kumoh national institute of technology
Networked Systems Lab.

Godwin Brown Tunze, Thien Huynh-The, Jae-Min Lee, and Dong-Seong Kim, Sparsely Connected CNN for Efficient Automatic Modulation Recognition, IEEE Transaction on Vehicular Technology, 2020.(Major Revision)
By : Godwin
Date : 2020-09-08
Views : 103

Dear Prof. Kim:

The review of your paper

VT-2020-02150: Sparsely Connected CNN for Efficient Automatic Modulation Recognition

has been completed. Below please find comments from the reviewers.

Editor's Comments: Three independent reviewers evaluated the paper and pointed out some issues that need to be carefully addressed and/or clarified. In a possible resubmission, the authors must make sure that all the raised concerns have been addressed satisfactorily.

Based on the reviewers' comments, publication of the paper in its present form is not recommended. It is recommended that you resubmit your manuscript, WITHIN TWO MONTHS from the date of this email, as revised in accordance with the comments. If you fail to do so, your paper will be considered withdrawn from the review process.

Please note that this recommendation of resubmission should not be treated as an implication of eventual acceptance of your paper. It only refers to a deferral of the editorial decision to see whether the reviewers' comments could be addressed. In the event that your revision is accepted, the manuscript must be in a format that is in accordance with IEEE final submission requirements. For detailed final submission requirements, kindly refer to the information for IEEE Authors guidelines contained on the website http:/www.ieee.org/pubs/authors.html.

To submit your revision, please first log onto Manuscript Central ( https://mc.manuscriptcentral.com/tvt-ieee). In the Author Center, you will see the entry "Revised Manuscripts with Decisions" under "My Manuscripts". Click this entry and you will see an entry with the current paper ID included in the list "Manuscripts with Decisions" at the bottom of the screen. Then click the "Create a revision" button in the entry for this paper to start submitting your revision.

For your paper submission, please make sure your final manuscript (including main body, figures, tables, references) is in double-columned with fonts no smaller than 10pt in the final IEEE typesetting with maximum page length 16 pages for regular papers.

Additionally, you MUST submit a response file alongside you manuscript that details the changes you have made and addresses the issues raised by the reviewers. REVISION WITHOUT THIS FILE WILL BE REJECTED. The file must be added in Step 2 of the submission as "OPTIONAL SOURCE FILE - MAIN DOCUMENT."

The revised manuscript should follow our submission policy. The complete revised manuscript (including main body, figures, tables, references, and appendix) of a Regular Paper should be in general limited to 35 pages. If the revised paper does not comply our submission policy, it may be considered withdrawn without the review process.

As the system keeps all the files from your earlier submission(s) for this paper, please remember to delete any files which are not part of this new revised submission.

Sincerely yours,

Daniel Benevides da Costa
Editor, IEEE Transactions on Vehicular Technology

Reviews for VT-2020-02150

*Please note that there may be attached files that contain additional
comments from the reviewers. To see these files, please log on to
your "Author Center" at the Manuscript Central web site
( https://mc.manuscriptcentral.com/tvt-ieee) and select this paper to
see and download the files.

Reviewer: 1

This paper proposed a convolutional neural network (CNN) called SCGNet for low-complexity and robust modulation recognition in intelligent communication receivers. However, it still needs major modifications before acceptance.

1. There are a lot of language mistakes that need to be addressed. Please proofread the paper carefully.

2. Compare and cite the proposed algorithm against the latest work in the same domain, such as CSFL: A novel unsupervised convolution neural network approach for visual pattern classification , Face recognition: A novel un-supervised convolutional neural network method , A Benchmark Dataset and Learning High-level Semantic Embeddings of Multimedia for Cross-media Retrieval , Optimization of CNN through novel training strategy for visual classification problems, Unsupervised pre-trained filter learning approach for efficient convolution neural network.

3. The training of CNN based algorithm is very computational expensive. Justify, the reason behind author choice of CNN based algorithm for automatic modulation recognition.

4. Explain why the author choose multiple convolution layers simultaneously in their architecture. However, we are aware that convolution operator extracts low-level features in the initial layers and high-level features in higher layers. The benchmark CNN architecture follows alternatively convolution layers. Justify?

5. The experimental results in this paper are too simple and there is a lack of specific comparison of different methods. It is difficult to explain the superiority of the proposed SCGNet model. The experimental results give too few examples, and more comparisons and examples can be added.

6. Figure 1 about convolution is not very expressive. It is recommended to improve it to illustrate better the concept of convolution.

7. In my opinion, it will help a lot showing a simple example for every convolution type to illustrate it. The description of some convolution types is superficial (e.g., grouped convolution).

Reviewer: 2

This paper proposes a new convolutional neural network (CNN) called SCGNet for low-complexity and robust modulation recognition in intelligent communication receivers. The idea seems interesting. I have some suggestions:

1.- An initial table with all the acronyms could help to understand the manuscript.
2.- As the model is a sparse network, I recommend the authors to provide a state of art disscusion about Deep Efficient Models as: EfficientNet, ShiftResNet, TresNet, SEResNeXt50, MixNet, MobileNet, etc.
3.- Please, include a link to the code in the paper
4.- Please, include a deeper discussion about the FLOPs.

Reviewer: 3

1. If support vector machines (SVMs), then neural networks (NNs) not NN.
2. (i.e.., CNN with few connections) -> N (i.e., CNN with few connections).
3. (*) represents the convolution operation -> * represents the convolution operation

1. Fig. 1. the grouped convolutional layer has a similar feature as in Video Foreground Extraction Using Multi-View Receptive Field and Encoder–Decoder DCNN for Traffic and Surveillance Applications, IEEE Transactions on Vehicular Technology, 2019. It is recommended to discuss this work in the literature review.
2. Can you justify why the proposed SCGNet has lesser inference speed than DrCNN, although it has lower trainable parameters than of DrCNN?