Effective Conversion of a Convolutional Neural Network into a Spiking Neural Network for Image Recognition Tasks

Round 1

Reviewer 1 Report

SNN is very good topic. This paper proposed a stream learning method using spiking neural model. The topic somehow does not appropriate for this publication, and the technical novelty of the paper is somewhat novel. Its contribution is moderately significant and the coverage of the problem sufficiently comprehensive and balanced. But, the bibliography is inadequate, and the overall organization of the paper could be improved. However, I have some questions below: 

1. The abstract can be extended to explain the idea more evident how to be related to concept drift than the current vision. 

2. What is the main contribution of the paper? The overall organization of the paper could be improved so that the article can be easy to understand. 

3. I suggest adding some ablation study. 

4. I strongly recommend authors to release the source code along with the submission, since the learning based projects are typically open-source oriented to facilitate a fair assessment of the performance of the proposed methods for the community. 

5. Why your proposed method can obtain better performance than others? I would appreciate a broader discussion on why the proposed method performs better than the others. 

6. They need to compare state-of-the-art methods and need to use more metrics. 

7. Some state-of-the-art works on temporal coding of spiking neural networks are missing,
a. Computational mechanisms of pulse-coupled neural networks: A comprehensive review
b. Feature-linking model for image enhancement
c. New spiking cortical model for invariant texture retrieval and image processing

8. Last but not least, the author should pay attention to several grammar issues and inappropriate usages. The authors should carefully proofread this paper and correct all the typos in the revision. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Review

Propose a method for SNN to CNN conversion with less loss and higher SNN accuracy.

Minor Comments
pg 1. ln 23, add a sentence or two describing what is SNN and its history with reference. 
pg 1. ln 30-31, provide a reference for the 3 main approaches.
pg 2. ln 49 and 52, provide the quantitative amount of how much more accurate and less loss 
pg 3. ln 72, explain and provide a reference for IF neurons
pg 3. ln 103, explain "unstable"
pg 4. activation, regularization, and pooling are all standard concepts in Deep Learning, no need to go into details describing each, providing references is sufficient
pg 8. ln 206, add a reference for the two data sets.
pg 8.  ln 214, what is the CNN network you've used?  is it the one described in Fig 6?  but Figure 6 doesn't show the exact details of the network
pg 10. fig 7,8, number on graphs too small to read
pg 15. fig 14, use the same colour scheme to be consistent, also consider users who will print out the paper without colour printer (ie. use dash lines in addition to colour)
pg 16. ln 235,236,338 remove "very" 

Major Comments
pg 1.  add more background on SNN.
pg 16. provide more explanation on the success of the conversion method over others.  Also, discuss the cons of the conversion approach in comparison to the other methods.  Provide recommended future research direction and improvements.

Overall
The author developed a novel CNN to SNN conversion scheme that showed significant improvement over existing methods.

Author Response

Please see the attachment

Author Response File: Author Response.docx