This reviewer's personal approach, nomenclature examples: p1c1h0.8 = means page 1, column 1 80% of the way down the page (very approximately) C2. = means Comment section number 2 WEAKNESSES (note that actions by the authors are NOT required for the points) ++---------------------------++ ACTIONS REQUESTED OF THE AUTHORS In my opinion, this paper requires a substantial re-write before it will be acceptable to a conference. NOTE: I have provided several comments below in case these may be of use to the authors. *************************** *************************** COMMENTS ONLY - actions by the authors are NOT required for the points listed below, to the end of the review. Perhaps some of these comments will be helpful in some way. (Main paper contributions, positive aspects, observed deficiencies, and suggestions on how to improve them:) PERTINENCE of the paper for IJCNN2017 Anchoraage, Alaska : As per the "Call for Papers" listed at http://www.ijcnn.org/call-for-papers : "... IJCNN 2017: The annual International Joint Conference on Neural Networks (IJCNN) is the flagship conference of the IEEE Computational Intelligence Society and the International Neural Network Society. It covers a wide range of topics in the field of neural networks, from biological neural network modeling to artificial neural computation. ..." >> This paper addresses a NN [challenge, techniques] that are of relevance to IJCNN2017 ++---------------------------++ C1. STRENGTHS OF THE PAPER: Derivations from Equations (1) through (18) p2c1h0.9 Q-learning p3c1h0.9 Euclidean -> Riemann space -> G Fisher information matrix p3c2h0.1 Equation (11) Kullback-Leibler divergence (KL - text should spell this out!!) I really liked your derivations here - it's quite nice, although I cannot say what parts are the authors' original innovations (which should be pointed out!!). Figures [2,3 7,8,9] - explain well what is happening during the search, as well as roblems that arise. ++---------------------------++ C2. WEAKNESSES: (again, changes to the paper are not require for these comments) There are far too many [spelling,grammar, sentence structure] mistakes to warrant paper acceptance. I usually make an attempt to provide suggestions to correct the writing, but the extent of change required is simply too great in this case. +-----+ p2c2h0.7 rt definition "... While training, in all of the time t we know the target observation o and current observation ot , after at executed observation would change to ot+1 . we would elevate the distance between ot and o, ot+1 and o, then a reward rt would given based on distances to tell the agent its a good action or bad. so in every cycle we can get a set (ot , at , ot+1 , rt ). ..." p4c2h0.05 "... we evaluate action ats reward based the distance between new observation windows center and target windows center as follow criterion ..." Perhaps I am mis-interpreting the situation, but tis is a huge problem - it seems that the DISTANCE to the target is given, but not the direction. In other words, the system is given half the answer at each step (supervised learning with a very simple, half defined goal), and it makes NO use of the image information at the current step. So what are the agent policies' learning, given that each agent is specific to one photo, and nothing is learned about the images, not even the target image? It seems to me that a more general goal should be unsupervised learning for each photo - i.e. find the face!!! However, the authors' work may be useful when a match is sought for a partial face image among a large database of photos, each of which has a trained agent. But that really only makes sense if the trained agent already knows exactly where the face is - hopefully by a fully unsupervised algorithm. There isn't much use for the authors' algorithm in that case? I imagine that the policies would best succeed simply by deriving a gradient for the distance to target. A hand-made algorithm could do a really good guess with step-offs in only three directions (say 60 degrees apart) on the starting point? An evolved algorithm should be able to easily derive that over a few generations for the full database (again, in the same semi-supervised way). +-----+ Lost in the statistics? - As I stated under "C1 Strengths" above, I like the developments in the paper. however, I get the feeling that the statistical representations and computations may be too far diverced from mportant realities of the dataset, and that there is a weak undeerstanding about whant the algorithm is actually doing and achieving. Furthermore, I really wonder if there is much to be "ported" from the range of agents produced that would be at all useful to a different set of targets (other than the general approach). +-----+ Eye saccadic movements A large amount of work has been done on saccadic eye movements. Even though most of this does not relate to the current paper, given the goal of "Active face detection" it seems to me that there are important lessons and comparisons to be taken from that field, and that it should be at least mentioned. I suspect that results from that area would both provide better reductions in computational costs, as well as providing context for the portions of an image being searched for, which the current paper does not do. ++---------------------------++ C3. QUESTIONS: (no need to answer) I would have questions about the nature and performance of the agents on new images, the use of agents to train other agents, and the questions as to whether much [information, policy] re-use is occuring. ++---------------------------++ C4. DETAILS and GRAMMAR: (again, changes to the paper are not required for these suggestions) As per my comment in "C2. Weaknesses" : There are far too many [spelling,grammar, sentence structure] mistakes to warrant paper acceptance. I usually make an attempt to provide suggestions to correct the writing, but the extent of change required is simply too great in this case. p3c2h0.05 Equation (5) - seems to missing a "d_pi(theta_old)" on the Right-Hand_side. p4c2h0.75 Figures 4 _and_ 5 are not referenced in the text, albeit partial explanations are provided. ++---------------------------++ C5. REFERENCES (using a quick web search, as opposed to checks using Scopus or standard indexes. I do not have access to CrossRefs "CrossCheck" via Elsevier's "iThenticate") C6. MATH CHECKS - step-by-step Given post-review-deadline time constraints, I did not attemp my usual check on references, nor did I do a step-by-step re-derivation or check on the authors' algorithms. ************************** C7. LIMITATIONS OF THIS REVIEW Reviewer's expertise on the subject: Low I apologize to the authors for not providing suggested changes to the wording of this paper, but there are simply too many to be able to afford the time for that. Also, please note that this review is a "post-review deadline" effort, given that one or more reviewers were not able to submit their review by the deadline. The timeframe now does not allow as much effort on my part as I put into my other peer reviews for IJCNN2017. www.BillHowell.ca *************************** C8. THOUGHTS: (again, changes to the paper are not require for these) *************************** CONFIDENTIAL COMMENTS for review chair / committee use only: None other than the list of Ratings