Profile-Splitting Linearized Bregman Iterations for Trend Break Detection Applications

Round 1

Reviewer 1 Report

The authors treat the problem of high computational time of the Linearized Bregman Iterations (LBI) algorithm for trend break detection applications. This is due to a quadratic increase of the processing time with respect to signal size (N). 

To tackle this problem, the authors proposed a profile-splitting methodology enables blocks of data to be processed simultaneously. Indeed, instead of analyzing the profile as a single N-dimensional vector, the algorithm evaluates multiple M-dimensional vectors that, together, compose the original data.

Based on the proposed method, a Linearized Bregman Iterations algorithm hardware implementation for large datasets is studied.

 

The proposed methodology has been tested on fiber fault detection application using Optical Time Domain Reflectometry (OTDR) dataset.

 

The proposed technique sounds and shows decent results. I have several minor concerns:

 

1- The paper is not well organized, for instance, there is a very short section number 5 compared to other sections. Furthermore, this section is skipped from the paper organisation at the end of the section 1 (Introduction)

2- I believe that the fused lasso techniques should be included in the introduction section as an efficient method, all the following papers should be included: 

Bleakley, K., & Vert, J. P. (2011). The group fused lasso for multiple change-point detection. arXiv preprint arXiv:1106.4199.

Vert, J. P., & Bleakley, K. (2010). Fast detection of multiple change-points shared by many signals using group LARS. In Advances in neural information processing systems (pp. 2343-2351).

Rida, I., Jiang, X., & Marcialis, G. L. (2015). Human body part selection by group lasso of motion for model-free gait recognition. IEEE Signal Processing Letters, 23(1), 154-158.

3- It would be nice to visualize the different regularization, you can include them and cite the following papers:

Mairal, J. (2010). Sparse coding for machine learning, image processing and computer vision (Doctoral dissertation, Cachan, Ecole normale supérieure).

Rida, I., Al-Maadeed, N., Al-Maadeed, S., & Bakshi, S. (2018). A comprehensive overview of feature representation for biometric recognition. Multimedia Tools and Applications, 1-24.

4- Can the proposed algorithm be tested on DNA sets ?

5-  The langage should be further checked and improved, there are some typos

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The article is well written and I have no objection.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

I thought after reading the abstract that this paper would fit in my field of study, but unfortunately that was not the case. I must be honest and leave the proper revision of the scientific content and originality to my colleague reviewers.

That being said, I did my best to provide useful feedback. First of all, I think the abstract is a bit misleading and should state clearly what the reader will find: a thorough analysis on the efficiency of the algorithms when applied to this particular kind of problems, as well as their implementation on FPGAs. Something along what you can read on the conclusions.

I also was a bit confused about the whole analysis using the pure LBI algorithm. I understand that it may contain some interesting data related to the problem at hand, but, as it is an O(N^2) algorithm, it scales badly when N goes high. That is something which I think is quite well-known, and that is why the Split Bregman Technique was proposed. I seem to recall it is well-known too and somewhat often used. The authors may consider simplifying and clarifying this part, but they’d better see if other reviewers with more experience in this field think the same.

So, if I understood correctly (which is something that may not be), the authors adapt this split technique to the problem of trend break detection. This and the discussion on the impact of the parameter lambda and the length of the split is the important point of this first part of the paper, in my opinion.

The resulting algorithm is parallelizable, so the next sections about using dedicated hardware using FPGAs, are very relevant and interesting.

There are some minor things regarding the presentation. Figures are too large (texts in axes are enormous) and the use of only colour to distinguish between sets of data makes it difficult once the paper is printed on grayscale or for colour-blind readers. I usually like to add different symbols to aid in this.

There are some minor English corrections to do, but the paper is quite well written. Some typos also exist such as the missing reference in line 182, but nothing of importance.

After a minor revision, and if the rest of reviewers agree, the paper may be published.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf