Exploring Artificial Intelligence in Smart Education: Real-Time Classroom Behavior Analysis with Embedded Devices
Round 1
Reviewer 1 Report
1. The title seems interesting but abstract doe snot reflect the same. The abstract is generic and discussing embedded devices and technologies for teaching process. Plus some information is on the class behaviour. Consider reviewing the abstract to reflect the actual work
2. Introduction is starting with the class behaviour analysis, then requirement, and challenges, and a solution. But where is the information on the smart education ? artificial intelligence link ? deep learning models? embedded devices. The introduction does not reflect the title and more on teaching quality and student development
3. Solution should be after the analysis has been conducted and methods but here it is before.
4. Contribution is too long and unnecessary information is presented which has no relevant research sources
5. Literature is on AI technology and education. Looks some relevant then computer vision technology applications started and then author moved in to section 3 with principles. It is too short and vague
6. All principles here does not make any sense as it is random information presented with no relevant aspect with the paper. Further the neural network discussion is very generic. There are points as introduction to concepts such as attention mechanism but not linked back with the previous sections
7. Material is providing solution structure but this all is based on computer vision not the Deep Learning Model Optimization Design Applied to Embedded Devices. There is lack of information on the dataset as well as how authors reached this point of work
8. Methods has improvements and privacy protection too. But nothing about deep learning methods. There are some points in 5.1.2 on the dataset but this is emotion recognition dataset with computer vision elements and facial expression, that is a total different field of work.
9. Discussion is in similar manner.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Congratulations to the authors on producing this article for review. The following observations are provided for feedback on the article submitted for review and intended to assist the authors in identifying opportunities for improvement:
1 How well the main question is addressed by the research: at section 1.4 (and lines 292-294, 532-534) the purpose of the research is identified in the sense of what is being done. The research question, or background, are explained extensively prior to this section but do not specifically identify a single or small number of focused, concise research questions. It is suggested that the authors can clarify the focus for the research by articulating, at the start of section 1.4, the specific research questions to be answered. Sometimes drawing the attention of the reader with a sentence starting, “The research questions are …” provides a simple and clear method of making the purpose and focus of the research clear and ability to determine the level of achievement in answering these in the conclusions is made much easier.
2 This research appears both original and relevant to the field, addressing a gap in knowledge. The evidence and argument for the need for the research is presented well indicating limits of previous research and challenges yet to be addressed.
3 This research appears to offer a successful solution to challenges faced in previous research in relation to retaining security of data.
4 Without data on a control group it is not possible to infer whether the claims associated with efficiency and effectiveness of teaching compared to existing methods without AI can be suitably validated when assessing improvement in teaching quality.
5 The methodology is clear and justified.
6 Whilst there is considerable data arising from undertaking the research there is little discussion arising from the output and there is limited evaluation of the results.
7 Whilst there are conclusions, because the focus of the research is not sufficiently clear nor compared with a control group it is not possible to ascertain the robustness of the research claims.
8 Additional comments on content, tables and figures:
· Line 85, sentence “Teachers have a heavy teaching workload and cannot spend a lot of time.” Appears incomplete – what is it that the teachers cannot spend a lot of time on? Lines 86-88 indicate this may be classroom analysis. Consider writing lines 85-88 again to remove ambiguity of what is meant.
· Line 102, try to avoid the use of words and abstract nouns that cause ambiguity because they are unclear e.g. huge. Similar issues with ‘high’ at lines 197 and 201. Line 241 sentence “As seen in Figure 4, there has been a large increase in” substitute large with the absolute number from Figure 4 or give the percentage increase as the data is in Figure 4 and ‘large’ is an interpretation.
· Last sentence of paragraph at lines 124-125 is duplication, this sentence could be removed.
· Line 135 space required between Figure and 1.
· Lines 226-227, consider rewriting to remove the use of informal language: currently, “Computer vision is a current research hotspot in AI. Automated facial recognition (AFR) is a hot topic in AI applications and is also widely applied on campus and in education.” It is suggested that a more formal and appropriate sentence could, for example, be “Computer vision is currently an area of intensive research activity in AI with automated facial recognition (AFR) a priority in AI applications in the context of education.”
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
As a tremendous progress in Modern education, author(s) introduced advanced 13 devices and Technologies into the teaching process. Embedded devices plus artificial intelligence is an advanced solution to the problem of classroom behavior analysis, however, AI models that can run efficiently in this environment. We convert student images to high-frequency images that cannot be recognized by the naked eye through image conversion technology. The authors optimized lightweight PIDM model that works well locally on embedded devices. PIDM embedded MR devices work and display analysis data in real-time without interfering with teachers' teaching.
This paper seems that it proposes technology way to support the education climate through technology devices. However, this paper includes many formulas as related computational aspect of the process. Is there should be included many formulas such as presented in the paper. This situation shows mixed detail that technological information is handicap for readers to follow main subject of the paper.
The authors explained that applications of AI in education fall into two main categories as improving learning effectiveness and the learning experience and Teaching management and assessment.
Consequently, the authors proposed a solution model that was an application of the AI to embedded devices, which avoids the problem of data leakage and protects student privacy. In addition, , the authors adopted a privacy-preserving method based on Fourier high-channel filtering. In order to protect the biological features of the analyzed objects.
The authors tested the embedded devices to use MR devices for classroom behavior analysis. The training set contains 28,709 samples and the test set contains 3,589 samples. Each sample is a grayscale image of a roughly centered head with emotion annotation, which is suitable for face emotion recognition model training.
This experiment tested the PIDM model in a variety of embedded environments, the effect of the PIDM model in terms of model size, accuracy, response speed and privacy-preserving treatment on accuracy. Regarding the results, the PIDM could run well in both of the above embedded environments. The the low-frequency information of the images has no effect on classification recognition. The experimental results show that the PIDM model meets the design requirements, and runs in an embedded environment with reliable accuracy in the difference in model accuracy is less than 1% in both embedded environments. The images processed by Fourier transform did not affect the model's learning of image features while privacy was protected. Experiments demonstrate that the processed images are not recognizable to the naked eye and do not affect model training and recognition. The authors explained that they will try to use techniques such as Network quantization, and knowledge distillation for model optimization in future studies. This sentence shows that current study has been planned as an initial stage for the future study. Also, the authors stated to improve the generalization ability of the model.
As a result, this paper is interesting in education area. For this reason, the limitations and implications sections should be embedded in the paper as applicable model in future study. Therefore, I suggest the authors to enhance the limitations and implications sections in the paper.
I congratulate the authors and good luck.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
The manuscript entitled “Exploring Artificial Intelligence in Smart Education: Deep Learning Model Optimization Design Applied to Embedded Devices” has been thoroughly reviewed. It discusses the application of deep learning model optimization design in smart education. The paper highlights the progress of modern education and the introduction of advanced devices and technologies, specifically focusing on the application of deep learning model optimization design in smart education. Although the general idea of the research looks clear, some work are suggested to assess its feasibility and validity before moving forward. Below shows the reviewer’s major concerns.
1. The manuscript lacks empirical evidence to support its claims. While the authors discuss their methodology for developing their proposed solution, they do not provide any data or results to demonstrate its effectiveness.
2. This work does not address some of the ethical concerns associated with using AI in education, such as bias or discrimination. However, this is not a required topic and could be a boost to this study.
3. The manuscript is highly technical and may be difficult for readers without a background in computer science or AI to understand. So a suggestion is to modify this manuscript with some context and background information to the techniques.
4. Some sections of the paper could benefit from more detailed explanations or examples. For example, the Privacy Protection on page 9.
5. There are some grammatical errors throughout the manuscript that could be corrected with careful proofreading. Therefore, careful proofreading is suggested.
6. While the authors acknowledge some of the challenges associated with implementing AI in education, they do not provide a detailed discussion of potential solutions to these challenges.
7. This paper could benefit from more discussion of how their proposed solution compares to existing solutions in the field. Instead of just listing those solutions.
8. The authors do not provide a detailed discussion of potential limitations or areas for future research related to their proposed solution. However, I insist on this discussion because this is a critical component of using AI on human subjects.
In summary, this paper discusses potential future directions for research in this area, which could help guide future studies in this field. While the manuscript provides valuable insights into how AI can be used to enhance smart education through deep learning model optimization design applied to embedded devices, it could benefit from more empirical evidence and a broader discussion of ethical considerations. Additionally, the technical language used throughout the paper may make it less accessible to some readers.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
1. Author just added one line in the abstract. That is not an improvement
2. 3.1 is in material and method, the comment was on introduction. Follow the previous comment:
Introduction is starting with the class behaviour analysis, then requirement, and challenges, and a solution. But where is the information on the smart education ? artificial intelligence link ? deep learning models? embedded devices. The introduction does not reflect the title and more on teaching quality and student development
The introduction comment is still valid. Deep learning and CNN under material and method does not make sense.
3. Authors moved principles to appendix but it included experiments too. Now the manuscript has no experiment details
4. There is no deep learning used in this paper. No model optimization or anything as mentioned earlier. The paper is not based on deep learning approaches. It is a simple paper with questionnaire rating and providing statistics. Nothing else in the paper.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The authors have done well in addressing the extensive observations identified with the result that this is a much more robust and detailed research paper. The remaining issues can be resolved at proof reading stage (lines 195, 199 and 203 for example).
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
After reading this revision, I agree that the authors have addressed most of my comments with sufficient evidence. Therefore, I am ok with the following proceeding of this manuscript.
Author Response
Thanks for your comments and help.
Round 3
Reviewer 1 Report
1. Title of the study has been updated now with removal of deep learning and CNN, the title makes some sense
2. Abstract now include the light weight models. Explanation is needed on this and plus the embedded devices. Emphasize the novelty of your approach. Briefly mention the unique aspects of your lightweight model PIDM, and how it differs from existing methods or models used in smart education and emotion recognition.
3. There is difference between emotion recognition and classroom behavior. this needs to be further differentiated
4. As you have updated the title, the literature review section needs to be updated too with the details of the lightweight approaches and methods. You mentioned PDIM but no background or further details have been provided. Further, you have mentioned in 1.4 "We propose the solution of embedded devices plus AI to apply intelligent and portable embedded devices to analyze students' behavior in classroom teaching.". However, its not clear what is meant by embedded device.
5. 2.3 is repetitive as 1.4. No added value
6. Surprisingly, authors have added model optimization without any models. Model optimization was going to be used when there are deep learning or AI models. However, the section 3.1 does not make sense. Authors added DSC but without any link with the paper. MobileNet is provided too but MobileNet is a computer vision model open-sourced by Google and designed for training classifiers. It uses depthwise convolutions to significantly reduce the number of parameters compared to other networks, resulting in a lightweight deep neural network. Now the question here is that authors mentioned that they removed deep learning and the model itself is a deep learning neural network. Where are the parameters and classifiers. ?
Further, To use MobileNet for emotion detection, you would need to fine-tune the pre-trained MobileNet model on an emotion recognition dataset. Common datasets used for emotion recognition include the Extended Cohn-Kanade (CK+) dataset, the Facial Expression Recognition (FER) dataset, or the Japanese Female Facial Expression (JAFFE) dataset.
Once fine-tuned on the appropriate dataset, the MobileNet model can be employed to analyze facial expressions in real-time and classify them into different emotions, such as happiness, sadness, anger, fear, surprise, and disgust.
It's worth noting that there are multiple versions of MobileNet, such as MobileNetV1, MobileNetV2, and MobileNetV3. Each version offers different trade-offs in terms of model size, computational requirements, and performance. Depending on your specific application and device constraints, you can choose the most suitable MobileNet version for your emotion detection task.
You need to provide details on these. Additionally, you have mentioned attention module, CBAM but you have not provided details on how these are incorporated with the previous models.
7. You have provided the solution structure but none other details on fine tuning and model use are presented
8. 3.3.2 just provided the PIDM model but all these details below just are the literature information. No actual results
9. Now back to the experiments as this was supposed to be the emotion detection + behavior analysis. You need to provide the results and pictures. Which dataset is used? Table 3 provided 5 different models but other than MobileNet and PIDM, the others are not discussed earlier or supported with literature. Table 5 is the pruning method results. After all of the other items, this is not making sense here.
10. None of the figures have model name. Only training curve is for PIDM.
Final Note: Dear Authors, this paper has some actual content, but its is all mixed up. There is effort needed to align the content and shape the content in form of a research article rather than mixing all items together. I am providing it a major recommendation this time to have the next version properly made up. You need to spend few weeks rather than few days and submit back a weak paper.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
