kumoh national institute of technology
Networked Systems Lab.

Heidy Indrayani, Ade Pitra Hermawan, Sanjay Bhardwaj Rizki Rivai Ginanjar, Dong-Seong Kim, Jae-Min Lee, "Split Super-resolution for CSI Reconstruction using Transfer Learning in MIMO System", IEEE Wireless Communications Letter (R) (Resubmit Soon)
By : Ade Pitra
Date : 2020-09-19
Views : 70

Dear Prof. Jae-Min Lee:

The review of the referenced manuscript, WCL2020-1046, is now complete. I regret to inform you that based on the enclosed reviews and my own reading of your manuscript, I am unable to recommend its publication in IEEE Wireless Communications Letters.

As you can find below, the reviewers have strong concerns in the novelty and the contribution of this work. The reviewers also point out some important explanations and performance comparisons are missing.

Please note that unlike full paper transactions journals, a letter submission to IEEE Wireless Communications Letters does not undergo a major revision. A binary decision is made on the submission, either acceptance or rejection, by Editor. However, you are encouraged to revise and resubmit your manuscript.

Reviewer: 1

Comments to the Author
The paper investigates the research issue of CSI feedback with deep learning. The authors propose a transfer learning-based CSI decoder, in which a trained super-resolution CNN is used for CSI reconstruction. The proposal outperforms the existing scheme, CsiNet, in the indoor environment. However, the paper has several technical problems, which are listed below:
*The new idea is to use SRCNN with a split block. The novelty and contribution are too small.
*SRCNN is a poor SR method compared with state-of-the-art SR method, e.g., VDSR, PConv, DRRN, etc. The authors carefully discuss the advantage of SRCNN.
*The transfer method is not explained. What dataset is used to train SRCNN? If that is an image, the authors should elaborate on why there is a correlation between image and CSI.
*The authors mention, A small number of filters are used in each layer to achieve better performance against the existing system. However, the number of filters at decoders in CsiNet is less than the proposal, i.e., 24*2=58 in CsiNet (64+32+1)*2=194 in the proposal. The authors should carefully investigate the structure of NN. It would be better to show the number of total parameters (weights and biases) in both CsiNet and the proposal.
*Kernel size of the first convolution at the encoder is set to 1x3 while CsiNet utilizes 3x3 kernel. 1x3 kernel can not extract features of vertical directions. The authors should explain the reason.
*The performance evaluation of CsiNet is done in both indoor and outdoor environments. The authors might want to add performance evaluation in outdoor environments.
*There are many existing works on CSI compression with CNN. The authors should compare the proposal with the following methods.

Reviewer: 2

Comments to the Author
The authors proposed to use a super-resolution reconstruction method in image processing for CSI reconstruction, where the reconstruction process is based on convolutional neural network.

This work is easy to follow, but its technical contribution is not sufficient for a letter and the most important question, why this method improves the performance and how it works, is not answered. I would like to share the following comments:

1. SRTL is used to reconstruct the CSI information, but is it the only machine learning approach to get the performance? Are there any other learning algorithms better than it? Are there any specific structure of the SRTL-CSI making it unique for CSI construction?
2. According to the statement To strengthen the structure quality of each CSI, the input size of the CSI (2x32x32)is split into (1x32x32, 1x32x32) respectively. Why splitting can strengthen the structure quality?
3. The database was generated from [11]. What is the specific scenario? Does it have some patterns that are suitable for the SRTL-CSI structure? Is it an urban/suburban area? Do we have to modify the SRTL-CSI structure when the scenario changes?
4. The comparison is not sufficient. Please compare the proposed approach with the followings:
T. Wang, C. Wen, S. Jin and G. Y. Li, "Deep Learning-Based CSI Feedback Approach for Time-Varying Massive MIMO Channels," in IEEE Wireless Communications Letters, vol. 8, no. 2, pp. 416-419, April 2019
J. Kang, O. Simeone, J. Kang and S. S. Shitz, "Joint Signal and Channel State Information Compression for the Backhaul of Uplink Network MIMO Systems," in IEEE Transactions on Wireless Communications, vol. 13, no. 3, pp. 1555-1567, March 2014.