Advancements in Complementary Metal-Oxide Semiconductor-Compatible Tunnel Barrier Engineered Charge-Trapping Synaptic Transistors for Bio-Inspired Neural Networks in Harsh Environments
Round 1
Reviewer 1 Report
You can find my comments and suggestions in the attached PDF file.
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment
Thank you for your kind review
Author Response File: 
Author Response.docx
Reviewer 2 Report
Summary:
This study systematically tested the electrical characteristics of an electronic synapse made from a CMOS-compatible, high-k tunnel barrier, and SOI-based charge-trapping transistor. The results demonstrated that the synaptic weight could be controlled via sequential gate pulses. The synapse also showed robustness, with stable operation over 10K program/erase cycle, and at high temperature. Finally, the functional capability of the synapse was demonstrated by simulating an artificial neural network for digit recognition.
The short/long term memory property showed is an unfavorable feature. It correlated with synaptic weight, complicating independent control of these two factors, resulting challenging limitation for real-world application. However, it’s a good starting point for further improvements.
Suggested revision:
1. In Section 3.3 for simulation of ANN for digit recognition, there are important implementation details missing, please clarify.
What is the basic assumption of how the synapse was used? For example, which terminal is connected to the presynaptic input, is it just the same gate electrode that used to control the weight or two separate gate electrodes one for receiving input signal and the other one for controlling the synaptic weight. What is the dynamic range of the input signal, if it’s strong enough to change the weight (cause charge trapping or de-trapping)?
And what is the learning rule for training the ANN? Is the learning unsupervised or supervised? In supervised manner, the back-propagated error signal was used to change the weight, while unsupervised learning could be implemented by the intrinsic synaptic property as shown in Fig 3.
2. Figure 4d showed that the memory length (short/long term) depend on the synaptic weight, which is an unfavorable feature, should be discussed.
Author Response
Please see the attachment
Thank you for your kind review
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
I appreciate the author's detailed explanation in the response, which clarified my confusion about ANN simulation.
And there are a few misunderstandings I want to clarify.
1. There are two suggestions in my previous review comment, not sure how the first suggestion was separated into 3, which causes the misinterpretation. So in the reviewer response, suggestions 2 and 3 are explanations for suggestion 1 and they are the points I wish the author could clarify as missing implementation details in the manuscript. Now I understand as the ANN was trained with supervised learning through error-backpropagation assuming no experience dependent synaptic plasticity.
2. The authors did not get the point of my second suggestion (suggestion 4 in the author response). So I want to further clarify. The memory length (short/long term) depend on the synaptic weight is not a desirable feature because we want the weights and memory length to be independently controllable.
Author Response
Please see the attachment
Author Response File: Author Response.docx
