the proposed approximation method avoid exponential computational and storage complexity? Additionally, the reviewer hope that in the revised version, the authors can provide a discussion on the differences between the approximations designed in this paper and the estimation methods in [Proctor et al., 2022], or if the approximations in this paper are merely direct adaptations of the latter's estimation methods.

Questions

How does the proposed architecture handle scalability challenges as the number of qubits increases?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors have discussed the limitations.

Summary

This paper introduces a neural-network-based architecture for quantum capability learning. The idea is to utilize the architecture to predict rates of errors in quantum circuits. The authors compared their qpa-NN with previous CNN-based method and elucidated their improved performance.

Strengths

1. qpa-NN outperforms the previous CNN method in predicting circuit success probability.

Weaknesses

1. Given that neural network-based methods for predicting circuit success probability have been previously discussed in the literature, the novelty of the approach is not sufficiently convincing. 2. The explanation of how the qpa-NNs leverage graph structures could benefit from further elaboration to enhance clarity. 3. The scope of the considered noise types is limited, and the authors have benchmarked their results exclusively on small-scale devices.

Questions

1. Have the authors compared their qpa-NN with the so-called “stability baseline model”, (which was mentioned to be better than CNN-based method Hothem et al. [2023c]) ?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

There does not seem to be negative social impact of this theoretical research.

Summary

The paper presents an approach to improve the state of the art in quantum capability learning, which is the task to predict the prowess for error when running a specific quantum algorithm given a fixed quantum computer. The tackle this, the authors introduce some specializations on (graph) neural networks that fit the nature of quantum computers and their errors especially well. The new approach yields better results than purlely CNN-based methods on a specific synthetic and empirical data.

Strengths

The approach is very interesting from both a quantum and a machine learning point of view. The issue of quantum capability learning is described in geat detail and the innovation of the approach becomed clear. In the method to intertwine neural network architectures with their subject of training, especially when that subject is a quantum computer, I see great potential for further investigation.

Weaknesses

To be frank, I consider the issue of quantum capabilty learning (especially as it is motivated in this paper) a rather artificial one, given the quite small number of potentially useful quantum algorithms. And I think this leads to some rather weak arguments, mostly in the motivation of this paper. However, quantum capability learning (as the authors examine it in this paper) is still an interesting and relevant topic. I suggest cutting down on practical promises and focusing on the pure challenge of predicting a quantum computer's behavior using classical neural networks, which I consider a sufficient motivation for the study of this topic. While the performance-based analysis has good results, I would have wished for a more in-depth take that can actually clarify the speculation posed in the introduction: "Our qpa-NNs' improved performance is likely largely due to their improved ability to model the impact of coherent errors..."

Questions

None.

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

No comments.