Explainable Deep Learning: A Visual Analytics Approach with Transition Matrices
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors1. The authors must clearly present their contribution and motivation in introduction section.
2. Missing literature review and analysis of the literature review.
3. The result will be improved by comparing it with related work.
4- In conclusion, need to present the problem solved by this manuscript.
Comments on the Quality of English Language
need to be improved
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe work is nice with the idea of constructing a transition matrix between the feature spaces of DL and ML models as formal and mental models, for improving the explainability of separating hyperplanes for classification tasks.
My main concern is the involved two tasks (MNIST, and Irs) are all focusing on image classification, which can be interpreted by many methods such as CAM, and SHAP analysis. Thus what is the main advantage of using a transition matrix for such tasks? And can the methodology be extended to explain regression tasks?
Comments on the Quality of English LanguageThe language is fine.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors propose an approach to improve the explainability of convolutional (CNN) models. This method is weakly validated by two chosen data sets IRIS and MNIST.
On the other hand, the proposed method is only applied to convolutional neural networks. For this reason, I suggest the authors modify the title since they do not present conclusions or experimentation for other deeplearning models: transformers, recurrent networks, etc.
The authors do not compare the proposed method with others already existing in the literature and only apply it to two data sets that are not complicated.
The empirical validation is limited, making it challenging to assess the effectiveness of the proposed method. To establish credibility, conduct extensive experiments across diverse datasets or scenarios. Present comparative studies against existing approaches, indicating statistical significance, and demonstrating the robustness and superiority of the proposed method.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsIts better to improve the result.
Comments on the Quality of English Language
Acceptable
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for Authors
As I suggested in my first review, I believe that the application of only two data sets is insufficient and incomplete. More examples and more complexity are needed to evaluate the goodness of the method.
On the other hand, as the authors comment in their review report if their method can be applied to other deep-learning models, it would also be good for the study if the authors add some of these examples. I encourage authors to include this idea.
As far as I am concerned, these suggestions are essential to validate the method and accept the paper.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 3
Reviewer 3 Report
Comments and Suggestions for AuthorsThe article has improved considerably with the changes made by the authors and I propose that it be accepted for publication in this journal
Author Response
Dear Reviewer,
Thank you for your valuable feedback and suggestions. We have carefully revised the manuscript to address the issues you raised.
We have thoroughly proofread the entire manuscript, with a particular focus on the revised sections. All typos you pointed out have been corrected, including the ones at line 424 (now "three" instead of "two") and line 461 (now "efficacy" instead of "efficiency"). Additionally, we have meticulously checked for any other typographical or grammatical errors throughout the text. All revised parts in the text are highlighted in green.
Next, we have comprehensively expanded the discussion on the limitations of our proposed approach, as per your suggestion. We have thoroughly addressed all the issues raised by the reviewers, providing detailed explanations and acknowledging the potential limitations and areas for further improvement.
Finally, in the revised limitations section (lines 791-816), we have explicitly discussed the importance of the quality of the input deep learning models and the requirement for separability between classes in the learned feature representations. We have also highlighted the dependence of our visual analytics techniques on the model's ability to accurately represent data with distinct class separations.
We believe that these revisions have addressed all the concerns raised by the reviewers and have significantly strengthened the manuscript. We appreciate your valuable feedback, which has helped us improve the quality and clarity of our work.