Total-Ionization-Dose Radiation Effects and Hardening Techniques of a Mixed-Signal Spike Neural Network in 180 nm SOI-Pavlov Process
Round 1
Reviewer 1 Report
Strong aspects:
This research focuses on the influence of radiation on the behavior of neuromorphic hardware which is a quite new approach and of high importance for future space missions.
Weak aspects
- The results shows that the radiation changes the frequency of neurons. However the authors does not present how this affects the learning process and the performance of the SNN to process information.
- The model of the brain's behavior during Pavlov experiment is superficial because it takes into consideration only first phase (reflex formation)
- The introduction section is too short. The authors should focus on the existing neuron models implemented in hardware and on previous studies of the radiation influence on neuromorphic hardware. See: doi: 10.1109/TNS.2018.2886793
An electronic neuron model that uses capacitors presented in a journal of MDPI is: doi.org/10.3390/s21082730
In addition the authors should present rigorous approaches to explain the Pavlov reflex such as: doi 10.1093/cercor/bhl152
- The authors should present the relation between the proposed circuit for weight adjustment and the mechanisms of learning of biological synapses such as spike timing dependent plasticity (STDP).
- The signal diagrams in figs 6 and 8 should be presented without the information given by the oscilloscope. This information should be presented using labeled horizontal and vertical axis as in fig 14.
- It is not clear how the input stimuli mimicking the sign of food and Ring of the bell produces training of the SNN. More details should be given related to the learning process when the first phase of the Pavlov experiment is simulated.
- Note that according to the Pavlov experiment the reflex is lost when the ring of bell is present without food. It is not clear how this experimental phase is modeled using the SNN.
- The font size in fig 14 should be higher
Author Response
Dear reviewers,
Thank you very much for your kindly comments on our manuscript entitled “Total-Ionization-Dose Radiation Effects and Hardening Techniques of a Mixed-Signal Spike Neural Network in 180nm SOI-Pavlov Process” (ID: electronics-1718068). There is no doubt that these comments are valuable and very helpful for revising and improving our manuscript. In what follows, we would like to answer the questions you mentioned and give detailed account of the changes made to the original manuscript.
Comment1: The results shows that the radiation changes the frequency of neurons. However, the authors do not present how this affects the learning process and the performance of the SNN to process information.
Response: Thanks so much for your useful comments. These parts can be found in Page7 “The spike information transmitted in SNN has no special difference between each other in terms of amplitude and width, which can be regarded as a series of points over time. However, the spike rate is extremely important for information coding and communication of SNN.” This is due to the fact that neural coding is usually associated with an average number of spikes per unit of time. Abnormal frequencies that are too high or too low can cause errors in neural coding and possibly incorrect changes in synaptic weights
Comment2: The model of the brain's behavior during Pavlov experiment is superficial because it takes into consideration only first phase (reflex formation)
Response: Thanks so much for your useful comments. Complex brain behaviors require more resources on the chip than our experiments can fund. We also hope to have the opportunity to do in-depth work in the future.
Comment3: The introduction section is too short. The authors should focus on the existing neuron models implemented in hardware and on previous studies of the radiation influence on neuromorphic hardware. See: doi: 10.1109/TNS.2018.2886793
Response: Thank you for pointing out this problem in our manuscript. According to the revised content, we have revised this section. However, much of the existing related work (As you have mentioned) is based on RRAM, which differs significantly from SOI-based work.
Comment4: An electronic neuron model that uses capacitors presented in a journal of MDPI is: doi.org/10.3390/s21082730
Response: Thank you for your recommendation. We have read the article and we think it is a fantastic job. We have gained a lot of inspiration and motivation.
Comment5: In addition, the authors should present rigorous approaches to explain the Pavlov reflex such as: doi 10.1093/cercor/bhl152
Response: Thank you for your suggestion. It is really true that “STDP rule“ is indeed closer to the biological interpretation model, however it is difficult to implement at the chip level. Examples include TrueNorth, HICANN, Neurogrid, ROLLS, Minitaur , etc. From the point of view of large-scale integration, STDP-related circuits are not yet developed enough for space applications.
Comment5: The authors should present the relation between the proposed circuit for weight adjustment and the mechanisms of learning of biological synapses such as spike timing dependent plasticity (STDP).
Response: Thank you for your suggestion. These parts can be found in Page4 and Page5. The picture below shows the SDSP on the left and the STDP on the right.
The difference between the two algorithms lies only in the pattern of changes in synaptic weights. The two algorithms differ only in the pattern of change in synaptic weights. SDSP focuses on the population characteristics (frequency) of spikes, whereas STDP focuses on each spike.
Comment6: The signal diagrams in figs 6 and 8 should be presented without the information given by the oscilloscope. This information should be presented using labeled horizontal and vertical axis as in fig 14.
Response: Thank you for your suggestion, and the signal diagrams in Figures 6 and 8 are actual experimental diagrams, which are difficult to change due to equipment problems. However, we have added graphs of the simulation results to demonstrate the experimental content. (Page 7)
Comment7: It is not clear how the input stimuli mimicking the sign of food and Ring of the bell produces training of the SNN. More details should be given related to the learning process when the first phase of the Pavlov experiment is simulated.
Response: Thank you for your suggestion. Following your suggestion, we have increased the description of Fig.5-6. (Page 6-7)
Comment8: Note that according to the Pavlov experiment the reflex is lost when the ring of bell is present without food. It is not clear how this experimental phase is modeled using the SNN.
Response: Thank you for your suggestion. We have adjusted the description of Fig.6 in the revised version. (Page 6)
Comment9: The font size in fig 14 should be higher
Response: Thank you for your suggestion. We have increased the size of Figure 14.
(Page 12)
Author Response File: 
Author Response.docx
Reviewer 2 Report
This paper presents an detailed evaluation of TID radiation effects on a Mixed Signal Spike Neural Network.
The results are discussed in details, explaining the circuit relations and dependences.
The main contributions of the manuscript are to present the implementation of a mixed-signal spike neural network in 180 nm SOI, evaluating the TID radiation effects.
This part the authors describe with deep details, following a strong approach of discuss the circuits reasons for the effects.
It is a considerable contribution.
However, the hardening techniques are not presented with the same level of details. This part of the article deserves more attention and discussion.
My main questions and suggestions are:
1) I appreciate the hardware description and evaluation of TID effects. However, about the hardening techniques, I do not understand if the proposed approaches have been evaluated or if the authors are proposing approaches to be future tested and implemented. About these point:
1.1) How the TID effects can affect the behavior of the additional auxiliary charging module?
1.2) Are the layout techniques also implemented in 180 nm and tested in the same conditions?
1.3) The approach of increase the length of the devices will change the operation of these devices to long channel behavior. Why the authors have opted for this approach? How approach was adopted to define the new transistor length?
1.4) If the hardening techniques are future work, or under evaluation, I suggest to rename the manuscript removing the "and Hardening Techniques", and, changing the Subsection 3.3 for a section of discussion about Radiation-Hardening Techniques
2) The simulated results presented in the Subsubsection (3), pg 10, is not clear. What are the circuit under evaluation? How the parameters are varied in the Monte Carlo simulations? Are just the firing rate observed? How is it observed? Also, the results are merely presented, and I miss a deep discussion about it.
3) Figure 4 is very helpful to the understanding of how the algorithm work. I suggest to avoid the repetition of the names on the spike monitoring module devices, and the weight update devices. Renaming the M1- M4 devices on one of these modules will improve the clearness of the text.
4) How the evaluation stands from the related work? What are the other circuit implementations related in the literature? What are the hardening approaches proposed in the literature?
5) I suggest to reorder the Section 2.2, starting presenting the figure and explaining the circuit, to them, present the model.
6) There are some parameters on the equations that are not explained in the text.
6.1) How are defined the values to the a, b, c and d on Equations (1) and (2), to reflect the recovery rate, reset values, etc?
6.2) How are defined the threshold values on Equations (4) and (5)?
7) About the experimental details: What are the reasons ( or references) to justify the radiation dose rate , and total doses applied?
Some minor details:
8) Adopt on all the manuscript the same pattern to refer to the 180 nm process.
9) Pg. 2: Section3 -> Section 3
10) Introduces the subsubsections goals on the Subsection 2.2.
11) Pg 3: "The current Iin(t) charges the neuron membrane capacitor Cmem, when ..." -> Break in a new sentence: ". When...."
12) Pg 3: The third paragraph is confused. Please, rewrite to clear present the Leak circuit.
13) On the same page (pg 3) the author repeat the same sentence in two points. "The current Iin charges the neuron". Please avoid the direct repetition of the same phases.
14) Pg 3: "... is presented in Fig.3", "which ..."
15) Equations, Figures, Tables are proper names that must be capitalized.
16) Enumerate the blocks on the text on the spike-based learning algorithm description and on the Mixed-signal Pavlov SNN. For example:
It consist of three parts: 1) spike neuron ....
17) Remove unnecessary subsection titles on the Experimental details and results. The text can better flow without the subsections, mainly because these subsections are very short.
18) The definition of the variables on the equations must be reviewed. For example, after introduce the equations, you can define the parameters and variable just below the equations, utilizing:
"
where:
x are ...
y is ...
"
It helps the reader to see the equation and variables together.
19) The figure numbers must be revised. The major of the references are incorrect. It makes very difficult to follow the explanations.
Author Response
Dear reviewers,
Thank you very much for your kindly comments on our manuscript entitled “Total-Ionization-Dose Radiation Effects and Hardening Techniques of a Mixed-Signal Spike Neural Network in 180nm SOI-Pavlov Process” (ID: electronics-1718068). There is no doubt that these comments are valuable and very helpful for revising and improving our manuscript. In what follows, we would like to answer the questions you mentioned and give detailed account of the changes made to the original manuscript.
Comment 1.1: How the TID effects can affect the behavior of the additional auxiliary charging module?
Response: Thanks so much for your useful comments. As can be observed from our experiments, the sensitive nodes of the circuit are usually associated with capacitance, whereas the auxiliary charging module does not contain capacitance. Although TID effects cannot be avoided, compared to neuron circuits, the auxiliary charging module is more stable.
Comment 1.2: Are the layout techniques also implemented in 180 nm and tested in the same conditions?
Response: Thanks so much for your comments. Yes., we used the same 180nm-SOI process.
Comment 1.3: The approach of increase the length of the devices will change the operation of these devices to long channel behavior. Why the authors have opted for this approach? How approach was adopted to define the new transistor length?
Response: Thanks so much for your comments. This approach is used for the purpose of reducing leakage. When the length of the transistor is increased, the transductance of the transistor is also increased and the leakage current of the neuron will be suppressed. The exact length is determined by the process and must be such that the neuron output F-I curve at extreme process angles meets the design objectives.
Comment 1.4: If the hardening techniques are future work, or under evaluation, I suggest to rename the manuscript removing the "and Hardening Techniques", and, changing the Subsection 3.3 for a section of discussion about Radiation-Hardening Techniques.
Response: Thanks so much for your comments. We have revised this part. (Page 10) As you say, we have only completed the simulations and the hardened chips are still in production.
Comment 2: The simulated results presented in the Subsubsection (3), pg 10, is not clear. What are the circuit under evaluation? How the parameters are varied in the Monte Carlo simulations? Are just the firing rate observed? How is it observed? Also, the results are merely presented, and I miss a deep discussion about it.
Response: Thank you for pointing out this problem in our manuscript. We have adapted the circuit based on the previous method (fig) and carried out a Monte Carlo simulation with the Candece Virtuoso software.
We set up a leakage supply to equate the TID effect and observe the stability of the output frequency under parameter fluctuations. The simulations are used to verify the resistance of the neuron to fluctuations, similar to in subthreshold (doi: 10.1109/TCSI.2020.3035575)
Monte Carlo simulations of neuron circuits usually observe only the output frequency, and no special changes have been made in this paper.
(Page 11)
Comment 3: Figure 4 is very helpful to the understanding of how the algorithm work. I suggest to avoid the repetition of the names on the spike monitoring module devices, and the weight update devices. Renaming the M1- M4 devices on one of these modules will improve the clearness of the text.
Response: Thanks to your suggestion, we have revised this section. (Page 5)
Comment 4: How the evaluation stands from the related work? What are the other circuit implementations related in the literature? What are the hardening approaches proposed in the literature?
Response: Thanks to your suggestion, much of the existing related work (See: doi: 10.1109/TNS.2018.2886793) is based on RRAM, which differs significantly from SOI-based work, and Reference [7-10].
Comment 5: I suggest to reorder the Section 2.2, starting presenting the figure and explaining the circuit, to them, present the model.
Response: Thanks to your suggestion, we have revised the Section 2.2(3). (Page 4-6)
Comment6.1: How are defined the values to the a, b, c and d on Equations (1) and (2), to reflect the recovery rate, reset values, etc?
Response: Thanks to your suggestion, we have revised this section. The Izhikevich model is a classical model with the following parameters normally used:
a = 0.02
b = 0.2
c = -65.0
d = 8.0
In practical circuit design, c is usually set to gnd rather than a negative potential
(Page 2)
Comment 6.2: How are defined the threshold values on Equations (4) and (5)?
Response: Thanks to your suggestion, V1, V2 and V3 are 300mV, 900mV and 1.5V respectively, VW_THR, VMAX and VMIN are 900mV, 1.8V and 0V respectively, the setting of these parameters is mainly determined by the process and can be adjusted according to the test results during actual work. (Page 6)
Comment 7: About the experimental details: What are the reasons (or references) to justify the radiation dose rate, and total doses applied?
Response: This section references some of NASA's experimental setups, with an upper limit of 1000k (1M) and gradually increasing irradiation dose.
https://radhome.gsfc.nasa.gov/radhome/papers/tidpart.html
Comment 8: Adopt on all the manuscript the same pattern to refer to the 180 nm process.
Response: Thanks to your suggestion, we have revised this section.
Comment 9: Pg. 2: Section3 -> Section 3
Response: Thanks to your suggestion, we have revised this section. (Page 2)
Comment 10: Introduces the subsubsections goals on the Subsection 2.2.
Response: Thank for your comments, we have adjusted the description of Subsection 2.2 in the revised version. (Page 2)
Comment 11: Pg 3: "The current Iin(t) charges the neuron membrane capacitor Cmem, when ..." -> Break in a new sentence: ". When...."
Response: Thanks to your suggestion, we have revised this sentence. (Page 3)
Comment 12: Pg 3: The third paragraph is confused. Please, rewrite to clear present the Leak circuit.
Response: Thanks to your suggestion, we have removed the incorrect description. (Page 3)
Comment 13: On the same page (pg 3) the author repeats the same sentence in two points. "The current Iin charges the neuron". Please avoid the direct repetition of the same phases.
Response: Thanks to your suggestion, we have revised this section. It was rewritten as” The neuron membrane capacitance Cmem is charged by the current IIN” (Page 3)
Comment 14: Pg 3: "... is presented in Fig.3", "which ..."
Response: Thanks to your suggestion, we have revised this sentence. (Page 3)
Comment 15: Equations, Figures, Tables are proper names that must be capitalized.
Response: Thanks to your suggestion, we have revised these sections. Especially Equation (4), Equation (5)
(Page 5)
Comment 16: Enumerate the blocks on the text on the spike-based learning algorithm description and on the Mixed-signal Pavlov SNN. For example:
Response: Thanks to your suggestion, we have revised these sections. (Page 5)
Comment 17: Remove unnecessary subsection titles on the Experimental details and results. The text can better flow without the subsections, mainly because these subsections are very short.
Response: Thanks to your suggestion, we have removed unnecessary titles.
(Page 7-8)
Comment 18: The definition of the variables on the equations must be reviewed. For example, after introduce the equations, you can define the parameters and variable just below the equations, utilizing:
Response: Thanks to your suggestion, we have defined the parameters and variable just below the equation.
Comment 19: The figure numbers must be revised. The major of the references are incorrect. It makes very difficult to follow the explanations.
Response: Thanks to your suggestion, we have revised these references. (Page 12-13)
Author Response File: Author Response.docx
