Decoding Visual Motions from EEG Using Attention-Based RNN

Round 1

Reviewer 1 Report

The authors investigate four major points of a) using four visual stimuli movements, b) using attention mechanism in conjunction of raw EEG data, c) using data augmentation methodology to increase limited data, d) using multi-trail classification.

I believe the authors have been successful to present systematically their four major points. In my opinion the investigation of using data augmentation is the most important part of the authors’ investigation. They show the feasibility of generating and using such data in training of shallow and deep networks where ERC, randomly concatenation, shows the best performance. I do not believe the results indicate significant improvement, see Figure 9, but it shows it works. This will pave the way to use new ambiguity functions in t.ex. ERC approach.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This research falls into a very popular domain that relates to EEG signal recognition/classification for various kinds of purposes. Various methods can be used for the purpose of signal classification, they range from some statistical methods, can use Bayes-based approach up to some AI methods (that includes Deep Learning). The Authors decided to go the Deep Learning way and use RNN for their classification/detection method. So, although this paper is in some part kind of "one of many" (with respect to the methodology, which is: record a signal, classify and check if the results are correct), then this paper is still relevant and will be of scientific interest. 

The original contribution, as indicated by Authors (lines 17-19), is a novel online data augmentation method randomly averaging EEG sequence to generate non-existing signal (based on the existing one) that, after merging with the existing samples, provides enough data for precise signal classification (many classification methods need certain amount of data to do their job and if there is not enough data, some alternative methods must be used). 

When it comes to some criticism which I have regarding this paper it is mainly located in the paper structure. This paper has no Related work section as such, the Introduction section contains quite a few references but this is not done the way the Related work section is usually constructed. So there is no review of similar solutions and explaining how this research relates to those other solutions. Having no such section results in things like in line 127, where the Authors state out of the sudden "Similar to previous research [8,29] (…)". I have initially thought that these references are other Authors' publications but to my surprise that was not the case. These were completely different publications and I have no clue how they relate to the Authors' solution. Having commented on this in Related work section would probably help a lot. So my first recommendation is:

**Recommendation 1 - try to extract Related work section from the Introduction section if possible or add Related work section to help to understand how this research is placed compared to other researchers' work. I understand that this may be my personal preference to have Related work section in a paper, but I think this would quite substantially improve this paper quality. So, although this is not a must, I would strongly recommend to do it.

The next thing is about references. Authors mention LSTM but somehow fail to reference in line 180 their own publication [18] in this specific context as if they have not published a related work while they did it. Not self-citing in such case raises concerns regarding re-using some previous work and forces the review to have a review the other publication to make sure this is not the case. In this case it was not the case but my next recommendation would be:

**Recommendation 2 - please go through the references and if there is any own previous work that relates to the current research, it should be explained what the relation is. Having Related work section would solve the problem as this is perfect place to properly refer all other pieces of work (by Authors' or others) that relate to the current work.

Other than that, I have no serious concerns regarding the paper. It is written in very good and comprehensible English, I have only found one little mistake that would need to be fixed:

• Line 165 - there is "Date preprocessing" whilst it surely should be "Data preprocessing".

Maybe one more proof reading would reveal some more of such simple mistakes so I would recommend to do it just one more time.

The results are presented in a clear way, they are commented on. The conclusions are a bit brief but summarise the research.

So, I would recommend to accept the paper after considering (and potentially, implementing) my remarks.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The article is found to have no novelty involved. The proposed methodology seems to exist already and have been published by many researchers. Hence I suggest the authors come up with some novel ideas and results.

Apart from that, the article involves too complex sentences and words. Hence I suggest the authors to use simple words so that the reading community will be comfortable. If these queries are addressed and the article may be revised, then the article may be considered for processing.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

I recommend the article may be accepted