Design and Evaluation of CPU-, GPU-, and FPGA-Based Deployment of a CNN for Motor Imagery Classification in Brain-Computer Interfaces
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe presented manuscript is very well structured, and the experimental design appears to be adequate. Congratulations to the Authors for the work. In order to consider the present manuscript for the final pubblication, few issues need to be addressed:
1. Among the material and methods section, the EEG preprocessing, it could be helpful to read about the EEG rejected epochs percentage, since a visual inspection was used to remove artefacts.
2. I suggest the Authors to include how it was selected the time resolution for the EEG epoch segmentation.
3. The references provided in the Introduction regarding the BCIs state of the art appear slightly outdated. I would suggest to consider more recent related works.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper provides an insightful exploration of deploying CNNs for MI classification in BCIs across different hardware platforms, CPU, embedded GPU, and FPGA. The reduction in power consumption achieved by the FPGA sounds very exciting and indicates a promising direction for efficient BCI deployments.
However, the overall analysis and performance evaluation of this paper seems to be quite preliminary. The paper lacks a broader analysis of the trade-offs between accuracy, power consumption, and inference time across more varied tasks or conditions. More comprehensive comparison and detailed analysis of the computational cost, including memory usage and latency, on each hardware platform would be beneficial.
Limited dataset size (only 5 participants) may affect the generalizability of the findings. The experiment would benefit from a larger, more diverse participant pool to validate the findings across different demographic and neurological profiles. (It would be interesting to see the 29 dataset performance on the deployment)
The focus on a single type of neural network architecture may limit understanding of performance across different models. A more extensive comparison of different neural network architectures could provide a better understanding of how specific models perform on various hardware platforms.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsI think the revision enhanced the quality of the paper. I don't have any further comments.