Hardware Demonstration of SRDP Neuromorphic Computing with Online Unsupervised Learning Based on Memristor Synapses
Round 1
Reviewer 1 Report
Authors present an interesting implementation of SRDP learning based on memristor synapses. I understand the work as a continuation of a previous work (ref. [4]), where the behaviour of the memristor as a synapse was shown. In this work, authors have advanced in the hardware implementation of the neuromorphic systems, although some calculations still remain on an external computer. The work is very interesting, the results are relevant and the implementation original and timely. So, in my opinion, the manuscript deserves publication in Micromachines.
My main concerns are about some missing explanations about some relevant (in my opinion) issues. The manuscript is based too much on ref. [4] and authors cite it for further explanations, but this makes the reading a more difficult task.
- For example, in lines 124-125 a leaky resistance and a capacitor and so on are referenced, but without explanation. Although they appear in Ref. 4, an explanation and figure in the current manuscript would make it more readable.
- The same comments can be applied in lines 213-214. Which specific previous theoretical and simulations results are cited? Please, provided here the main guidelines.
- Although the parameters Pg, Pr, Pn and so on are defined in Table 1, the reading would be easier if they are specifically defined or explained in the text.
- A little mistake, “V_{g}” in line 144 (g as a subscript).
I think that the contents of sections 1-4 are relevant enough and the manuscript could be published even without section 5. I understand the efforts of the authors for providing and additional study about non-binary data image and multilevel memristors. However, as the methodology is completely different from the previous section (it is a simulation study, without any hardware implementation, except for Figs. 8a and 8b, where authors shows that binary memristors does not work properly for gray-scale patterns following the proposed architecture), more information would be necessary. How were the simulations carried out? How were defined the memristor conductance levels? Equally separated? Was variability also taken into account? Do the different conductance levels overlap? How would they be programmed? So, in my opinion, the sentence “Classification of gray-scale and color-scale patterns is simulated, and 98.4% and 98.0% accuracy is achieved based on the analog property of memristor synapse.” is too categorical taken into account that there are too many open questions and that the accuracy of such neuromorphic systems could strongly depends on these issues.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
This paper shows SRDP neuromorphic system with online unsupervised learning method. Printed circuit board and memristor chips are used to verify the models. This results are very meaningful and technically there are no issues from my knowledge. so I recommend the acceptance at a current form.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
