Quantum Vision Transformers for Quark–Gluon Classification

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors propose a hybrid quantum-classical vision transformer architecture to distinguish between quark-initiated from gluon-initiated ejections. There are several aspects of the paper that could be improved:

 

1、

In the comparison with VIT, only the small gap in AUC of the model is described. It is hoped to add more detailed information such as model size and weight in the comparison with VIT.

2、

The author uses the CMS Open Data dataset for training and testing. The diversity of the datasets used should be discussed to improve the applicability of the model.

3、

The authors lack a detailed comparative analysis of model complexity. There is no in-depth understanding of the trade-offs in model computational cost and performance.

4、

Lack of mathematical support. Adding mathematical formulas can provide precise expressions, allowing readers to better understand the principles and working mechanisms of the model.

 

The authors introduce a novel quantum-classical hybrid method, and although the performance is slightly inferior to the traditional model VIT, the results are still encouraging. The model proposed by the authors provides a feasible direction for future research and requires further research and improvement to achieve better performance and applications.

Comments on the Quality of English Language

The authors propose a hybrid quantum-classical vision transformer architecture to distinguish between quark-initiated from gluon-initiated ejections. There are several aspects of the paper that could be improved:

 

1、

In the comparison with VIT, only the small gap in AUC of the model is described. It is hoped to add more detailed information such as model size and weight in the comparison with VIT.

2、

The author uses the CMS Open Data dataset for training and testing. The diversity of the datasets used should be discussed to improve the applicability of the model.

3、

The authors lack a detailed comparative analysis of model complexity. There is no in-depth understanding of the trade-offs in model computational cost and performance.

4、

Lack of mathematical support. Adding mathematical formulas can provide precise expressions, allowing readers to better understand the principles and working mechanisms of the model.

 

The authors introduce a novel quantum-classical hybrid method, and although the performance is slightly inferior to the traditional model VIT, the results are still encouraging. The model proposed by the authors provides a feasible direction for future research and requires further research and improvement to achieve better performance and applications.

Author Response

% REVIEWER 1

We thank the referee for carefully reading the manuscript and providing useful feedback. The revised version here has been modified in accordance with the referees' comments and suggestions. 

Comment: 1. In the comparison with VIT, only the small gap in AUC of the model is described. It is hoped to add more detailed information such as model size and weight in the comparison with VIT.

Response: We thank the reviewer for this suggestion. In "Section 2.2. Model," we have clarified the size of the quantum circuits employed, and we have also extended "Section 3. Results" with a discussion of the ROC curve.

Comment: 2. The author uses the CMS Open Data dataset for training and testing. The diversity of the datasets used should be discussed to improve the applicability of the model.

Response: We thank the referee for the suggestion. We have significantly expanded section 2.1 with the description of the dataset used in the analysis.

Comment: 3. The authors lack a detailed comparative analysis of model complexity. There is no in-depth understanding of the trade-offs in model computational cost and performance.

Response: The objective was to compare the hybrid and classical architectures at similar complexities. We commented on this in the abstract and also added text on page 5 where the number of hyperparameters is being computed.

Comment: 4. Lack of mathematical support. Adding mathematical formulas can provide precise expressions, allowing readers to better understand the principles and working mechanisms of the model.

Response: We thank the reviewer for this suggestion.  We have expanded the introduction with a brief review of the basic theoretical concepts of quantum computing, including mathematical definitions of the quantum operations applied.

Comment: The authors introduce a novel quantum-classical hybrid method, and although the performance is slightly inferior to the traditional model VIT, the results are still encouraging. The model proposed by the authors provides a feasible direction for future research and requires further research and improvement to achieve better performance and applications.

Response: We agree, and thank the referee for appreciating the novelty and potential significance of this line of research.

Reviewer 2 Report

Comments and Suggestions for Authors

1.       Abstract:

·       The abstract does not highlight the innovation of this research, and how it contributes to the academic society.

·       What problem is this research aiming to solve, and how is it different from what has been done previously? For instance, what are the challenges in the available Quantum Vision Transformers? How does this research solve this challenge?   

2.       Introduction:

·       Line 18-19: The purpose of quantum machine learning is not limited to “using models that are partially or fully executed on a quantum computer by replacing some subroutines of the models with quantum circuits.” This is not true; on the contrary, the purpose of quantum machine learning is to apply the unique properties of quantum mechanics to enhance the capabilities of “traditional” machine learning algorithms, tackling complex systems, and exploring new directions, to name a few – please see:  https://doi.org/10.1103/PhysRevLett.122.040504

·       Why were quantum vision transformers applied? What limitations do vision transformers have? And how do quantum vision transformers bridge the gap?    

3.       Method:

·       Please provide a schematic diagram summarizing the approach and/or steps followed.

·       Please highlight how this research is innovative and how it is different from

·       How was the diagram inspired by Vaswani et al. [13] and Dosovitskiy et al. [11]? Is this a new model different from both? How is it different from both? And what advantages does it have over both?

·       For reproducibility and repeatability, please provide which SDK you used (e.g., IBM QISKIT or Google Cirq), the number of qubits, what VQA you applied, etc.

·       How sample sizes for training were chosen as follows, what best practices were followed:

Distribution	Percentage
Training	76%
Validation	9%
Testing	15%
Data Set	100%

 

·       After training, why didn’t you add RoC and discuss if the system was overfitted or underfitted / biased and what precautions/approaches were applied to ensure reliable training?

4.       Conclusion:

·       The conclusion needs to be rewritten to reflect all changes, emphasizing on the novelty and the contribution of this research.   

Comments on the Quality of English Language

The English language is fairly adequate. 

Author Response

% REVIEWER 2

We thank the referee for carefully reading the manuscript and providing useful feedback. The revised version here has been modified in accordance with the referees' comments and suggestions. 

Comment: 1. Abstract
· The abstract does not highlight the innovation of this research, and how it contributes to the academic society.
· What problem is this research aiming to solve, and how is it different from what has been done previously? For instance, what are the challenges in the available Quantum Vision Transformers? How does this research solve this challenge?  

Response: We thank the referee for the suggestions. The abstract has been expanded by taking these suggestions into account. The changes are highlighted in the revised version.

Comment: 2. Introduction:
· Line 18-19: The purpose of quantum machine learning is not limited to “using models that are partially or fully executed on a quantum computer by replacing some subroutines of the models with quantum circuits.” This is not true; on the contrary, the purpose of quantum machine learning is to apply the unique properties of quantum mechanics to enhance the capabilities of “traditional” machine learning algorithms, tackling complex systems, and exploring new directions, to name a few – please see:  https://doi.org/10.1103/PhysRevLett.122.040504
· Why were quantum vision transformers applied? What limitations do vision transformers have? And how do quantum vision transformers bridge the gap?    

Response: We thank the referee for the suggestions. The introduction has been revised by taking into account these suggestions. The mentioned paper has also been consulted and added as a reference. The changes are highlighted in the revised version.

Comment: · Please provide a schematic diagram summarizing the approach and/or steps followed.

Response: Figures 3 and 4 contain three diagrams that depict the different components of the proposed system.

Comment: · Please highlight how this research is innovative and how it is different from

Response: We think the last question got cut off and is unavailable to us from the reviewer's report. In any case, in the revised version we expanded the abstract, introduction and conclusions to better explain the new elements of this research.

Comment: · How was the diagram inspired by Vaswani et al. [13] and Dosovitskiy et al. [11]? Is this a new model different from both? How is it different from both? And what advantages does it have over both?

Response: The illustration is inspired by the one presented in Vaswani et al. [13] and Dosovitskiy et al. [11] in that the drawing style is the same. In Dosovitskiy et al. [11] they also say “The illustration of the Transformer encoder was inspired by Vaswani et al. (2017).”. We modified the figure caption of Fig. 3 to clarify this.

Comment: · For reproducibility and repeatability, please provide which SDK you used (e.g., IBM QISKIT or Google Cirq), the number of qubits, what VQA you applied, etc.

Response: The SDKs used (JAX, Flax, TensorCircuit) are mentioned in the last paragraph of section 2.2 Model.

Comment: · How sample sizes for training were chosen as follows, what best practices were followed:
Distribution Percentage
Training     76%
Validation    9%
Testing      15%
Data Set    100%

Response: We thank the reviewer for this question. The sizes are the same as used in Andrews et al. This is now mentioned in the revised version. 

Comment: · After training, why didn’t you add RoC and discuss if the system was overfitted or underfitted / biased and what precautions/approaches were applied to ensure reliable training?

Response: We thank the reviewer for this question. This discussion is now included in the revised version.

Comment: · The conclusion needs to be rewritten to reflect all changes, emphasizing on the novelty and the contribution of this research.   

Response: We thank the reviewer for the suggestion. The conclusion has been revised by taking into account the suggestions from all three reviewers.

Reviewer 3 Report

Comments and Suggestions for Authors

According to abstract Authors presents a hybrid quantum-classical vision transformer architecture which employs variational quantum circuits to compute the required matrices in the attention calculations, as well as in the multi-layer perceptron’s.

In general, paper seems to be very interesting, but in current very short version I cannot recommend its publication. Authors provided the source code for numerical experiments, so it is possible to reproduce the numerical results, which is a convincing argument for the high evaluation of the article.

However, in reviewed paper because of its length many aspects are omitted Additionally, the manuscript in current form lacks quality in terms of scientific presentation. I will recall one of missed description connected with variational quantum circuits shown at Fig .4. Authors should consider also other form of such circuits because generated states from circ at Fig.4 may be in different form that data used under process of learning. But without analysis in paper, we do not have enough information about correctness of such circuits. This aspect is neglected by Authors, and it is difficult to score the final result of accuracy. Some other theoretical definitions of used data and quantum transformations as unitary and measurement operations are also omitted. Paper does not hold detailed information about structure and model of proposed hybrid solution which try to connect quantum and classical machine learning techniques.

For above reason, lack of many scientific details, I cannot recommend publication of the manuscript in the current form.

Author Response

% REVIEWER 3

We thank the referee for carefully reading the manuscript and providing useful feedback. The revised version here has been modified in accordance with the referees' comments and suggestions. 

Comment: According to abstract Authors presents a hybrid quantum-classical vision transformer architecture which employs variational quantum circuits to compute the required matrices in the attention calculations, as well as in the multi-layer perceptron’s. In general, paper seems to be very interesting, but in current very short version I cannot recommend its publication. Authors provided the source code for numerical experiments, so it is possible to reproduce the numerical results, which is a convincing argument for the high evaluation of the article.

Response: We thank the reviewer for the positive evaluation of the paper. We have expanded the paper in the revised version to provide additional details and information.

Comment: However, in reviewed paper because of its length many aspects are omitted Additionally, the manuscript in current form lacks quality in terms of scientific presentation. I will recall one of missed description connected with variational quantum circuits shown at Fig .4. Authors should consider also other form of such circuits because generated states from circ at Fig.4 may be in different form that data used under process of learning. But without analysis in paper, we do not have enough information about correctness of such circuits. This aspect is neglected by Authors, and it is difficult to score the final result of accuracy. Some other theoretical definitions of used data and quantum transformations as unitary and measurement operations are also omitted. Paper does not hold detailed information about structure and model of proposed hybrid solution which try to connect quantum and classical machine learning techniques.

Response: We thank the reviewer for these comments. We have expanded the introduction with a brief review of the basic theoretical concepts of quantum computing, including mathematical definitions of the quantum operations applied. We have also clarified that the size of the VQCs used should coincide with the size of the layers in the neural network in order to avoid the possible incompatibility between sizes mentioned in the comment. Finally, we have also added more information about the size of the quantum circuits used. 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Good to see the revised version  

Author Response

We thank the referee for the positive feedback. No changes needed.

Reviewer 2 Report

Comments and Suggestions for Authors

The approach can be further improved should the machine learning is further elaborated showing layers structure. 

Comments on the Quality of English Language

English is fairly adquate 

Author Response

We thank the referee for the feedback. In the revised version we have modified the figure caption of Fig. 4 which shows the model architecture.

Reviewer 3 Report

Comments and Suggestions for Authors

Authors given some additional remarks into reviewed paper and they answered on comments given at the first review. Although the paper is still quite short, but in my opinion introduced changes improved the final score of paper, therefore the reviewed paper may be accepted for publication.

Author Response

We thank the referee for the positive feedback. No changes were requested.

Round 3

Reviewer 2 Report

Comments and Suggestions for Authors

In the figure caption, the word inspired is still unclear. The authors need to clearly describe, whether this is a diagram made by them based on the approaches of the cited papers or was directly taken from them. If the diagram was made by them, then the authors need clearly state which part of the illustrated approach in the diagram is related to the first cited article (i.e., Vaswani), and which is related to the second one (i.e., Dosovitskiy).      

Comments on the Quality of English Language

The English Language is fairly acceptable.  

Author Response

We thank the referee for the suggestion. We have reworded the figure caption - taking out the reference to Vaswani et al and clarifying the difference from Dosovitskiy et al.

Reviewer 3 Report

Comments and Suggestions for Authors

Authors given some additional remarks into reviewed paper and they answered on comments given at the first review. Although the paper is still quite short, but in my opinion introduced changes improved the final score of paper, therefore the reviewed paper may be accepted for publication.

Author Response

We thank the referee for the positive feedback. No changes were requested.