Design of a Configurable Spike-Encoding Circuit Based on Focal Plane Array

Round 1

Reviewer 1 Report

This study presented the concepts related to spiking neural networks (SNNs) and their potential applications in artificial intelligence, which highlights the challenges and innovations associated with using SNNs, particularly in the context of artificial vision systems.

While the approach you described for interfacing with spiking neural networks (SNNs) using spike-coded readout has several advantages, there are also potential drawbacks and challenges associated with this approach. Here are a few considerations:

1. Noise Sensitivity: Spiking neural networks can be sensitive to noise due to their event-driven nature. The conversion of analog signals to spike signals might introduce some noise, which could affect the accuracy and reliability of the neural network's processing.

2. Complexity of Circuitry: Designing and implementing circuits that can accurately convert continuous analog signals into spike-coded representations can be complex and resource-intensive. The more intricate the circuitry, the more difficult it might be to ensure its proper functionality and reliability.

3. Limited Precision: Spike coding inherently reduces the precision of the input data. Analog-to-digital conversion typically allows for higher precision, while spike-based representation involves discretizing the data into binary spike events. This might impact the network's ability to distinguish fine-grained differences in input data.

Author Response

Sincere appreciation for your insightful comments on our manuscript. We have revised the manuscript based on your comments and provided detailed responses to each of them in the attachment. We trust these responses will help you locate these revisions in the manuscript more quickly.

Author Response File: Author Response.docx

Reviewer 2 Report

 

The manuscript presents a circuit capable of providing a spike-coded readout data from an array of infrared detectors at the pixel level, which is then sent for processing to spiking neural network. The manuscript presents two pulse encoding methods together with circuit implementation. The circuit can be configured to switch between temporal and frequency encoding. The readout circuit eliminates the use of conventional Analog to Digital converters.

I have the following comments for the authors, which may improve further the manuscript quality:

• I find Figure 4 hard to read. It should be improved.

• It is more appropriate to compare your circuit performance with conventional readout circuits of infrared focal plane array detectors.

• How much is the spikes duration? Would it affect the operation of the following SNN?

• Would the size and number of focal plane array pixels affect the overall performance of your readout circuit?

• How fast you can switch (configure) between temporal and frequency encoding?

 

 

 

 

 

 

 

 

 

Author Response

Sincere appreciation for your insightful comments on our manuscript. We have revised the manuscript based on your comments and provided detailed responses to each of them in the attachment. We trust these responses will help you locate these revisions in the manuscript more quickly.

Author Response File: Author Response.docx

Reviewer 3 Report

The paper numerically investigated SNNs with readout integrated circuit (ROIC) of infrared focal plane array (IRFPA). The concept is very interesting. However, the presentation is confusion. Please reconsider the structure of the manuscript and address the following comments.

 

1.         The main point is the relation between SNNs and IRFPA is unclear. The title uses “Spike Encoding Circuit Based on Infrared Focal Plane Array”. The conclusion section shows “Our solution eliminates the counter design and saves space for the focal plane array pixel area when compared to the technical solution of the digital pixel readout circuit used in conventional sensor arrays.” Is the purpose of the study to eliminate the counter design using SNNs or to construct SNNs by using IRFPA readout circuit?

2.         The reviewer cannot understand which ROIC of IRFPA was considered in the manuscript because the are some types of ROIC such as DI and CTIA.

3.         Please explain the relation between Figs. 1,2,3,5,6, and 7 and IRFPA.

4.         There are some typos. Please revise English.

There are some typos. Please revise English.

Author Response

Sincere appreciation for your insightful comments on our manuscript. We have revised the manuscript based on your comments and provided detailed responses to each of them in the attachment. We trust these responses will help you locate these revisions in the manuscript more quickly.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

The paper is now much clearer. However, it is still confusing.

1.         What is the relation between the circuit of Fig. 14 and Figs. 10 and 11 (scheme B)?

2.         Are Fig. 4 and Fig. 3 concept models? Has they already been demonstrated or not. If just concept, the author must prove the concept.

3.         Please clarify what was demonstrated or what was not demonstrated. The discussion and conclusion are still confusing.

4.         In conclusion, the authors claimed, “Compared with standard digital FPA readout circuits, our scheme produces image information encoding that is more suitable for SNN processing.” Can the standard digital FPA ROIC produce encoding for SNN? If it is true, please explain the standard digital FPA for SNN.

Some sentance is still consfusing.

Author Response

Sincere appreciation for your insightful comments on our manuscript. We have revised the manuscript based on your comments and provided detailed responses to each of them in the attachment. We trust these responses will help you locate these revisions in the manuscript more quickly.

Author Response File: Author Response.docx