Multimodal Quanvolutional and Convolutional Neural Networks for Multi-Class Image Classification
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors1. The abstract lacks clarity in the contribution of the research. The author should briefly state the novel aspects of the hybrid quantum-classical neural networks (HNNs) and their importance in the image classification tasks.
2. In introduction section, include a brief discussion on the limitations or challenges of existing machine learning approaches in addressing complex classification tasks, thereby setting the stage for the need for novel architectures like HNNs.
3. Elaborate the class distribution within the datasets utilized in this research. How were the author addressing the potential bias in the dataset, because that may affect the evaluation of the HNNs?
4. Provide comprehensive details in to rationale behind the choice of hyperparameters for quantum processing. How do these choices influence the performance of the HNNs?
5. How does the interpolation technique used to address the size mismatch between the output of the quanvolutional layer and the input of the classical model affect the fidelity of the input data? Were any alternative approaches considered, and if so, what were their advantages and disadvantages?
6. What challenges or limitations are associated with using a quantum device as the first layer of a HNN, particularly in terms of computational complexity and scalability?
7. Provide architecture diagram for the proposed method with layer information for better understanding.
8. The manuscript provides a detailed description of the experimental setup and results, but include more mathematical formalism or equations to elucidate the underlying principles of the proposed hybrid neural network (HNN) architecture.
Author Response
Please, find our report as to corrections to our manuscript according to the reviewer's comments (many thanks for them!).
>> 1. The abstract lacks clarity in the contribution of the research. The author should briefly state the novel aspects of the hybrid quantum-classical neural networks (HNNs) and their importance in the image classification tasks.
FIXED, by extending abstract.
>> 2. In introduction section, include a brief discussion on the limitations or challenges of existing machine learning approaches in addressing complex classification tasks, thereby setting the stage for the need for novel architectures like HNNs.
FIXED, by adding paragraph to introduction section.
>> 3. Elaborate the class distribution within the datasets utilized in this research. How were the author addressing the potential bias in the dataset, because that may affect the evaluation of the HNNs?
FIXED, by extending "datasets" section in multiple places.
>> 4. Provide comprehensive details in to rationale behind the choice of hyperparameters for quantum processing. How do these choices influence the performance of the HNNs?
FIXED, by adding quantum circuit diagram and extending "Data pre-processing process" section.
>> 5. How does the interpolation technique used to address the size mismatch between the output of the quanvolutional layer and the input of the classical model affect the fidelity of the input data? Were any alternative approaches considered, and if so, what were their advantages and disadvantages?
FIXED, by extending "Hybrid Neural Networks" section and outlined impact of interpolation as a possible future research topic.
>> 6. What challenges or limitations are associated with using a quantum device as the first layer of a HNN, particularly in terms of computational complexity and scalability?
FIXED, by extending "conclusion" section.
>> 7. Provide architecture diagram for the proposed method with layer information for better understanding.
FIXED, by adding architecture diagram and extending "Hybrid Neural Networks" section.
>> 8. The manuscript provides a detailed description of the experimental setup and results, but include more mathematical formalism or equations to elucidate the underlying principles of the proposed hybrid neural network (HNN) architecture.
FIXED, by extending "Materials and Methods" section in multiple places with diagrams that illustrate architecture approaches that were used.
Thank you again for your valuable comments that helped us to improve the quality of the paper.
On behalf of all authors and with the best regards,
Yevhenii Trochun
Reviewer 2 Report
Comments and Suggestions for AuthorsThe study's focus on the potential advantages of using HNNs in image classification tasks can be considered a very interesting and valuable research.
However, there are some aspects of the study that need to be criticized.
1.The abstract provides a general description of how HNNs are used, but details of the methodology can be given.
2. Information such as how quantum and classical neural networks are brought together, how the training process is carried out, and on which features of the model are tested should be given.
3. More information can also be provided on how pre-trained models such as ResNet, EfficientNet, and VGG-16 are selected and used.
4. More details need to be provided regarding the diversity of datasets and the applicability of HNNs in real-world scenarios.
5. It would be helpful to provide more information about the limitations of the study and suggested directions for future research. In particular, more information can be provided about the conditions under which HNNs perform better and about potential studies to increase the real-world applicability of this technology.
Additionally, sources other than quality journals should be added to the study. (Exmaple:
https://www.sciencedirect.com/science/article/abs/pii/S0208521622000432
https://www.sciencedirect.com/science/article/pii/S2666285X21000558
)
Author Response
Please, find our report as to corrections to our manuscript according to the reviewer's comments (many thanks for them!).
>> 1. The abstract provides a general description of how HNNs are used, but details of the methodology can be given.
FIXED, by extending abstract.
>> 2. Information such as how quantum and classical neural networks are brought together, how the training process is carried out, and on which features of the model are tested should be given.
FIXED, by extending and improving wording of "Hybrid Neural Networks" section.
>> 3. More information can also be provided on how pre-trained models such as ResNet, EfficientNet, and VGG-16 are selected and used.
FIXED, by improving "Architectures selection" section.
>> 4. More details need to be provided regarding the diversity of datasets and the applicability of HNNs in real-world scenarios.
FIXED, by improving "Datasets" section.
>> 5. It would be helpful to provide more information about the limitations of the study and suggested directions for future research. In particular, more information can be provided about the conditions under which HNNs perform better and about potential studies to increase the real-world applicability of this technology.
FIXED, by extending "Conclusion" section.
Thank you again for your valuable comments that helped us to improve the quality of the paper.
On behalf of all authors and with the best regards,
Yevhenii Trochun
Reviewer 3 Report
Comments and Suggestions for AuthorsThe research's focus on hybrid quantum-classical neural networks (HNNs) is timely, given the growing interest in leveraging quantum computing's potential to solve complex computational problems more efficiently.
Strengths:
Innovative Approach: The manuscript successfully introduces an innovative approach to image classification by integrating quantum computing into neural networks, which is a novel area of research with significant potential.
Comprehensive Analysis: The study provides a comprehensive analysis of the HNNs' performance across different datasets, including CIFAR100 and Satellite Images of Hurricane Damage, which demonstrates the versatility of the proposed approach.
Practical Application: The application of HNNs to real-world datasets and comparison with classical models provide valuable insights into the practical viability and potential advantages of quantum-enhanced machine learning models.
Areas for Improvement and Suggestions:
Methodological Transparency: While the manuscript outlines the overall approach, more detailed descriptions of the quantum operations, the specific quantum circuits used, and their integration with classical layers would enhance the reader's understanding and reproducibility of the research.
Comparative Analysis: The comparison between HNNs and classical neural networks (CNNs) could be expanded. Specifically, a deeper analysis of cases where HNNs underperform or outperform classical models would offer valuable insights into the strengths and limitations of quantum computing in this context.
Limitations and Challenges: The discussion could benefit from a more detailed examination of the limitations and challenges associated with implementing HNNs, including quantum hardware limitations, the scalability of quantum operations, and the implications of noise and error rates in quantum computations.
Future Directions: While the conclusion mentions future research directions, a more detailed discussion on potential improvements in quantum layer design, optimization techniques for HNNs, and exploration of other quantum operations could provide a clearer roadmap for future research.
Author Response
Please, find our report as to corrections to our manuscript according to the reviewer's comments (many thanks for them!).
>> Methodological Transparency: While the manuscript outlines the overall approach, more detailed descriptions of the quantum operations, the specific quantum circuits used, and their integration with classical layers would enhance the reader's understanding and reproducibility of the research.
FIXED, by adding diagrams of quantum circuit and HNN architecture together with explanations in sections "Data pre-processing process" and "Hybrid Neural Networks".
>> Comparative Analysis: The comparison between HNNs and classical neural networks (CNNs) could be expanded. Specifically, a deeper analysis of cases where HNNs underperform or outperform classical models would offer valuable insights into the strengths and limitations of quantum computing in this context.
FIXED, by extending "Complex HNNs based on reference classical models" section.
>> Limitations and Challenges: The discussion could benefit from a more detailed examination of the limitations and challenges associated with implementing HNNs, including quantum hardware limitations, the scalability of quantum operations, and the implications of noise and error rates in quantum computations.
FIXED, by extending "Conclusion and Future Work" section.
>> Future Directions: While the conclusion mentions future research directions, a more detailed discussion on potential improvements in quantum layer design, optimization techniques for HNNs, and exploration of other quantum operations could provide a clearer roadmap for future research.
FIXED, by extending "Conclusion and Future Work" section.
Thank you again for your valuable comments that helped us to improve the quality of the paper.
On behalf of all authors and with the best regards,
Yevhenii Trochun
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThe corrections requested in the previous revision have been made.
However, when Figure 7 is examined, the performance graphs appear to be seriously inaccurate.
The performance values of the proposed Quantum+CNN model are seen to be low and it does not contribute to the literature. Also, the bottom-left subgraph in this graph is not acceptable. This model needs to be re-examined. It should also be reproduced in other graphs.
In addition, the graphics in figure 8, figure 9, figure 10, figure 11, figure 12 and figure 11 should be recreated. If there aren't errors in the presentation of this study, it is clearly seen that the Qauntum+CNN model does not contribute to the performance. This needs to be re-examined.
You can also get samples from the sources I recommend for loss and accuracy charts.
Author Response
Thank you again for your valuable comments.
We added the following text to the subsection that contains the figures in question:
"Despite the lower maximum accuracy metrics and higher number of epochs needed for HNNs for reaching top classification accuracy, compared to the classical reference models, as shown on Fig.7 - Fig. 14, HNNs can still have major advantages comparing to a classical NN. Main advantages of HNNs are better power efficiency and significantly higher computation speed provided by quantum hardware. These advantages can be highly beneficial and provide significant value even in light of lower accuracy metrics in a wide number of applications.
In this subsection, we studied the impact of dataset size, dataset complexity and HNN structure on the accuracy metric and compared it to the metrics shown by the reference classical models. The main goal of this subsection is not to find HNN configuration that will surpass the accuracy metric of classical NN but to study the impact of various parameters on HNNs performance."
We hope that this will help to better illustrate our intent behind this part of our research.
On behalf of all authors and with the best regards,
Yevhenii Trochun
Round 3
Reviewer 2 Report
Comments and Suggestions for AuthorsIt is necessary to explain in more detail why HNN is used. As seen in the graphs, it seems that it contributed to the study at least a little, but it was not at an acceptable level. His contribution to the literature should be more emphasized.
Performance parameters (f1-score, precision, recal etc.) results should also be added. These parameters are included in the resource I recommend adding. Additionally, new resources can be added. For example, I recommend a resource on CNN (https://www.sciencedirect.com/science/article/pii/S0208521622000432).
Comments on the Quality of English LanguageThere does not appear to be a serious problem with the language of the study. But it is recommended to read it again.
Author Response
Thank you for your valuable comments!
Your comments are insightful and very appreciated (many thanks for them again!). We will study performance parameters of HNN in greater detail in our subsequent research (including recal, precision, f1-score and cross-validation metrics).
Our current research is more of a feasibility study of the approach of building HNNs with quantum device as a first layer of neural network.
Thank you for proposing relevant to our research article, we included it to our reference list.
Kind regards,
Yevhenii Trochun
Round 4
Reviewer 2 Report
Comments and Suggestions for AuthorsThe revisions requested in the previous evaluation were made in the Manuscript. In these revisions, the performance of the Quantum+CNN model is below the Classic CNN model, although it is not a serious contribution to the literature, it is considered appreciable. Acceptable.
