Handwritten Digits Recognition Based on a Parallel Optoelectronic Time-Delay Reservoir Computing System
Round 1
Reviewer 1 Report
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The paper describes the results of numerical evaluation of a handwritten digits recognition implementation using optoelectronic time-delay reservoir computing. The modelling performed in the paper show that it is possible to achieve up to 97.8% recognition accuracy with theoretical speed of 10^6 recognitions per second, and with relative stability with regards to physical parameter values.
There are some things which should be improved in the manuscript:
1. The references given in the first paragraph are quite outdated, especially given that the reference [4] is given when quoting computing resource consuption of classical computing systems, the reference being from 1994. The beginning of the introduction needs to be edited to include more recent results.
2. In the introduction there is a weird notion that there is an intent to find a process which could be implemented on a non-traditional hardware. I think this is simply a badly worded phrase - surely the main objective is to create a high-performance computing systems, but the implementation on a non-traditional hardware is not a goal in itself.
3. In multiple places in the article the authors mention a method "zoning 2 features" which is unclear from the text. From the description I think it is better described as "2x-downscaled features", in order for it to be clearer. The first appearance is on page 2 line 62.
4. Couple of typos: page 3 line 100 "concludes" -> "contains", page 3 line 109 "boarder" -> "border", page 3 line 114 "cast of extract" -> "cast of extracting"
5. It would be good to add future research directions in the conclusions.
Overall the comments I gave seem to fall under "minor revision" category, thus I would recommend accepting this paper for publication after these concerns are rectified.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Reservoir computing has received more attention as an efficient way for information processing and an emerging area in neuromorphic computing development, especially quantum reservoir computing. Authors gave a substantial review on time-delay reservoir computing system (TDRC) and its state-of-art in recognizing handwritten digits. Using an optoelectronic time-delay reservoir computing system, the authors numerically investigated its performance in handwritten digit recognition and found its robustness to the reservoir parameters when using histograms of oriented gradients (HOG) features in recognition. However, we expect the following points should be addressed:
1. Beyond the current review, a broader review and state-of-art of reservoir computing are expected to benefit more readers.
2. The correspondence of Figure 1 and Eq. 1 & 2 should be addressed: how each component in the schematic diagram of the optoelectronic time-delay system in Figure 1 is modelled and how they are reflected in Eq. 1 and Eq. 2 in page 3. For example, what y(t) is and what it represents in Figure 1.
3. The output weight Wout is calculated by the ridge regression method and Y is calculated in the formula in Page 4. It is not clear how the winner-takes-all approach is implemented. The details are expected.
4. On the page 4, “…the number of virtual nodes N is set to 1500, 700, 200, and 144 700, respectively, which satisfies an input-to-neuron ratio about 1:2…”, why the ratio is required to satisfy, or it is just set by authors in their experiments.
5. In the Figure 5, the correspondences are between vector x1, …, xk and elements of matric x is not clear. Also, how the output digits are calculated, please address it together with the point 3.
6. The mean and standard deviation in Figures 6 and 7 are mentioned. Please specify how the tests are sampled, especially for the Figure 7, which needs to sample 3 parameters.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The comments and questions are addressed and publication is supported.
Author Response
We sincerely acknowledge you very much for your positive evaluation to this work.
Reviewer 3 Report
Thanks for the revision. We have following points for authors to further consider:
The clarity of Figure 1 has been greatly improved. However, Eq. 1 & 2 are directly presented by referring a reference. This is a photonics Journal. We feel it is good to brief how each optoelectronic component in Figure 1 is modelled and eventually formed in Eq. 1 & 2.
In page 5 lines 179-180,”…that it is due to the gradients of an image are more useful than its 179 pixels…” could be better expressed as: “”…that it is due to the gradients of an image, which are more useful than its 179 pixels...”
Author Response
Responses to Reviewer 3 Comments:
1. The clarity of Figure 1 has been greatly improved. However, Eq. 1 & 2 are directly presented by referring a reference. This is a photonics Journal. We feel it is good to brief how each optoelectronic component in Figure 1 is modelled and eventually formed in Eq. 1 & 2.
Response:
We acknowledge you very much for your professional advice. In the revised version, Figure 1 is re-edited as Fig. 1(a) and Fig. 1(b), which show the schematic diagram of system setup and operation process of RC. The first paragraph of Section “2. System model” (lines 83-108) are modified. Two equations (Eq. 1 and Eq. 2) are added to explain the models for two main components (modulator and the filter) of the system. Meanwhile, a paper is cited as Ref. [29] which presents more detailed derivation process.
2. In page 5 lines 179-180,”…that it is due to the gradients of an image are more useful than its 179 pixels…” could be better expressed as: “”…that it is due to the gradients of an image, which are more useful than its 179 pixels...”.
Response:
We acknowledge you very much for your kind reminders. The corresponding expressions (lines 187-188) have been modified.
Once again, thank you very much for your professional comments and suggestions.