Edge AI in Practice: A Survey and Deployment Framework for Neural Networks on Embedded Systems

Round 1

Reviewer 1 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

no

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report (Previous Reviewer 2)

Comments and Suggestions for Authors

Journal: Electronics (ISSN 2079-9292)

Manuscript ID: electronics-4032978

Title: Edge AI in Practice: A Survey and Deployment Framework for Neural Networks on Embedded Systems

Authors: Ruth Cordova-Cardenas * , Álvaro Gutiérrez * , Daniel Amor *

 

This review article focuses on recent developments in the deployment of deep-learning models to embedded devices with limited resources, exploring optimization techniques and explaining the different hardware platforms used, as well as lightweight architectures. A five-step process is suggested to guide the design, optimization, implementation and testing of effective AI systems on edge and embedded hardware. After reviewing the manuscript, my specific comments are below:

 

1. Abstract:

• The review should clearly demonstrate the methodological rigor of the approach and the novelty of its proposed framework in comparison to past surveys.
• Also, the paper claims that it employs a "five-stage methodology," but it does not provide sufficient information on how this methodology was derived from the literature.
• Moreover, it lists multiple topics, but there is no clear indication of quantifiable outcomes, patterns/trends, or gaps in the literature reviewed.

 

2. Introduction: It is necessary to explicitly compare its scope, methodology, and contributions with the top 8-10 surveys that are widely used in this field.

 

3. Literature Search Methodology:

• The PRISMA methodology is described, but it does not provide the actual query strings, date ranges, or inclusion-exclusion rules in a reproducible way.
• The search yielded 60 chosen studies, yet the article lacks a summary table that arranges/categorizes them by task, hardware or methodology (while also including contributions are required for conducting systematic reviews).
• Why the methodology does not account for cloud implementations; thus, the review may be biased because some hybrid edge-cloud works are relevant.

 

4. Historical Background:

• The historical section is too narrative and lacks of quantitative evidence, such as performance improvements, hardware evolution timelines, and parameter/latency comparisons.
• This section contains general information that is not directly linked to the research questions, and therefore, it should be shortened or restructured to emphasize only those transitions that impact embedded AI.

 

5. Core Concepts & Metrics:

• It does not have a structured way (taxonomy) of categorizing model-level, system level, and hardware-based metrics.
• There are several crucial terms (such as TOPS/W/MHz, GOPS/4 GHz), but they are not critically discussed in terms of weaknesses and inconsistencies.

 

6. State of the Art – Optimization Techniques:

• You should provide a comparative table that summarizes compression ratios, accuracy drops and hardware compatibility across studies.
• Several important emerging techniques, such as NAS and joint pruning-quantization optimization, mixed-precision inference, have been mentioned but not critically summarized.

 

7. State of the Art – Architectures:

• The architecture overview combines classification, detection, transformers, and compact networks without a consistent comparison framework (e.g, accuracy-latency trade-offs; memory footprints or hardware suitability).
• Also, the architectural overview does not provide common benchmarks for all hardware platforms or a single set of models, which limits the reported outcomes.

 

8. Hardware & Frameworks:

• There is no organized/structured comparative analysis, which contains aspects like (peak power, real-world power level, memory bandwidth, latency scaling, and supported bit-widths).
• Moreover, a summary table that compares supported precision values, operator coverage and hardware compatibility, as well as performance of edge workloads, is not included.

 

9. Discussion and Future Work:

• The discussion does not present a comprehensive critical analysis of the studied studies and fails to clearly articulate the convergence trends, contradictory findings, or methodological weaknesses of current literature.
• The future work outlines broad research objectives; it does not transform them into practical research questions or issues with implementation that are grounded in the evidence reviewed.

 

10. The paper has several grammar and typographical issues need to be addressed.

 

Comments on the Quality of English Language

The paper has several grammar and typographical issues need to be addressed.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report (Previous Reviewer 4)

Comments and Suggestions for Authors

The revised manuscript adequately addresses previous review concerns and demonstrates significant improvement in both the conceptual organization and empirical grounding of its primary methodology. The main contribution is the proposal and formal validation of a five-stage methodology for optimizing and deploying neural networks in resource-constrained embedded systems, supported by a new real-world case study and substantial clarifications. While the five-stage framework is now empirically grounded, its novelty is still moderate as the true innovation lies in systematization rather than in new technical mechanisms. The methodology leverages existing state-of-the-art practices in a structured format.

Overall, the revised manuscript merits a positive evaluation for its clarity, depth, and demonstration of real-world methodology application. The addition of empirical analysis and constructive organization resolve the deficiencies identified in previous round.

Although the discussion of emerging techniques is expanded, future trends (such as TinyML, sub-4-bit quantization, on-device learning, and neuromorphic computing) are realistically outlined as medium/long-term priorities. The analysis is robust, but practical timelines and barriers could be discussed in greater detail.

The authors should consider broadening the set of deployment case studies and providing further quantification of how their methodology compares to alternative frameworks, which would add depth to their claims of systematic advantage.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript offer guidances in designing a optimal deep learning system for developers. However, the five-stage methodology from requirement definition to final deployment is a commonly known processing procedure. Furthermore, the widely used models are familiar for researchers. 

1. In Table 4, which datasets are used for the experimental results such as MAP of different models?
2. For different tasks, the metrics are different. For example, Map50 for object detection, Rank1 and Rank5 for individual identification, etc.
3. The range of review is  too broad, this makes the analysis and discussion not profound enough.

Author Response

Please refer to the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Journal: Electronics (ISSN 2079-9292)

Manuscript ID: electronics-3893770

Title: Edge AI in Practice: A Survey and Deployment Framework for Neural Networks on Embedded Systems

Authors: Ruth Cordova-Cardenas * , Álvaro Gutiérrez * , Daniel Amor *

 

The article is an exhaustive review of Edge AI, focusing on how neural networks can be efficiently deployed on embedded systems with high memory, power, and processing requirements. It reviews state of the art model compression techniques, inference optimization methods, hardware platforms, and software frameworks, and then proposes a practical five-stage methodology to guide developers from requirement definition to final deployment of optimized models at the edge. After reviewing the manuscript, my specific comments are below:

 

 

1. Abstract:

• How does the proposed five-stage methodology differ from existing surveys?
• Highlight and identify the gaps or future research directions?
• To demonstrate the scope of the review, include quantitative evidence (number of papers examined, timeframe, or evaluation measures).

 

2. Introduction and Background:

• Include recent references on Edge AI and how this review advances beyond them?
• Express the criteria for selecting studies to be reviewed.
• Consider moving a portion of the early history of ANNs to an additional section and emphasizing the advancements in embedded Edge AI during the last decade.

 

3. Methodology:

• Give a methodology for the identification, filtering and classification of reviewed papers (databases, search terms and timeframe). Now, the review is seen as narrative rather than systematic.
• Add a PRISMA-style flow diagram or table can be used to summarise the number of papers at each selection stage and increase reproducibility.

 

4. Core Concepts and Metrics:

• The section on metrics does not make any reference to standardized benchmarks (such as MLP Perf Tiny or Edge AI Benchmark)?
• Some metrics are established but not demonstrated on actual systems. Please include a similar number that shows latency, throughput, energy consumption, and memory usage for at least two platforms.

 

• State of the Art Techniques, Architectures and Frameworks:
• In the model compression section, Table 2 mixes advantages and disadvantages but skips recent hybrid techniques (e.g., quantization-aware pruning, neural architecture search for compression). Include them or explain why they are excluded.
• YOLO versions are the primary focus of the CNN section, but fails to showcase transformer-based vision models like MobileViT and EdgeFormer. Give at least one paragraph that compares these emerging models with CNN.?.
• The RNN and transformer discussion should contain performance trade-offs for edge hardware, rather than solely focusing on general characteristics. Provide a resource utilization table (FLOPS and latency) for RNN vs. transformer models on typical MCUs/NPUs.
• Only DeepliteRT is cited in the frameworks section of this article, not any other framework such as Apache TFLM, Glow or TVM. Provide a brief overview of these or explain why they are not included.

 

5. Proposed Five-Stage Methodology:

• Explain how the five stages are utilized in a real embedded deployment by providing at least one use-case example or hypothetical scenario to illustrate their application.
• Enhance the feedback-loop explanation by showing how metrics feed back into earlier stages (Figure 2 is currently generic)..

 

6. Discussion and Future Directions:

• The limitations posed by tools and datasets, such as immaturity and lack of standardization, are not critically considered when discussing emerging trends?
• Add a "Research Agenda" section that details/summarizes the top 4-5 open challenges (e.g, security, energy-conscious NAS, standardized metrics) and practical recommendations for future work.

 

7. There are some recurrent grammatical and typographical issues need to be addressed.

 

8. Follow the Electronics journal reference style exactly.

Comments on the Quality of English Language

There are some recurrent grammatical and typographical issues need to be addressed.

Author Response

Please refer to the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

A very good and valuable work, which presents important issues in a very precise manner. Congratulations on a job well done. 

However, I find that the work lacks a methodology for selecting the analyzed literature, i.e., which databases were searched using which keywords and how the articles presented in the work were selected from the returned results. 

In summary, I believe that after minor corrections, the work can be published. .

Author Response

Please refer to the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

Concern 1 (Most important)

The five-stage methodology proposed in Section 6 represents the manuscript's primary contribution. While individual optimization techniques are well-established, the structured integration into a practical deployment methodology offers moderate novelty. The framework's value lies in systematizing the complex decision-making process rather than introducing fundamentally new techniques.
However, the section lacks experimental validation of the proposed five-stage methodology:

• No case studies demonstrating the framework's effectiveness
• No comparison with existing deployment approaches
• No quantitative metrics proving the methodology's advantages

The authors must provide at least 2-3 detailed case studies showing the methodology applied to different scenarios (e.g., autonomous vehicles, medical devices, IoT sensors).

This is fundamentally a review paper but claims to propose a "methodology" without empirical validation. For this reason, the authors should either:

• Reframe as a pure survey paper, or
• Add substantial experimental sections validating their methodology
• Provide detailed implementation guidelines with performance benchmarks

 

Concern 2

The paper proposes a thorough coverage of optimization techniques (pruning, quantization, knowledge distillation) with detailed comparison tables. However, some hardware accelerator metrics discussion (Section 3.2.2) relies heavily on theoretical peak values without sufficient real-world context.

 

Concern 3

The energy efficiency comparisons in Table 12 need more recent data (some references are from 2018-2020).  Plenty of works have been proposed in the past five years considering the exponential outspreading of AI.

 

Concern 4

The discussion of recent transformer adaptations for edge devices is very limited. The same happens for the coverage of emerging quantization techniques beyond binarization/ternarization. Future trends are well-identified but need more critical analysis of feasibility timelines.

 

Comments on the Quality of English Language

The English language quality requires improvement. Specific issues include:

• Inconsistent technical terminology usage
• Several grammatical errors throughout 
• Some sentences are overly complex and could be simplified for clarity
• Technical accuracy in some descriptions needs refinement

Author Response

Please refer to the attachment.

Author Response File: Author Response.pdf