Improvement for Convolutional Neural Networks in Image Classification Using Long Skip Connection
Round 1
Reviewer 1 Report
Paper presents an improvement to the skip connection principle first presented in DenseNet. The main difference that authors add additional addition blocks and longer connections between higher-level blocks. Research is interesting, as authors show the potential of this technique to improve accuracy, reduce learning duration and computational cost of deep neural network model trained for an image classification task.
The main strength of this article is well-covered methods. On the other hand, there are more weaknesses:
• the figures are very primitive and not showing enough details about the size of layers or their type, therefore the article looks unfinished;
• there are also lots of singular vs plural errors, verb tense errors, and overall errors, skipped words,
• some abbreviations are not used consistently, or look different in various parts of the article.
• the actual numbers of figures are not referenced in the text, but instead, non-existent figures referenced.
• the literature review is quite weak, the newest articles are 2017, and most networks authors are referencing are more than 5 years old, authors could include some newer models such as Efficient net, MobileNetV2 and V3 instead of MobileNet.
• I believe that comparison of different size networks using the same count of epochs for learning and same learning parameters is not fair, because originally their authors used other learning parameters; also bigger networks require a longer time for learning. I agree that authors want to show that proposed networks learn faster, but when they should not compare networks by accuracy, because longer training may provide better results for bigger networks, that are more prone to vanishing gradient problem.
• while defining the proposed model architecture, the article is not comparing it to DenseNet properly nor in Tables nor Figures, therefore it is hard to see actual contribution to model architecture. The reviewer had to open an article about DenseNet and compare tables. Change itself is very small, yet it gives better results, therefore I believe that this article can be interesting if the weaknesses solved.
There is a line by line comments, that covers some of the mistakes and suggestions to authors:
Line 8: "about" can be omitted
Line 9: convolutional networks
Line 13: suggest changing "... save both parameters ..." to "... reduce both count of parameters ...."]
Line 15: dataset -> datasets
Line 16: (minified version of ) suggest removing brackets and changing to "...reduced version of ImageNet."
Line 17 suggest changing "weight" to "parameters" or at least correcting to "weights"
Line 19: convolution ->convolutional
Line 22: and 23: visions -> vision
Line 23: "convolution ... network" -> "convolutional ... networks"
Line 24: When referencing initial of the name is not mandatory.
Line 26: abbreviation should be used instead of convolutional neural networks
Line 28: "vanishing gradient" is more often used than "infinitesimal gradient", I suggest using first as it is more common
Line 30: it is unclear why ' ' are used for vanishing gradient because it is a common phrase in the field of artificial neural networks
Line 31: 'and' should be inserted between "connections" and "show".
Line 39: suggest changing "return" to "decrease"
Line 43, 74: inconsistent capitalization of the first letter in same phrase Depthwise separable convolution through the article.
Line 50: today's .... architectures
Line 52: ....equal to 2...,
Line 53: ending sentence and paragraph by comma
Line 61: models -> model that
Line 63: ...concatenation of all ...
Line 70: the strong fact is presented but no reference.
Line 72: firstly -> first
Line 77: adopts .... builds
Line 78: AlexNet is an old network, there is no use to compare to it, as you already comparing to DenseNet, that is much better.
Line 85: Resnet's
Line 88: later -> subsequent
Line 90, 92, 94: these sub captions should be moved to figure caption, only (a), (b), (c) should be left. Also, the figure formating does not comply with the author's guidelines of MDPI. The b) image is out of place it seems.
Line 96: densely -> Dense connections
Line 97: connectivity -> connections
Line 98: different with -> differently than/different from
Line 104: stayed -> stays
Line 106: from -> feature of
Line 131: suggest changing those -> ones
Line 147: "to downsampling" -> "to downsample"/"downsampling"
Line 159: Figure 8 non existent in article
Figure 3: a) If improvement to DenseNet is described here, why not showing actual DenseNet architecture instead of some generic model of unknown architecture without no details? It is very unclear what is difference between DenseNet and your proposed model architecture this way.
Line 168 and 169: inconsistent use of S-CONV, s-Conv. Please use capitalization consistently.
Table 1: There in rows of Block 1 and Block 2 tree identical fields are presented, therefore can be reduced to one as in other rows with identical block parameters
Line 174: build -> built
Line 198: Figure 10 is not existent.
Line 205: I suggest placing a reference or full address instead of hyperlink or in addition to a hyperlink. As researcher often print articles to read and printed version of this paper will have no link available.
Figure 6: Readability can be improved by adding major and minor gridlines.
Line 223: consuming -> consumption
Author Response
Please see the attachment
Author Response File: 
Author Response.doc
Reviewer 2 Report
-The paper should have 15-20 pages.;;;
-Discussion Section is missing;;;
-please add colorful picture of measurements (optionally);;;
-please add block diagram of the proposed research step by step ;;; what is the result of paper?;;;
-please add block diagram of the proposed method;;;
-please add formula for Accuracy?;;;;
-please add photo/photos of application of the proposed research ;;;;
-figures - please add labels, describe it better.;;;;
-please add sentences about future analysis;;;
-references should be 2018-2021 Web of Science about 50% or more (30 references at least);;
-Please compare with other methods, justify. Advantages or Disadvantages;;;
for example:
1) Fault diagnosis of electric impact drills using thermal imaging, Measurement, Volume 171, 2021,
https://doi.org/10.1016/j.measurement.2020.108815
-Conclusion: point out what are you done;;;;
Author Response
Please see the attachment
Author Response File: Author Response.doc
Round 2
Reviewer 1 Report
Thank you, for providing a revised paper and solving most of the paper’s weaknesses, and language is also improved. Article quality was improved substantially. After viewing all the paper I see some things that, I would suggest to improve:
- New materials as graphs and ablation study are added to discussion only. This material, in my opinion, should appear in section 4 "Experimental Results and Analysis". The discussion should not include any new results or experiments as an ablation study. I suggest moving all new results to results related section, analyse them there, and if needed discuss them in the discussion section.
- The conclusions section is weak. Please check if you do all points of good conclusions: 1. Providing the last word on the problems being solved in your paper; 2. Summarizing your discussed results and conveying the larger implications of your study (actually I see implications, but not summarizing); 3. Demonstrating the importance of your ideas (you could provide some numeric arguments, how long skip connections improve speed and resource consumption); 4. Introducing possible new or expanded ways of thinking about the research problem (the future plans are presented, so this point is made).
Table 1: Here are still the same values in all columns in line with 'Block 1' and 'Block 2'. That should be one as in other rows with the same parameters.
Line 103: Pixel4 edge GPU. Please check if it is really GPU because the reviewer checked and found out that it may be TPU (Tensor Processing Unit) usually "Google Edge TPU" is used often, and it is integrated into Pixel 4 smartphone, besides ordinary GPU that is in every smartphone currently. Please check and correct if needed.
Figure 11: Readability would be improved if the grid added to the graph.
Author Response
Please see the attachment
Author Response File: Author Response.doc
Reviewer 2 Report
116 identity as showed in Figure 1: - shown
-please correct english
Author Response
Please see the attachment
Author Response File: Author Response.doc
