Approximate Computing-Based Processing of MEA Signals on FPGA

Round 1

Reviewer 1 Report

This paper proposes an approximate computing hardware system for the signal processing of microelectrode arrays (MEAs). The proposed system is implemented on FPGA. There are multiple approximate modules proposed to reach a trade-off between accuracy, area, power, and delay. The experiments also prove the claim. However, there are some concerns needed to be addressed:

 

1.     The novelty needs to be clarified more. It is mentioned that the approximate computing is not widely used in the field of MEA signal processing, but the way that the proposed adders is design is not novel enough. Is the novelty only about the application? Please discuss more about this.

2.     The motivation needs to be explained more. This work claims that the proposed system can provide optimal performance gains in area, power consumption, and latency. It is easy to understand that low latency is important in such signal processing. But are area and power the timely issues for such system? If not, the system can be simply improved by add more parallel processing units without the need of the proposed approximate computing methods?

3.     In the experiments, please discuss more about the test samples. Do these samples cover enough scenarios?

4.     For the full system evaluation, different adders are evaluated under different mode, but is there any guidance of choosing which adder for the full system (since it is not necessary to implement all of them in one system for different modes)?

5.     In the experiments, modes 3 and 4 have relatively low accuracy when using the proposed adders. Will this accuracy loss affect the final decision on the data from the MEA system? If so, how to improve it?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this work, approximation computing on FPGA is used to construct signal processing algorithms, and the accuracy, latency, and area of these algorithms are compared to those of the accurate system. After testing the system with genuine biological signals, the authors assessed how well the three approximation adders worked. The work is good, but the following are some recommendations for enhancements.

1. Several of the chapters are too lengthy. For example, the evaluation results of approximation adders might be utilized with condensed explanations. Some graphs are so evident that additional explanation is not necessary. The majority of the graphs may also be linked together by chapter to improve reading.

2. This paper's concluding chapter is neither comprehensible nor clear. The chapter that wraps up the study should go over a couple of other significant findings.

3. The paper's creativity might be improved. Though it can lower latency and power use when added to FIR filters, its scientific value is not obvious.

4. Some more work could be added to this paper. For instance, it would be wise to compare several unlisted approximation adders.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper investigates the use of approximate arithmetic for the processing of data acquired by micro-electrode arrays, which is a novel contribution. The paper, however, has several weaknesses as detailed below.

 

1) The considered detection algorithm is extremely simplified: just a comparison with a fixed threshold. This approach is likely to give unsatisfactory results in applications with reduced signal to noise ratio. It is true that an hardware implementation requires simple algorithms, but some spike detection techniques requiring very low hardware resources have been presented in literature (see below). These techniques should be referenced and possibly compared with the simple approach proposed in the paper.

 

[A] Liu, Z. et al. “A Novel Algorithm for Online Spike Detection”. In Proceedings of the Materials Science, Engineering and Chemistry Conference (MATEC), Nanjing, China, 24–26 May 2018; Volume 173, p. 02017.

 

[B] Saggese, Gerardo, et al. "Comparison of Sneo-Based Neural Spike Detection Algorithms for Implantable Multi-Transistor Array Biosensors." Electronics 10.4 (2021): 410.

 

[C] Zhang, Z. et al. “Adaptive Spike Detection and Hardware Optimization towards Autonomous, High-Channel-Count BMIs” J. Neurosci. Methods 2021, 354, 109103.

 

[D] Saggese, G. et. al “A Low Power 1024-Channels Spike Detector Using Latch-Based RAM for Real-Time Brain Silicon Interfaces”. Electronics 2021, 10, 3068

 

 

2) The dataset used is an actual neural data recording. Therefore, this dataset does not include a ground truth to compare with to estimate the accuracy of detection. However, the evaluation results show 100% accuracy for exact adder in Tab. 6. Please comment on this.

 

 

3) The implementation of the FIR filter (fig. 4) requires 5 multipliers. A multiplier (while simplified since one of the operands is constant) is much more complex than an adder, but this point is completely overlooked in the paper. The advantages / disadvantages / improvements related to the use of approximate multipliers should in the considered application should be clearly stated in the paper. There are several simple approximate multipliers proposed in literature - some of them are very simple to be described and implemented (an example is referenced below).

 

[E] A. G. M. Strollo et al. "Approximate Multipliers Using Static Segmentation: Error Analysis and Improvements," in IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 69, no. 6, pp. 2449-2462, June 2022

 

 

4) There are several approximate full adders proposed in Literature. The CPredA seems ineffective, since the does not consider the possibility of propagation between carry-in and carry-out. Lower-part OR (where the carry is discarded altogether and the sum is computed as the OR of A, B can also be investigated):

 

[F] Y. Guo, H. Sun, and S. Kimura, “Design of power and area efficient lower-part-OR approximate multiplier,” in IEEE Region 10 Conference. IEEE, 2018, pp. 2110–2115.

 

 

5) The results show advantages mainly in terms of speed and area. However, for the considered application the advantage in speed is irrelevant since the data-rate is in the order of tens of Kilo Hertz. Thus, reducing the maximum delay from, say, 6ns to 3ns, is not useful. This should be clearly stated in the paper. It should also be stressed that this kind of investigation would be more useful in the case of full-custom implementation (maybe, for implantable devices), where power reduction is the highest concern a a significant power gain can be expected by using approximate arithmetic circuits.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The author answered my previous questions well. There is one reminding minor issue:

The author claim "The proposed processing system uses approximate computing to provide optimal performance gains". Please try to avoid using the word "optimal" since it is too absolute (It is like "the best", which is hard to be proved).

Author Response

Manuscript ID: electronics-2151292

Type of manuscript: Article

Title: Approximate Computing Based Processing of MEA Signals on FPGA

Authors: Mohammad Hassan, Falah Awwad *, Mohamed Atef, Osman Hasan

 

 

Dear Editor,

We are grateful to the referees for their valuable comments, which helped us enhance the quality of our manuscript. Our responses to the referees' comments are shown below and the changes are highlighted in the manuscript.

 

Referee 1:

We thank the referee for the careful review of our manuscript and for the great efforts.

The author answered my previous questions well. There is one reminding minor issue: The author claim "The proposed processing system uses approximate computing to provide optimal performance gains". Please try to avoid using the word "optimal" since it is too absolute (It is like "the best", which is hard to be proved).

The word “optimal” is modified through the whole manuscript.   The whole manuscript is also revised for spelling and grammar mistakes.   Also, all references have been checked for their relevance to the contents of the manuscript.   All changes are highlighted in the revised version.

 

Author Response File: Author Response.pdf

Reviewer 3 Report

My suggestions have been implemented in the revised version of the paper

Author Response

Manuscript ID: electronics-2151292

Type of manuscript: Article

Title: Approximate Computing Based Processing of MEA Signals on FPGA

Authors: Mohammad Hassan, Falah Awwad *, Mohamed Atef, Osman Hasan

 

 

Dear Editor,

We are grateful to the referees for their valuable comments, which helped us enhance the quality of our manuscript. Our responses to the referees' comments are shown below and the changes are highlighted in the manuscript.

 

Referee 3:

We thank the referee for the careful review of our manuscript and for the great efforts.

My suggestions have been implemented in the revised version of the paper.

The whole manuscript is revised for spelling and grammar mistakes.   Some paragraphs are also enhanced.   Also, all references have been checked for their relevance to the contents of the manuscript.   All changes are highlighted in the revised version.

 

Author Response File: Author Response.pdf