A Lightweight Deep Learning Model for Automatic Modulation Classification Using Residual Learning and Squeeze–Excitation Blocks
Round 1
Reviewer 1 Report
The authors proposed a light weight deep learning based automatic modulation classification scheme and tested for two datasets. The following points needs to be addressed during review:
1. In contributions QMA64 is mentioned. It is QAM-64.
2. The mathematical modeling/analysis of residual learning and squeeze excitation network needs to be added.
3. As metioned in the simulations, the proposed network misclassifes 16-QAM to 64-QAM. Present standards prefer higher order QAMs like 256, 1024 etc. In such case, proposed model fails. Justify.
4. The authors are expected to add higher order QAMs in simulations and record classification accuracies.
5. The proposed work needs to be compared with recent works.
6. The choice of parameters needs to be justified.
Author Response
- In contributions QMA64 is mentioned. It is QAM-64.
Thank you for the careful reading. The typo has been corrected.
Refer to page: 2, line 64.
- The mathematical modeling/analysis of residual learning and squeeze excitation network needs to be added.
Thank you for the insightful comment. The mathematical modeling of residual learning and squeeze and excitation network have been updated in the revised manuscript.
Refer to section 3.1 and 3.2 (Equations 1 – 9, Pages: 6 – 7, Lines: 224 – 279).
- As mentioned in the simulations, the proposed network misclassifies 16-QAM to 64-QAM. Present standards prefer higher order QAMs like 256, 1024 etc. In such case, proposed model fails. Justify.
Thank you for your comment. Authors agree with the reviewer’s concern that the proposed network misclassifies 16-QAM as 64-QAM. However, we would like bring this in reviewer’s knowledge that the dataset chosen for this study i.e., RML 2016.10A has been identified by several researchers for the close distance constellation diagrams of QAM16 and QAM64. A detailed discussion has been added in the revised manuscript.
Refer to the section 4.4 ( Page: 13 Lines: 462 – 482).
- The authors are expected to add higher order QAMs in simulations and record classification accuracies.
Thank you for your suggestion. Currently there is lack of availability of a public dataset containing very high ordered QAM schemes. Furthermore, as mentioned above, we plan to test the performance of the proposed model on various communication dataset in future.
- The proposed work needs to be compared with recent works.
Thank you for your comment. A thorough comparison of proposed work with recent studies has been included in the updated manuscript.
Refer to Table 2 and Table 5 (Pages: 11 and 12).
- The choice of parameters needs to be justified.
Thank you for the comment. To justify the choice of several hyperparameters of the model, we present the results of the ablation study. The objective was to obtain the combination of the residual, squeeze-excitation and fully connected layers, that results in highest accuracy at the expense of lowest number of parameters. In particular, 5 models with varying number of residual, squeeze-excitation and fully connected layers along with variations in size of kernels, number of filters, type of subsampling were tested. The models in the ablation study and their structure have been shown in Table 3, whereas the corresponding performance of each of the model in terms of accuracy, number of parameters and its weight has been shown in Table 4.
Refer to Table 3 and Table 4 (Pages: 11 – 12).
Reviewer 2 Report
In the manuscript, a lightweight AMC network based on Residual Learning and squeeze-and-excitation architecture is proposed. The proposed method is logical and clarity. Some concerns are as following:
1. Add contrast with AMC method based on ResNet in Figure 5 and 6.
2.Add contrast with some lightweight AMC method with pruning algorithm in table 2.
Author Response
In the manuscript, a lightweight AMC network based on Residual Learning and squeeze-and-excitation architecture is proposed. The proposed method is logical and clarity. Some concerns are as following:
- Add contrast with AMC method based on ResNet in Figure 5 and 6.
Thank you for the comment. The results in Figures 5 and 6 have been updated. Relevant details and papers related to ResNet have been cited and included in comparative results.
Refer to Figs. 5 and 6 (Pages: 10-11).
- Add contrast with some lightweight AMC method with pruning algorithm in table 2.
Thank you for your comment. In the revised manuscript a comparison with the pruned algorithms for modulation classification have been added.
Refer Table 2 (Page: 10).
Reviewer 3 Report
This paper presented a deep learning algorithm for AMC inspired by the residual learning which has remarkable accuracy and great representational ability. A squeeze and excitation network is employed to improve performance. This paper is well written, but its novelty is very limited.
Comments:
1. The authors only introduced the residual learning networks utilized in AMC. The residual networks has already been adopted by AMC in the existing works. The novelty and the contributions of the proposed AMC method should be reclarified in detail.
2. The theoretical model and derivation of the proposed AMC method cannot be found in the paper.
3. The authors claimed that “Results show that the proposed model outperforms the existing methods in terms of accuracy and has up to 72.5% less parameters than the convolutional neural network designs.” However, the theoretical analysis of the computational complexity can not be found in this paper.
4. According to the simulation results, the accuracy of proposed AMC is still lower than the conventional FB-based AMC methods.
5. This paper should be carefully proofread.
Author Response
This paper presented a deep learning algorithm for AMC inspired by the residual learning which has remarkable accuracy and great representational ability. A squeeze and excitation network is employed to improve performance. This paper is well written, but its novelty is very limited.
Comments:
- The authors only introduced the residual learning networks utilized in AMC. The residual networks has already been adopted by AMC in the existing works. The novelty and the contributions of the proposed AMC method should be reclarified in detail.
Thank you for your comment. In the revised manuscript, the contributions and novelty of the proposed model has been elaborated in detail along with the mathematical models.
Refer to section 3.1 and 3.2 (Equations 1 – 9, Pages: 6 – 7, Lines: 224 – 279).
- The theoretical model and derivation of the proposed AMC method cannot be found in the paper.
Thank you for the comment. The mathematical derivations of the proposed method have been added to the revised manuscript.
Refer to section 3.1 and 3.2 (Equations 1 – 9, Pages: 6 – 7, Lines: 224 – 279).
- The authors claimed that “Results show that the proposed model outperforms the existing methods in terms of accuracy and has up to 72.5% less parameters than the convolutional neural network designs.” However, the theoretical analysis of the computational complexity can not be found in this paper.
Thank you for your comment. The reduction in the number of parameters was achieved by the choice of layers present in the model. The combination of residual and squeeze-excitation network and their configurations led to an increase in the accuracy at the expense of a smaller number of parameters. The updated manuscript also presents the ablation study that has been performed to justify the choice of number of residual stacks, configuration of squeeze-excitation network and several other parameters including number of fully connected layers and type of subsampling.
Refer to Table 3 and Table 4 (Pages: 11 – 12).
- According to the simulation results, the accuracy of proposed AMC is still lower than the conventional FB-based AMC methods.
Thank you for your careful reading. For the revised manuscript, we reperformed all the experiments in order to develop the ablation study. The revised experiments resulted in a better configuration of the model which shows superior classification accuracy than the contemporary FB methods. It is noteworthy that study was aimed to achieve comparable accuracy at the expense of less number of parameters to make it suitable for resource constrained devices. Table 2 clearly shows that the number of parameters in the proposed method are significantly less than contemporary FB methods. Refer to Section 4.2, Table 2 (Page: 10).
- This paper should be carefully proofread.
Thank you for the suggestion. The revised manuscript has been carefully proofread.
Round 2
Reviewer 2 Report
No other comments.
Author Response
We thank the reviewer for their careful consideration and valuable suggestions for our work.
Reviewer 3 Report
This paper presented a deep learning algorithm for AMC inspired by the residual learning which has remarkable accuracy and great representational ability. A squeeze and excitation network is employed to improve performance. The authors have majorly revised the paper. However, the novelty is still low.
Comments
1. Although this paper claims a new DL-based AMC method, the discussion in this paper mainly focuses on the residual learning network. What features of the signal are extracted for the network? The mathematical model of the signal features should be clarified.
2. The performance of the proposed algorithm should be compared to the traditional FB-based AMC approaches in accuracy and computational complexity.
Author Response
1. Although this paper claims a new DL-based AMC method, the discussion in this paper mainly focuses on the residual learning network. What features of the signal are extracted for the network? The mathematical model of the signal features should be clarified.
We would like to thank the reviewer for raising this concern. The architecture of the proposed design is such that it tries to learn the most representative features while enforcing a classification constraint. To emphasize this point in more understandable fashion, in the revised manuscript we provided the visualization for the tSNE embedding of the proposed model. Moreover, the motivation of using residual and squeeze and excitation blocks, and their feature extraction capabilities have been highlighted in the revised manuscript.
Please refer to Section 3.2.1, Pages 8-9, Lines 323-342.
2. The performance of the proposed algorithm should be compared to the traditional FB-based AMC approaches in accuracy and computational complexity.
Thank you for your comment. In the revised manuscript, a comparison of the proposed method with the traditional FB based methods have been added.
Please refer to Figure. 7, in Section 4.2, Pages 10-11, Lines 388-393.
