{"id":"0978a349-76be-4d3c-a177-84c2085edb00","arxiv_id":"1909.00560","paper_version":2,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper proposes a taxonomy and research roadmap for Edge Intelligence, dividing it into AI for edge and AI on edge, without presenting new empirical results.","lead":"This paper is a survey that sorts the intersection of edge computing and artificial intelligence into two buckets: using AI to improve edge systems, and running AI models on edge devices. It offers newcomers a roadmap of the field and lists open challenges, but it contains no experiments or new algorithms.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Taxonomy overlap: several works cited as 'AI for edge' are also 'AI on edge', so the central division may not be a partition.","rationale":"The reader's verdict of UNVERDICTED is appropriate for a survey and position paper with no empirical claim to test. However, the stress-test reveals a more specific and more central weakness than the reader's stated concern about literature selection bias. The paper's main contribution is the proposed division of Edge Intelligence into AI for edge and AI on edge, and the road-map in Fig. 1 depends on those two branches being meaningfully separable. The paper's own examples blur that separation: federated learning via over-the-air computation appears both as an intelligent wireless networking technique (AI for edge) and as a model-training framework or adaptation (AI on edge). The text even admits this dual classification. Because the central claim is about how to organize the field, an internally inconsistent taxonomy undermines the paper's utility even if the literature sample were perfectly representative. I therefore partially agree with the reader's diagnosis: the selection-bias concern is real but secondary. A conditional verdict is appropriate: the paper's contribution can be salvaged by clarifying that the two branches are analytical perspectives rather than a disjoint partition, and by adding an explicit treatment of boundary cases such as learning-driven communication for federated training. This is a conceptual revision, not a scientific retraction, so CONDITIONAL better reflects the needed adjustment than REJECT or UNVERDICTED.","tokens_in":19898,"tokens_out":6079,"duration_ms":52874,"concrete_test":"Build a cross-classification table from the paper's own citations: list every work cited in Section IV-A (AI for edge) and every work cited in Section V-A and Table II (AI on edge), then tag each work with the definitions in Section I. Count how many works satisfy both definitions, especially [38], [39], [40], and [41]. If more than two distinct works from the paper's own sample are dual-classified, the claimed division is not a partition and the road-map requires an explicit boundary rule, a third co-design category, or a revised claim that the two branches are complementary perspectives rather than disjoint research directions.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The road-map's utility depends on the two branches being separable: the Abstract says Edge Intelligence is divided into AI for edge and AI on edge, and Fig. 1 presents them as distinct left/right branches. The paper itself violates this partition. In Section IV-A1, Learning-driven Communication, including over-the-air aggregation for Federated Learning [38], [39] and importance-aware retransmission [40], is presented as AI for edge under Wireless Networking. Yet these same mechanisms directly serve model training on the edge, which is the defining concern of AI on edge in Section III-C and Section V-A. The text explicitly concedes the overlap: after describing [39], it states that this contribution 'can also be categorized as Model Adaptation in AI on edge, but it accelerates Federated Learning from the perspective of fast data acquisition.' If a representative set of cited works satisfies both definitions, then the boundary is not a property of the work but a choice of viewpoint, and the central claim that Edge Intelligence 'should be divided into' these two directions is correspondingly weakened. A reader cannot use the road-map to classify a new paper without already knowing which side the authors intend. This internal-consistency concern is more load-bearing than the coverage or selection-bias issue: even a complete literature sample would still leave these dual-classified works unresolved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a taxonomy for the emerging field of Edge Intelligence, dividing it into \"AI for edge\" (Intelligence-enabled Edge Computing) and \"AI on edge\" (Artificial Intelligence on Edge). It presents a research road-map (Fig. 1) with hierarchical decompositions: the AI-for-edge side is organized by Topology, Content, and Service layers, while the AI-on-edge side is organized by Model Adaptation, Framework Design, and Processor Acceleration. The paper reviews selected state-of-the-art works in several subtopics (wireless networking, service placement and caching, computation offloading, model compression, conditional computation, algorithm asynchronization, and thorough decentralization) and discusses grand challenges in both directions. The central claim is that this two-way division provides a useful organizing framework for an otherwise fragmented literature.","tokens_in":20148,"tokens_out":3189,"duration_ms":35122,"significance":"If the proposed taxonomy and road-map are accepted, they could serve as a useful conceptual scaffold for researchers entering Edge Intelligence, which is a real and rapidly growing area. The paper's distinction between using AI to improve edge systems and running AI models on edge systems is a natural and potentially clarifying framing. The paper also usefully collects representative recent work in both directions and highlights open challenges. However, the paper is a position/survey paper with no systematic methodology: it contains no derivation, no data, and no formal evaluation of its taxonomy. Its value rests entirely on whether the proposed categories are coherent and whether the selected examples are representative. The paper does not provide a systematic literature-selection protocol, and, more importantly, the two central categories overlap in the very examples used to illustrate them, which weakens the claim that Edge Intelligence \"should be divided into\" these two directions.","major_comments":[{"comment":"The central dichotomy between AI for edge and AI on edge is not a partition as claimed. In Section IV-A1, Learning-driven Communication is presented as AI for edge, yet the works cited there ([38], [39], [40], [41]) directly serve model training on the edge, which is the defining concern of AI on edge in Section III-C and Section V-A. The paper explicitly concedes this for [39], stating that \"This contribution can also be categorized as Model Adaptation in AI on edge, but it accelerates Federated Learning from the perspective of fast data acquisition.\" If a representative set of cited works satisfies both definitions, the boundary is a choice of viewpoint rather than a property of the work, and the road-map in Fig. 1 cannot be used to classify a new paper without prior knowledge of the authors' intent. This undermines the abstract's claim that Edge Intelligence \"should be divided into\" these two directions. The authors should either provide classification criteria that make the branches mutually exclusive, or reframe the two directions as complementary facets of a design space whose overlap is explicitly acknowledged and analyzed.","section":"Section IV-A1 and Fig. 1"},{"comment":"The selection of \"hot topics\" used to exemplify the road-map is not justified. Section IV-A states, without further support, that \"we only focus on these topics\" (wireless networking, service placement and caching, computation offloading), and Section V-A similarly narrows attention to Model Adaptation. There is no systematic literature-search protocol, inclusion criteria, or coverage metric, so the paper's claims about the \"state of the art\" and the completeness of its road-map rest on an unstated assumption that the chosen examples are representative. If the selection is biased, the road-map and the associated future-research guidance could mislead readers. The authors should describe how the surveyed works were identified and selected, or explicitly scale back the claim from \"state of the art\" to \"illustrative examples.\"","section":"Section IV-A and V-A"}],"minor_comments":[{"comment":"The phrase \"Model Spitting and Model Partitioning\" appears to be a typo; it should read \"Model Splitting and Model Partitioning.\"","section":"Section V-A (introductory paragraph)"},{"comment":"References [19] and [44] refer to the same paper (Chen et al., \"Spatiotemporal edge service placement: A bandit learning approach\"). The duplicate should be removed and citations should be unified.","section":"References [19] and [44]"},{"comment":"Table II lists \"ProNN (based on Stochastic Neighborhood Compression)\" but the text and reference [61] use the name \"ProtoNN.\" The table entry should be corrected.","section":"Table II"},{"comment":"The sentence \"Algorithm Asynchronization trys aggregating local models in an asynchronous way\" contains a typo: \"trys\" should be \"tries.\"","section":"Section III-C2"},{"comment":"Several self-citations ([20], [21], [25], [29], [45]) are used as representative examples of the proposed directions. In a survey with no explicit inclusion criteria, this creates an impression of selection bias, although self-citation itself is not inappropriate. The authors may wish to note how their own works were selected.","section":"Section IV-A1"}],"recommendation":"major_revision","confidential_remarks":"The paper's central taxonomy issue is real but fixable: the authors can either sharpen the definitions to make the branches exclusive or explicitly treat the overlap as a feature and discuss its implications. The second major comment about selection bias is also addressable by documenting the selection process. The paper is a survey/position paper rather than a technical contribution, so the editor should judge whether its conceptual contribution is sufficient for the journal's readership. The heavy reliance on self-citations in the examples is worth watching but not disqualifying."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a survey/position paper, not a research result, and its main contribution is a vocabulary: divide Edge Intelligence into AI for edge and AI on edge. That split has been useful to people entering the field, and the paper is a decent road-map. But the division is not a partition, and the authors themselves concede several of their examples land on both sides.\n\nWhat it does well: it organizes a lot of recent work into a coherent structure — left side Topology/Content/Service, right side Model Adaptation/Framework Design/Processor Acceleration — and it puts a multi-criteria QoE lens (performance, cost, privacy, efficiency, reliability) on top. The writing is clear, the grand challenges are sensible, and the examples are mostly representative of the subareas. There are no derivations or data to verify; it's a synthesis, and judged as a synthesis it's competent.\n\nSoft spots: the central claim that Edge Intelligence 'should be divided into' the two directions is softer than it looks. In Section IV-A1, Learning-driven Communication — over-the-air aggregation for Federated Learning [38], [39], importance-aware retransmission [40] — is filed under AI for edge, but the same mechanisms directly serve model training on the edge, which is the core of AI on edge. The text says of [39] that it 'can also be categorized as Model Adaptation in AI on edge,' and [41] is noted as sharing concerns with Model Adaptation. So the taxonomy is really two viewpoints rather than an exclusive split. If you're trying to classify a new paper, you often can't tell which side it belongs to without knowing the authors' intent. That is a real internal-consistency issue, and I'd ask the authors to address it.\n\nThe other soft spot is selection bias: the survey says it focuses on 'hot topics' but gives no systematic protocol or inclusion criteria. That's a minor issue for a position paper, but combined with the partition problem it means the road-map is more of a curated map than a faithful representation of the field.\n\nOn circularity: the self-citations are used as illustrative examples of the proposed directions. That's normal for a survey, not a circular argument.\n\nWho it's for: newcomers to edge intelligence who want a quick map; also useful for a reading group discussion about how to slice the field. It does not need to be cited as a technical result, but it's a reasonable survey citation.\n\nRecommendation: send it to peer review. The taxonomy needs a caveat that the two branches are complementary lenses, not a partition, and the example selection should be justified. With that revision, it's a solid survey.\n\nBest.","headline":"Readable survey and position paper on Edge Intelligence whose central two-way split is a useful coarse map but not a clean partition, since the paper's own examples overlap.","tokens_in":20658,"tokens_out":3708,"would_cite":true,"duration_ms":282569,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This survey argues that Edge Intelligence is best organized as two research directions — AI for edge and AI on edge — joined under one quality-of-experience goal.","keywords":["Edge Intelligence","Edge Computing","Artificial Intelligence","AI for edge","AI on edge","Federated Learning","Computation Offloading","Quality of Experience"],"falsifier":"A reader could settle the matter by building a systematic corpus of edge-intelligence publications with explicit inclusion criteria and counting how many fit one of the road-map's cells: if a substantial share of papers falls outside both the AI-for-edge and AI-on-edge branches — or fits both — the claimed division would fail as an organizing framework.","tokens_in":19717,"feed_emoji":"🗺️","tokens_out":7380,"duration_ms":56769,"temperature":0.7,"pith_summary":"Edge Intelligence — the integration of Edge Computing and Artificial Intelligence — has no formal definition, and the survey literature has grown piecemeal. This paper tries to establish a single organizing framework: split the field into \"AI for edge\" (using AI to solve edge-computing optimization problems) and \"AI on edge\" (running model training and inference on edge hardware), then map each direction onto a three-layer road-map with Quality of Experience as the shared goal. If the framework holds, it gives researchers a common vocabulary, exposes gaps (such as participator selection in federated learning), and separates the field's grand challenges into two coherent agendas. A sympathetic reader would take this as a proposal for how future edge-intelligence research should be positioned and compared.","feed_headline":"Survey maps edge intelligence into two AI research tracks","feed_subtitle":"One road-map unifies the fragmented edge-computing-plus-AI literature under a shared quality-of-experience goal.","key_machinery":"The load-bearing object is the research road-map of Fig. 1: a single two-sided diagram with Quality of Experience at the top. Reading left to right, the 'AI for edge' side is decomposed bottom-up into Topology, Content, and Service layers — each with named problems such as wireless networking, service caching, and computation offloading — while the 'AI on edge' side is decomposed top-down into Model Adaptation, Framework Design, and Processor Acceleration. The map does the argument's work by showing that every edge-computing resource decision and every edge-AI model-lifecycle decision can be placed in a named cell, so that 'hot topic' papers become instances of a category rather than isolated results.","core_discovery":"The paper's central claim is that Edge Intelligence is not the simple combination of edge computing and AI but a field with an internal structure, best captured by a two-way division. On one side, AI for edge (Intelligence-enabled Edge Computing) treats AI as an optimization tool: deep neural networks, deep Q-networks, multi-armed bandits, and reinforcement learning supply near-optimal solutions to edge-computing problems in topology (edge-site orchestration, wireless networking), content (data and service provisioning, placement, composition, caching), and service (computation offloading, user-profile migration, mobility management). On the other side, AI on edge (Artificial Intelligence on Edge) asks how the full model lifecycle — training and inference — can run on devices and edge servers with device-edge-cloud synergy, organized top-down into model adaptation (compression, conditional computation, asynchronization, decentralization), framework design (federated learning, knowledge distillation, model splitting), and processor acceleration. Both branches pursue the same top-level objective, Quality of Experience, defined by performance, cost, privacy (security), efficiency, and reliability. If the road-map is right, the field stops being a scatter of disconnected optimization papers and becomes two research programs with their own states of the art and grand challenges.","pith_inferences":["The clean two-way division may blur in practice: work at the communication–learning interface already straddles both branches, so the field may eventually need a third cross-cutting category, or an explicit treatment of such overlaps, as it matures.","The road-map's completeness is testable: a systematic survey with explicit inclusion criteria and a coverage metric could check whether published edge-intelligence work actually falls into the map's cells or spills outside them.","The five Quality-of-Experience criteria could be turned into a scoring rubric for comparing AI-on-edge systems across papers, since most current evaluations report only accuracy and latency."],"forward_implications":["Researchers gain a shared taxonomy: a new result can be classified as AI for edge or AI on edge and located in one layer, which makes comparisons across papers easier.","The road-map exposes empty cells — the paper states that no work yet targets participator selection in conditional computation — thereby pointing to concrete research openings.","Each branch carries its own grand challenge list: model establishment, algorithm deployment, and the optimality–efficiency trade-off on the AI-for-edge side; data availability, model selection, and coordination mechanisms on the AI-on-edge side.","The two branches interact: schemes like over-the-air computation sit in wireless networking (AI for edge) yet accelerate federated learning (AI on edge), so progress in one branch can feed the other.","A shared Quality-of-Experience rubric (performance, cost, privacy, efficiency, reliability) gives both branches a common evaluation language for otherwise disparate results."],"supporting_citations":[{"why":"Supplies the prior six-level definition of Edge Intelligence (cloud–edge co-inference to all-on-device) that this paper positions its own two-way division against.","marker":"[7]"},{"why":"Supplies the competing definition of Edge Intelligence as the capability to execute AI algorithms at the edge, motivating the need for a broader framing.","marker":"[8]"},{"why":"Foundational federated learning paper that anchors the AI-on-edge model-training branch and the privacy criterion of QoE.","marker":"[11]"},{"why":"Provides the data-splitting versus model-splitting taxonomy on which the Framework Design discussion of distributed training rests.","marker":"[30]"},{"why":"The tutorial on efficient DNN processing that the Processor Acceleration branch relies on.","marker":"[34]"},{"why":"The leading example of learning-driven communication, filling the wireless networking cell of the AI-for-edge map.","marker":"[15]"},{"why":"Example of multi-armed-bandit service placement decisions in the Content layer of AI for edge.","marker":"[44]"},{"why":"Example of DQN-based computation offloading for IoT devices, filling the Service layer cell.","marker":"[26]"}],"fun_headline_variants":["Edge intelligence splits into two tracks: AI for edge, AI on edge","Survey frames edge intelligence as AI for edge and AI on edge","Edge AI roadmap: two research paths, one quality-of-experience goal","Unifying edge computing and AI: a two-way division","Edge intelligence: AI as tool and AI as target"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The road-map is only as trustworthy as the hand-picked sample of papers behind it: the authors focus on \"hot topics\" without a systematic search protocol or stated inclusion criteria, so if those examples are not representative, the map and its claims about the state of the art would mislead.","fun_headline_variants_meta":{"raw":{"variants":["Edge intelligence splits into two tracks: AI for edge, AI on edge","Survey frames edge intelligence as AI for edge and AI on edge","Edge AI roadmap: two research paths, one quality-of-experience goal","Unifying edge computing and AI: a two-way division","Edge intelligence: AI as tool and AI as target"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000767,"raw_usage":{"total_tokens":3424,"prompt_tokens":990,"completion_tokens":2434,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":2347}},"tokens_in":606,"tokens_out":2434,"duration_ms":15413,"temperature":1.0,"reasoning_tokens":2347,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:44:54.151892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle the matter by building a systematic corpus of edge-intelligence publications with explicit inclusion criteria and counting how many fit one of the road-map's cells: if a substantial share of papers falls outside both the AI-for-edge and AI-on-edge branches — or fits both — the claimed division would fail as an organizing framework.","supporting_citations":[{"cited_title":"Edge intelligence: Paving the last mile of artificial intelligence with edge computing,","cited_arxiv_id":null,"evidence_quote":"Supplies the prior six-level definition of Edge Intelligence (cloud–edge co-inference to all-on-device) that this paper positions its own two-way division against."},{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Foundational federated learning paper that anchors the AI-on-edge model-training branch and the privacy criterion of QoE."},{"cited_title":"Not Just Privacy: Improving Performance of Private Deep Learning in Mobile Cloud","cited_arxiv_id":"1809.03428","evidence_quote":"The tutorial on efficient DNN processing that the Processor Acceleration branch relies on."},{"cited_title":"Wireless Data Acquisition for Edge Learning: Data-Importance Aware Retransmission","cited_arxiv_id":"1812.02030","evidence_quote":"Example of multi-armed-bandit service placement decisions in the Content layer of AI for edge."},{"cited_title":"Learning- based computation offloading for iot devices with energy harvesting,","cited_arxiv_id":null,"evidence_quote":"Example of DQN-based computation offloading for IoT devices, filling the Service layer cell."}],"review_version":1}