PulseNetOne: Fast Unsupervised Pruning of Convolutional Neural Networks for Remote Sensing
Round 1
Reviewer 1 Report
The paper proposes an algorithm to perform pruning of CNNs. The proposal is based on a clustering algorithm that groups convolutional filters by similarity, removing those that are very similar. This way, the number of parameters is reduced. After the pruning process, a retraining process is sill needed. The authors implemented their algorithm on two state of the art architecture achieving high success in different datasets. The experiments are exhaustive and convincing.
Nonetheless, I have some concerns:
- The paper is unnecessarily long. For instance, as AlexNet and VGG are well known, the authors should remove their sections and provide a citation instead. In addition, there is no need for Figures 1,2,3,4,5 and 6. A simple citation to the source paper is more than enough. Likewise, the confusion matrices could be ommited as they are not discussed. Or at least, remove the rows that show no noticeable confusions with other categories.
- Section 4.2 should be moved to an early section in the manuscript. For instance, Section 3.1 references PulseNetOne but it is not defined until Section 4.2. As far as I concern, I find this misleading and makes the manuscript hard to understand.
- I am concerned by the outcome of the experiments. The tested architectures are too weak (AlexNet) or have a lot of parameters as for today's standards (VGG). This fact, combined with the low number of samples of the datasets, could lead to an overfitting problem. This effect could be ammended by reducing the parameters of the architectures, leading, thus, to an improvement on the accuracy. In general terms, it is expectable that the accuracy of a network is reduced as the number of parameters is lowered. Nonetheless, the acuracy improved after the pruning process as the experimentation remarked. I think this is because the parameters reduction helps to fight against a potential overfitting issue. I suggest to involve a more modern, accurate and optimized network in the experiments to test the proposed pruning method. For instance, ResNet is more accurate than VGG and yield less parameters (23M vs 134M).
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 2 Report
In general, I have a very good impression from this paper. It deals with modern approaches to image processing and object recognition. Sufficient results are obtained concerning both improvement of correct recognition probability and simplification of the system structure. The method is intensively tested and the obtained results are impressive.
Meanwhile, I have quite many comments concerning this paper:
1) Is this really network compression or some other (better) term can be used instead of compression like complexity reduction, simplification, etc.?
2) What is image orientation? I started to understand only after reading several pages. Maybe, orientation of objects in images? This has to be explained or corrected.
3) In line 19, it is said "parts can be considered noise". Noise in remote sensing image processing has another meaning. Maybe, some pther term can be found?
4) "reconstructiong" in line 39
5) what are "hand-crafted features"? Please explain.
6) line 179 "the its classification"
7) I dislike the style of sentences like "[40] introduced a novel way...". This is in many places. Maybe, the authors if [40]...
8) The datasets MIT 67 and SUN397 can be hardly treated as remote sensing datasets.
9) data in red blocks like in Fig. 10 can be hardly read. Is it possible to present them in a more convenient way?
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
All the concerns I raised in my previous report have been addressed properly. No I feel the paper is ready for publication.