Beauty in the Eyes of Machine: A Novel Intelligent Signal Processing-Based Approach to Explain the Brain Cognition and Perception of Beauty Using Uncertainty-Based Machine Voting
Round 1
Reviewer 1 Report
The article is devoted to the issue of cognitive analysis of the results of the work of computer vision methods. The topic of the article is relevant. The structure of the article is classical (Introduction (including analysis of analogues), Models and methods, Experiments, Discussion, Conclusions). Due to the ill-conceived structure, the article is easy to read. The level of English is acceptable. Many of the figures in the article are small. Proportions in Fig. 1 and Fig. 4 are violated. Tables 4, 5 are massive and uninformative. They need to be taken out in applications, and generalizations should be used instead. The References section cites 59 sources, not all of which are up-to-date.
The following remarks can be made on the material of the article:
1. The authors hooked on an interesting topic, but as always, they reduced everything to artificial neural networks, instead of dreaming up. For example, it has long been established that a beautiful face of a person seems to other people the more beautiful, the more symmetrical are its left and right halves. In addition, the "golden ratio" was invented by the Greeks, and it still works. You can measure it for different elements of the face (“Oh, what a beautiful left eye this guy has”), and for the entire face as a whole. Here you can really use AI to automatically cast the “nose” from the “ear” in photographs. But this is optional. Sarah Connor's first rule is if you can't use AI, then don't.
2. The authors use voting and selection of individual parts of the face. It's fine. I have a question: is the Venus de Milo beautiful? Yes, you answer, of course. What if she doesn't have hands? How did the authors foresee the situation that a beautiful-faced person does not have a left ear? Averaging is not the best way to measure beauty. This will be confirmed by any member of the jury of the Miss ...
3. The authors use convolutional neural networks. I am an old hacker and I will immediately ask: How stable are convolutional neural networks with respect to input data? Are they easy to break? It turns out that it is possible to introduce perturbations that are almost imperceptible to the human eye into the input data, which, nevertheless, completely change the output of the neural network. So, such almost imperceptible perturbations that change the output of a neural network are called adversarial examples (or, often, attacks on this same neural network. You heard right. It’s “almost imperceptible perturbations.” How did the authors check their offspring for stability? I think, dear ladies will get an app to prepare their "perfect photo for perfect AI" very soon.
4. And of course, the authors need to prove that they have chosen the optimal metric, double black and balanced data, take into account overtraining and gradient decay. And these are all analytics. Beauty requires sacrifice.
Author Response
Thank you for taking the time to review our manuscript. Kindly find the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The presented work is described reliably and in detail. The topic raised will undoubtedly arouse interest. However, I have attention. I propose to present the two described (lines 248 and 533) in a graphical form. I also wonder if the results of making decisions about the beauty class were compared with the beauty evaluation by the volunteers, that is, were eye movements tracked during the evaluation? In this way, it would be possible to correlate the ways of making decisions by the machine with points, for example, on which the eyes stopped longer.
Author Response
Thank you for taking the time to review our manuscript. Kindly find the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
I formulated the following basic recommendations for the basic version of the article:
1: The authors hooked on an interesting topic, but as always, they reduced everything to artificial neural networks, instead of dreaming up. For example, it has long been established that a beautiful face of a person seems to other people the more beautiful, the more symmetrical are its left and right halves. In addition, the "golden ratio" was invented by the Greeks, and it still works. You can measure it for different elements of the face (“Oh, what a beautiful left eye this guy has”), and for the entire face as a whole. Here you can really use AI to automatically cast the “nose” from the “ear” in photographs. But this is optional. Sarah Connor's first rule is if you can't use AI, then don't.
2: The authors use voting and selection of individual parts of the face. It's fine. I have a question: is the Venus de Milo beautiful? Yes, you answer, of course. What if she doesn't have hands? How did the authors foresee the situation that a beautiful-faced person does not have a left ear? Averaging is not the best way to measure beauty. This will be confirmed by any member of the jury of the Miss ...
3: The authors use convolutional neural networks. I am an old hacker and I will immediately ask: How stable are convolutional neural networks with respect to input data? Are they easy to break? It turns out that it is possible to introduce perturbations that are almost imperceptible to the human eye into the input data, which, nevertheless, completely change the output of the neural network. So, such almost imperceptible perturbations that change the output of a neural network are called adversarial examples (or, often, attacks on this same neural network. You heard right. It’s “almost imperceptible perturbations.” How did the authors check their offspring for stability? I think, dear ladies will get an app to prepare their "perfect photo for perfect AI" very soon.
4: And of course, the authors need to prove that they have chosen the optimal metric, double black and balanced data, take into account overtraining and gradient decay. And these are all analytics. Beauty requires sacrifice.
The authors consistently gave answers to all of them. Despite the fact that these answers are not supported by sufficiently weighty mathematical analytics, I agree with the authors' arguments. I hope that in the next works they will use the tools of the theory of mathematical statistics to a greater extent. I recommend this version of the article for publication and wish the authors further creative success.
