Trade-Off Between Energy Consumption and Three Configuration Parameters in Artificial Intelligence (AI) Training: Lessons for Environmental Policy
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors studied Artificial Intelligence paper titled “Trade-off Between Energy Consumption and Three Configuration Parameters in The Artificial Intelligence (AI) Training: Lessons for Environmental Policy”. The paper is recommended for publication. The following issues may be helpful for the authors to address:
- The paper's structure needs to rearrange the introduction and literature review should be on section with a logical flow of literature.
- The literature review should be comprehensive and current, clearly demonstrating how this study addresses gaps in existing research. Additional references are needed to strengthen the study’s foundation.
- The research questions and objectives should be defined at the end of the introduction.
- The novelty should be well highlighted. How significant is its contribution to environmental policy and SDGs?
- The authors should add details on why MNIST is representative for broader AI applications and its limitations for complex models like facial recognition.
- The simulation scenarios (Table 1) lack specifics on how hyperparameters (e.g., learning rate, optimizer) were controlled.
- The energy savings (e.g., 200% with early stopping) are presented, but the statistical significance of results is unclear.
- There are some grammatical errors, and don’t highlight the words in bold in the text.
- The results and discussion should be in one section.
The paper is well written, but it requires proofreading and correction of some typos.
Author Response
Comments 1: The paper's structure needs to rearrange the introduction and literature review should be on section with a logical flow of literature. |
Response 1: Thank you for pointing this out. We agree with this comment. Therefore, we have rearranged and modified the introduction in section 1, pages 1-2, paragraphs 1-7, and lines 34-91.
|
Comments 2: The literature review should be comprehensive and current, clearly demonstrating how this study addresses gaps in existing research. Additional references are needed to strengthen the study’s foundation |
Response 2: Agree. We have, accordingly, added the references to emphasize this point. Please see section 2.1, page 5, paragraph 1, line 166-171; section 2.3, page 7, paragraph 1, line 244-255.
Comments 3: The research questions and objectives should be defined at the end of the introduction. Response 3: Thank you for pointing this out. We agree with this comment. We have explained the research question and the objectives in section 1, page 3, lines 92-98.
Comments 4: The novelty should be well highlighted. How significant is its contribution to environmental policy and the SDGs? Response 4: Thank you for pointing this out. We agree with this comment. We have explained the novelty and contribution in the introduction page 2-3 lines 81 – 91
Comments 5: The authors should add details on why MNIST is representative for broader AI applications and its limitations for complex models like facial recognition Response 5: Agree. We have, accordingly, added the explanation in Section 5, page 18, paragraph 1, lines 554 -557.
Comments 6: The simulation scenarios (Table 1) lack specifics on how hyperparameters (e.g., learning rate, optimizer) were controlled Response 6: Thank you for pointing this out. We agree with this comment. Therefore, we have added the model specification and scenarios explanation in section 3, pages 8-9, lines 321-330, and page 9, lines 335-338.
Comments 7: The energy savings (e.g., 200% with early stopping) are presented, but the statistical significance of the results is unclear. Response 7: Agree. We have checked and revised in Section 4.1, page 10 lines 363– 370 & Section 6. Conclusion page 21 line 648-650.
Comments 8: There are some grammatical errors, and don’t highlight the words in bold in the text. Response 8: Thank you for pointing this out. We have checked and revised it.
Comments 9: The results and discussion should be in one section Response 9: Agree. We have merged Results and Discussion into Section 4 (please see page 10, line 354), and made Section 5 focus on the Application of Lessons Learned in Use case scenarios (please see page 18, line 548).
|
4. Response to Comments on the Quality of English Language
|
Point 1: The paper is well written, but it requires proofreading and correction of some typos. |
Response 1: Thank you for pointing this out. We have checked the writing and corrected some typos |
5. Additional clarifications |
- |
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for Authors1. In the introductory section, it is recommended to increase the relevance and logic of the upper and lower sentences, for example, by introducing the relationship between AI and sustainable development before specifying relevant cases. 2. In the Literature Review section, there are fewer references to key studies in recent years, and it is suggested to add relevant literature to enhance the theoretical depth. 3. In the energy consumption section of AI, a simple explanation of Figure 1 is suggested, such as the reason for the jump in numbers in 2021. 4. In the Research Methods section, the experimental design only uses the MNIST dataset (simple classification task), which cannot represent the energy consumption characteristics of complex tasks, and it is suggested to extend to multiple types of datasets. 5. In the Trade off between Energy Consumption and Early Stopping Epochs section, there is a logical ambiguity in the original multiplier formulation, as here in Figure 4(a), which is proposed to be amended to read that the energy consumption is reduced to 45 per cent of the original value (or reduced by 55 per cent), which corresponds to the original energy consumption being 2.22 times that of the Early Stopping Energy Consumption. 6. In the example section, data needs to be added to illustrate the advantages of applying an AI time and attendance system based on facial recognition technology.
Author Response
Comments 1: In the introductory section, it is recommended to increase the relevance and logic of the upper and lower sentences, for example, by introducing the relationship between AI and sustainable development before specifying relevant cases |
Response 1: Thank you for pointing this out. We agree with this comment. Therefore, we have rearranged and modified the introduction in section 1, pages 1-2, paragraphs 1-9, and lines 34-116.
|
Comments 2: In the Literature Review section, there are fewer references to key studies in recent years, and it is suggested to add relevant literature to enhance the theoretical depth |
Response 2: Agree. We have, accordingly, added the references to emphasize this point. Please see section 2.1, page 5, paragraph 1, lines 166-171; section 2.3, page 7, paragraph 1, lines 244-255.
Comments 3: In the energy consumption section of AI, a simple explanation of Figure 1 is suggested, such as the reason for the jump in numbers in 2021. Response 3: Thank you for pointing this out. We agree with this comment. We have added the simple explanation of Figure 1 in section 2.1, pages 3-4, paragraph 1, lines 124-129.
Comments 4: In the Research Methods section, the experimental design only uses the MNIST dataset (simple classification task), which cannot represent the energy consumption characteristics of complex tasks, and it is suggested to extend to multiple types of datasets Response 4: We appreciate the reviewer’s insightful suggestion regarding the inclusion of more complex datasets to better represent diverse energy consumption characteristics. We also fully acknowledge the importance of validating the findings across more complex and diverse datasets. However, we were unable to extend the experimental design to additional datasets beyond MNIST in the current manuscript. The main concern is that we intentionally make this a recommendation for our future work. The choice of MNIST was intentional to allow controlled experimentation on the effects of training configurations (e.g., batch size, early stopping, data size) while minimizing external variability. As the primary objective of this study is to explore parameter-level trade-offs in energy consumption, we believe the use of a simple and well-understood dataset provides a clear and focused foundation for these analyses. It also mentions the limitations and future work, as in the Conclusion lines 680 – 684.
Comments 5: In the Trade-off between Energy Consumption and Early Stopping Epochs section, there is a logical ambiguity in the original multiplier formulation, as here in Figure 4(a), which is proposed to be amended to read that the energy consumption is reduced to 45 per cent of the original value (or reduced by 55 per cent), which corresponds to the original energy consumption being 2.22 times that of the Early Stopping Energy Consumption. Response 5: Thank you for pointing this out. We agree with this comment. We have checked and revised in section 4.1, page 10, lines 363-370.
Comments 6: In the example section, data needs to be added to illustrate the advantages of applying an AI time and attendance system based on facial recognition technology. Response 6: Thank you for pointing this out. We agree with this comment. Therefore, we have added the data in section 5, page 19, lines 589 -591; page 20, lines 603-606; page 20, lines 615-617; and page 21, lines 648-650.
|
4. Response to Comments on the Quality of English Language
|
Point 1: The English is fine and does not require any improvement. |
Response 1: Thank you for your comment. |
5. Additional clarifications |
- |
Author Response File: Author Response.pdf