Transposed Convolution as Alternative Preprocessor for Brain-Computer Interface Using Electroencephalogram

Round 1

Reviewer 1 Report

= Please make sure that the Introduction covers the following points:

- What is a Brain-Computer Interface (BCI), and how does it work?

- What are the typical steps involved in a BCI using electroencephalograms (EEG)?

- Why are there many disordered combinations of feature extraction and classification methods in BCI?

- How can neural networks be applied to classification problems related to BCI?

- What is a transposed convolution, and how is it used as a preprocessor in this study?

- How does using a transposed convolution in the first layer of a convolutional neural network (CNN) affect BCI classification accuracy?

- What were the results of this study, and how did they compare to conventional preprocessing methods?

- What are some potential applications of using a two-dimensional CNN with transposed convolution as an alternative to preprocessing for BCIs?

- What are some potential limitations or challenges of using a transposed convolution in BCI, and how could they be addressed in future research?

= If possible, please include in the introduction the references:

- Robust averaging of covariances for EEG recordings classification in motor imagery brain computer interfaces, Neural Computation, Vol. 29, No. 6, pp. 1631 — 1666, June 2017 (DOI: 10.1162/NECO_a_00963)

- An experimental study to compare CSP and TSM techniques to extract features during motor imagery tasks, Chapter 3 of the contributed book ''Signal Processing and Machine Learning for Brain-Machine Interfaces'', pp. 41 — 60, The Institution of Engineering and Technology (IET), 2018 (ISBN: 978-1-78561-398-2)

- An introductory tutorial on brain-computer interfaces and their applications, Electronics, Vol. 10, No. 5, Article No. 560, February 2021 (DOI: 10.3390/electronics10050560)  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This work proposed that by adding an up-sampling deconvolution step before the CNN feature extraction and classification, the overall feature classification accuracy of BCI EEG signals can be improved. We have the following comments.

1.       Please explain in more detail about the transposed CNN. What are the limitations? Does it have to have some relationship with the following CNN like using similar kernels, learning rates, or spatial-temporal parameter arrangement, etc.

2.       It will be helpful for the readers to provide a detailed explanation about the principles or reasons why adding the transposed convolution step helps to improve the classification accuracy.

3.       What will be the cost to impact the effective speed of BCI operations? For example, influence on the Information Transfer Rate (ITR). A plot of the dependence of the ITR to the accuracy before and after adding the extra step will be helpful.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf