MobileNets Can Be Lossily Compressed: Neural Network Compression for Embedded Accelerators

Round 1

Reviewer 1 Report

1. The Section 2 should be improved. The author should summarize the advantages and disadvantages of related methods and compare them.
2. The title of  Algorithm 1 should be placed above the content.
3. In the evaluation section, it is better to supplement the experimental comparison with other methods.

Author Response

Thank you very much for your valuable feedback!

We agree with your concerns, and have made the following changes to the manuscript.

 

1. We have improved Section 2 to emphasize the strengths and shortcomings of existing work, especially DeepSZ, in order to set the stage for our work.
2. We moved the title of Algorithm 1 above the content
3. We re-organized the evaluation section and added table 2 as well as section 5.6 to explicitly compare against existing work and emphasize our strengths.

 

We have also fixed minor grammar and spelling issues.

Thank you very much for helping us improve the manuscript!

Reviewer 2 Report

There is no description of the methods in the article

Author Response

Thank you very much for your help on improving the manuscript!

 

We understand our methods and novelty may not have been emphasized clearly enough.

We have edited the manuscript to emphasize our methods and contributions.

Specifically, we introduce a novel ZFP variant ZFPe, and show that its hardware implementation is efficient enough to remove the off-chip memory bottleneck.

We also added text in the evaluation section to describe how our design results in a more efficient hardware implementation compared to existing work.

Reviewer 3 Report

The article is well written and may be accepted in its current form. 

Author Response

Thank you very much for your kind words!

In our revised manuscript we tried to better present our methods and emphasize our novelty.

We hope our revised manuscript is even better than the original.

Reviewer 4 Report

1. The authors have evaluated their proposed design in terms of top 1 and top 5 accuracy metrics. However, it is unclear how much should these values be in practice?

2. Fig. 7 is confusing. Have the authors clipped the figure in between? It is better to use a different scale if the authors are unable to capture the entire figure or split the main figure into multiple sub-figures instead of clipping the figure.

3. Also, from the results, it is unclear how much is the gain of the proposed design over state of the art methodologies. The authors need to first clarify how much gain is considered a good amount of gain compared to state of the art and then show that they are able to meet that expectation. 

4. The discussion text for some of the results presented by the authors does not draw any significant insight. The authors need to clearly state their key findings and why those are important/new for the presented results.

Author Response

Thank you very much for your effort in improving this manuscript!

 

We have addressed your concerns in the following way:

1. In our literature search, Top-1 and Top-5 accuracy was the most common metrics to evaluate neural network models, and is used by the ImageNet challenge to compare accuracy. We have added text to emphasize it is a common metric.
2. We have split the graph into two, one for scientific data and one for neural network data, in order to remove the clipping. We were concerned a non-clipped graph may cover too much y-axis to be legible, but we were pleasantly surprised the split version is actually more readily understandable. Thank you for your suggestion.
3. We absolutely agree with your assessment regarding the evaluation section. We have added a significant amount of text and re-organized the section to better emphasize our strengths compared to existing work. We specifically added table 2 and section 5.6 to emphasize our novelty against the state-of-the art.

 

We also fixed many grammar and English issues.

Thank you again for your insightful suggestions on improving the manuscript!

Round 2

Reviewer 1 Report

The authors have addressed my concerns and I would like to see the paper published.

Reviewer 4 Report

I have reviewed the revised draft. The authors have made satisfactory changes to address my comments.