{"id":"da9bfc08-7b8f-4376-aaf7-34c233001715","arxiv_id":"2501.02001","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A dual-threshold, multi-exit local/server co-inference framework with a channel-adaptive offloading policy is proposed and tested on retinal images for rare-event classification.","lead":"An edge-AI paper proposes a dual-threshold, multi-exit scheme that detects rare events locally and offloads only confident tail cases to a server, claiming lower communication overhead and higher rare-event accuracy. A generalist might read it to see how 6G edge inference is balancing energy, bandwidth, and mission-critical detection.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The optimization analysis treats α→∞ logistic functions as if they had the finite α=1 derivative bounds; with the stated step-function model the Lipschitz, weak-convexity, and strong-convexity claims are unproven, so the central threshold-optimization guarantee fails.","rationale":"The reader's weakest assumption correctly identifies the central problem: the paper defines the threshold indicators through α→∞ logistic functions but analyzes them with α=1 derivative bounds, invalidating the Lipschitz, weak-convexity, and strong-convexity results. My read agrees with that assessment. I additionally find an independent flaw in Appendix C: the paper equates gradient Lipschitzness with weak convexity, which is not valid; Nesterov's inequality supplies only an upper quadratic bound, not the lower bound needed for weak convexity. Even if the authors were to replace the step functions by finite-α sigmoids, the claimed strong convexity of ft would still require a separate proof that is absent. The threshold-optimization algorithm and the 'equivalent strongly convex' reformulation are the central advertised contributions, so this is not a peripheral technicality. The experimental section shows plausible gains, and I credit the authors for reporting real CNN architectures and datasets, but the experiments alone cannot support the theoretical optimality and convergence claims. No machine-checked proof or reproducible code is provided that would independently verify the optimization analysis. Therefore the appropriate verdict remains REJECT; a revised paper that either analyzes the finite-α smoothed problem with explicit α-dependent constants and correct weak-convexity arguments, or honestly repositions the optimization as heuristic, could be reconsidered.","tokens_in":20234,"tokens_out":4175,"duration_ms":40695,"concrete_test":"Independently recompute the Hessian of ft(βℓ,βu) in Eq. (24) using the actual derivative bounds for σ(y)=1/(1+e^{-αy}), namely |σ'|≤α/4 and |σ''|≤α²√3/6. Substitute these into the Appendix A calculation of γ and observe that γ diverges as α→∞, so Lemma 3's constants are not finite for the model stated in Eq. (6). Then, for a finite α (e.g., α=10), numerically check Lemma 4's weak-convexity claim: sample pairs (βℓ,βu), compute the minimum eigenvalue of ∇²[f_acc(β) + (γ/2)||β||²], and verify whether it is nonnegative on a grid. If any negative eigenvalue appears, weak convexity fails even with a smooth sigmoid approximation, confirming that Proposition 1's strong-convexity conclusion is unsupported.","verdict_should_be":"REJECT","load_bearing_attack":"The load-bearing step is the smooth optimization analysis. In Eq. (6), σ(y)=1/(1+e^{-αy}) with α→∞, so the indicator functions in Eqs. (5), (7), and (8) are Heaviside step functions. Yet Appendix A derives bounds using |σ(x)|≤1, |σ'(x)|≤1/4, and |σ''(x)|≤√3/6, which are the α=1 logistic-function bounds. For finite α, the actual bounds are |σ'|≤α/4 and |σ''|≤α²√3/6; both diverge as α→∞. Consequently, the Lipschitz constant γ in Lemma 2 is not finite for the model actually defined in the paper, and the Lipschitz claims in Lemma 3 and the smoothness/weak-convexity claims in Lemma 4 are not established.\n\nThere is also an internal error that would survive smoothing with finite α: Appendix C infers γ-weak convexity from γ-smoothness by citing Nesterov's inequality (50). That inequality is an upper quadratic bound (f(y) ≤ f(x)+⟨∇f(x), y−x⟩ + (γ/2)||y−x||²), which follows from gradient Lipschitzness; it does not imply the lower quadratic bound required for weak convexity. Therefore the proximal-penalty function ft(βℓ,βu) in Eq. (24) is not shown to be strongly convex, Proposition 1 is unsupported, and the convergence rate in Remark 1 and the optimal-threshold structure in Proposition 2 do not follow. Since the abstract and the conclusion both rest on the 'equivalent strongly convex' reformulation, this is a fatal gap in the paper's central theoretical claim. The empirical comparisons may be suggestive, but they do not repair the analytical core.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an event-triggered edge co-inference framework for rare-event classification. A lightweight on-device CNN with early-exit blocks applies two confidence thresholds (βℓ and βu) so that head events are resolved locally while tail events are offloaded to an edge server for refined multi-class classification. The authors formulate a constrained optimization problem (P1) for choosing the thresholds, claim that the non-convex objective can be transformed into an equivalent strongly convex problem, and present an online proximal-gradient algorithm. Experiments on retinal images compare the dual-threshold detector with single-threshold, terminal-detection, and ideal-case schemes under different class-imbalance ratios, energy budgets, and SNR conditions.","tokens_in":20597,"tokens_out":8309,"duration_ms":82205,"significance":"If the theoretical core were valid, the paper would be a useful practical contribution: it couples early exiting with channel-adaptive offloading in a way that targets rare events, and it validates the architecture on real medical data with multiple CNN backbones and sensible baselines. The experimental design includes separate validation and held-out test sets, five test groups, and comparisons with three baselines, which are strengths of the empirical part. However, the optimization analysis is the load-bearing part of the contribution, and the current proofs do not establish the claimed guarantees. The experimental comparison is suggestive but cannot repair the analytical core, and no code or data are provided to support independent reproduction.","major_comments":[{"comment":"Eq. (6) defines σ(y) = 1/(1+e^{-α y}) with α→∞, so the indicator functions in Eqs. (5), (7), and (8) are Heaviside step functions, which are not differentiable and not Lipschitz. Yet Appendix A proves Lemma 2 using the bounds |σ(x)|≤1, |σ'(x)|≤1/4, and |σ''(x)|≤√3/6, which hold only for the α=1 logistic function. For finite α the derivative bounds scale as α/4 and α²√3/6, and they diverge as α→∞. Consequently, the Lipschitz constants in Lemmas 2–4, the strong-convexity parameters in Proposition 1, and the convergence rate in Remark 1 are not established for the model actually defined in the paper.","section":"Section IV-A, Eq. (6); Appendix A"},{"comment":"The passage from gradient Lipschitzness to weak convexity is not demonstrated. Eq. (50) is the standard one-sided upper quadratic bound; weak convexity requires the lower quadratic bound f(y) ≥ f(x) + ⟨∇f(x), y−x⟩ − (γ/2)||y−x||². Since the proof relies only on the displayed upper bound, the claimed weak-convexity of facc in Lemma 4, and hence the strong-convexity parameter η in Proposition 1, is unsupported unless the lower-bound argument is supplied explicitly.","section":"Appendix C, Eq. (50)"},{"comment":"Proposition 2 is not a consequence of the optimization problem P1. Eq. (31) merely counts how many events can be offloaded with the residual energy budget at the current transmission rate, and Eq. (30) restates the feasibility condition of Lemma 1. Neither formula uses the data-volume constraint θ, the offloading probability Poff, or the accuracy trade-off in P1, so labeling this the 'optimal offloading policy' is not justified. At most it is an energy-feasibility bound on the number of offloadable events.","section":"Section V-B3, Proposition 2, Eqs. (30)–(31)"},{"comment":"The paper claims that the original non-convex problem is transformed into an equivalent strongly convex one, but no equivalence theorem is given. The proximal-penalty function (24) is not shown to have the same solution set as P1, and the output selection in Algorithm 1 (line 16, minimizing ||β^{t+1} − β^t|| across iterations) is not accompanied by any convergence or feasibility guarantee for the original constraints (20)–(21). Thus the optimality of the proposed online algorithm is unproven even under the strong-convexity assumptions.","section":"Section V-B2, Algorithm 1, Eq. (24)"},{"comment":"The experimental section reports average accuracies without error bars, confidence intervals, or significance tests, despite the five-group test protocol described in Section VI-A. The claimed superiority of the dual-threshold mechanism over the baselines is therefore not statistically quantified. In addition, the single-threshold baseline is dropped from the imbalanced-ratio comparison in Section VI-C, and the 'ideal case' is an upper bound rather than a fair baseline, which limits the strength of the comparison.","section":"Section VI, Figs. 4–7"}],"minor_comments":[{"comment":"The logarithm base in the data-rate expression is inconsistent: Eq. (3) writes log(1+SNR), while Lemma 1, Eq. (22), and the experimental settings use log₂(1+SNR). Please make the base explicit and consistent throughout.","section":"Eq. (3) vs. Eq. (22) and Section VI-A"},{"comment":"The function fmiss(β) appears in the definition of ft but is never defined; the objective in P1 is facc(βℓ, βu). Please clarify whether fmiss is meant to be 1 − facc or a different loss.","section":"Algorithm 1, line 8"},{"comment":"The constant k (and hence k²) in the Lipschitz constant γ is never defined in the main text or in Appendix A, so the numerical value of γ cannot be evaluated.","section":"Lemma 2 and Appendix A"},{"comment":"The symbol B is used both for the bandwidth and for the constant defined in Eq. (28), which is confusing; please use a different symbol for one of these quantities.","section":"Proposition 1, Eqs. (25)–(28)"},{"comment":"There are numerous typographical and grammatical errors, including 'refered', 'early existing', 'trade-off indicates a trade-off', 'adaptes', and 'purposed'. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The expression for facc(βℓ, βu) is written in a confusing way because it contains M · M/Mtail without simplification; please state the final form explicitly and define Ptail before using it.","section":"Eq. (15)"}],"recommendation":"reject","confidential_remarks":"The manuscript is not publishable in its current form because the central optimization theory is invalid as written: the step-function model in Eq. (6) is incompatible with the smoothness assumptions used in Appendices A–D, and Proposition 2 is a restatement of the energy budget rather than a derived optimal policy. The empirical study is suggestive but does not repair the analytical core. If the authors later resubmit, the theory would need to be redeveloped with a finite smoothing parameter and explicit dependence of all constants on that parameter, and the experiments would need error bars and a justification of the dropped baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the dual-threshold early-exit plus offloading design is a sensible system-level idea and the experiments point in the right direction, but the optimization analysis that carries the paper's central claim is not valid. I'd treat this as a promising draft whose theory needs to be redone, not as a publishable result as it stands.\n\nThe genuinely new piece is the two-threshold uncertainty region on confidence scores in a multi-exit device, combined with channel-adaptive offloading of the tail class. That is a useful combination for the 6G mission-critical edge-inference setting. The experiments on two CNN backbones with a retina dataset show the dual-threshold scheme consistently lowering missing probability for the minority class relative to single-threshold and terminal detection, and the energy/SNR curves behave as one would expect. That is real evidence, though it is not overwhelming.\n\nThe soft spots are serious. Eq. (6) defines the logistic function with alpha→infinity, so all the indicator functions are Heaviside steps. Appendix A then uses the alpha=1 bounds |sigma'|<=1/4 and |sigma''|<=sqrt(3)/6. The actual bounds scale with alpha and alpha^2, so they diverge in the model as written. The Lipschitz, weak-convexity, and strong-convexity lemmas are unproven for the actual functions.\n\nThere is a second, independent mistake that survives smoothing. Appendix C cites Nesterov's inequality, which gives an upper quadratic bound, and uses it to conclude weak convexity, which requires a lower quadratic bound. That is a categorical error, so Proposition 1 and Algorithm 1's convergence guarantee do not follow even for a finite smooth sigma. Proposition 2's offloading formula is also basically the energy-budget identity rewritten as a policy; calling it the 'structure of optimal policy' is a restatement, not a derived result.\n\nThe experimental section could be stronger: no error bars, no run-to-run variance, the single-threshold baseline is dropped in the imbalanced-data experiments, and despite the abstract's claim of outperforming 'existing edge-inference approaches,' the plain split-inference baseline is not included. Those are fixable, but they add to the overall gap between claims and evidence.\n\nWho is this for? People working on device-server co-inference, early exiting, and 6G resource allocation will find the architecture worth discussing. But the math, as written, should not be relied on. A serious editor should send it to review, because the defects are specific and the system concept is interesting; the authors can either repair the proofs with a fixed finite alpha approximation or reposition the algorithm as a heuristic without the convergence claims. I would not cite the optimization results in their current form.","headline":"Sensible dual-threshold early-exit/offloading architecture and suggestive experiments, but the optimization theory is invalid as written — the central strong-convexity and convergence claims do not hold.","tokens_in":21118,"tokens_out":2887,"would_cite":false,"duration_ms":25837,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A channel-adaptive, event-triggered co-inference framework with dual-threshold early exiting classifies rare events more accurately than existing edge-inference approaches while transmitting far less data.","keywords":["rare event detection","edge inference","early exiting","computation offloading","dual threshold","channel-adaptive policy","split inference","long-tailed distribution"],"falsifier":"Compute the Hessian of the indicator functions in Eqs. (5) and (8) using the finite logistic function with $\\alpha \\to \\infty$: if the second derivative exceeds the constant $k^2$ used in Lemma 2, the Lipschitz, weak-convexity, and strong-convexity claims do not follow. A simpler check: run Algorithm 1 on a fixed dataset and compare its optimized thresholds to an exhaustive grid search over $(\\beta_\\ell, \\beta_u)$ under the same constraints; any significant gap in objective value would show the algorithm is not finding the claimed optimum.","tokens_in":19985,"feed_emoji":"📡","tokens_out":5463,"duration_ms":49277,"temperature":0.7,"pith_summary":"Rare but critical events (medical emergencies, security incidents) are exactly the cases where offloading to an edge server matters most, yet standard split inference transmits features for every event and floods the link. This paper proposes a co-inference framework in which a lightweight local model checks each event against two confidence thresholds at each early-exit block: confident head events stop locally, confident tail events are offloaded to the server for fine-grained classification, and low-confidence events move to the next block. The paper claims that optimizing these dual thresholds channel-adaptively maximizes end-to-end rare-event accuracy subject to data-volume and energy constraints, and that the non-convex threshold optimization can be rewritten as a strongly convex problem solved online. If correct, the framework would let mission-critical edge systems respond to rare events sooner and with less radio traffic than existing early-exit or terminal-detection baselines. Experiments on retinal image datasets with ShuffleNetV2/MobileNetV2 local models and a ResNet50 server support the claim.","feed_headline":"Dual-threshold offloading lifts rare-event accuracy, cuts traffic","feed_subtitle":"A device-server split that offloads only uncertain events beats standard early-exit schemes under tight energy budgets.","key_machinery":"The load-bearing mechanism is the dual-threshold multi-exit confidence detector. At each exit block $n$ the Softmax tail-confidence $C^{\\mathrm{tail}}_n(m)$ is compared with a lower threshold $\\beta_\\ell$ and an upper threshold $\\beta_u$; the paper encodes the resulting head/tail decisions as asymptotic indicator functions built from logistic functions $\\sigma(y) = 1/(1+e^{-\\alpha y})$ with $\\alpha \\to \\infty$, so that products of these sigmoids select the first block at which the event exits. These indicators convert event counts (miss probability, false-alarm probability, offloading probability, energy) into differentiable functions of the thresholds, which is what allows the optimization. The second piece is the proximal-point penalty method: adding a proximal term and quadratic penalties for the data-volume and energy constraints makes the objective strongly convex with explicit smoothness and convexity parameters, enabling accelerated gradient updates and a lookup table of optimal thresholds indexed by SNR.","core_discovery":"On the paper's own terms, the central discovery is that a dual-threshold early-exit detector, rather than a single confidence threshold, resolves the conflict between missing rare tail events and wasting communication on frequent head events. The device's intermediate classifiers output a tail-confidence score; scores below $\\beta_\\ell$ are labeled head and stop, scores above $\\beta_u$ are labeled tail and are offloaded, and scores between the two continue through deeper blocks. The paper derives indicators for correct head/tail detection with these thresholds, defines the offloading probability as a function of them, and formulates the end-to-end tail-classification accuracy maximization under data-volume and energy constraints. It then asserts that the non-convex problem becomes an equivalent strongly convex proximal-penalty problem, giving a convergence rate that improves with SNR, and that the resulting optimal policy has threshold-based structure: no offloading below an SNR feasibility bound, and an SNR-dependent number of offloaded events above it. Experiments with real medical datasets and two CNN families are presented as evidence that this mechanism beats single-threshold and terminal detection in tail-class accuracy while reducing computation and communication.","pith_inferences":["Editorial inference: the same dual-threshold logic could be applied to multi-class long-tail problems, with per-class upper and lower thresholds, though the paper only demonstrates binary head/tail detection.","Editorial inference: because the analysis models decision functions as infinitely sharp step functions while the proofs require bounded derivatives, a robustness check with finite-$\\alpha$ sigmoids, or with the actual neural network confidences, would show whether the convergence guarantees survive outside the idealization.","Editorial inference: the tradeoff between miss probability and offloading probability resembles selective prediction; connecting the thresholds to conformal prediction could provide distribution-free coverage guarantees for rare-event detection."],"forward_implications":["Events that are confidently head can stop at shallow blocks locally, so most normal traffic never consumes uplink bandwidth; only rare, uncertain events trigger offloading.","The SNR feasibility condition gives an explicit channel requirement: below it, no event should be offloaded, because transmission energy would break the budget.","Above the feasibility SNR, the optimal offloading policy is threshold-based: the number of offloaded events is set by the available energy, data rate, and optimized confidence thresholds.","Higher SNR improves the convergence rate of the online threshold optimization, so the algorithm adapts faster when the channel is good.","On medical retinal datasets, the dual-threshold scheme lowers tail-event miss probability and raises tail-class accuracy compared with single-threshold early exiting and terminal detection under the same offloading and energy constraints."],"supporting_citations":[{"why":"Supplies the early-exiting architecture (BranchyNet) that the dual-threshold detector extends.","marker":"[30]"},{"why":"Source of the dual-threshold idea for head/tail detection in an ANN classifier.","marker":"[34]"},{"why":"Provides the inexact proximal-point penalty method used to turn the non-convex threshold problem into a strongly convex one.","marker":"[37]"},{"why":"ShuffleNetV2 is one of the two local CNN backbones used in the experiments.","marker":"[31]"},{"why":"MobileNetV2 is the other local CNN backbone used in the experiments.","marker":"[32]"},{"why":"Provides the retinal image dataset used to train local and server models and evaluate rare-event accuracy.","marker":"[39]"},{"why":"Gives the energy consumption model for local computation and offloading used in the problem formulation.","marker":"[18]"},{"why":"Supplies the rate/SNR model and task-offloading resource allocation that underpin the feasibility condition.","marker":"[14]"},{"why":"Supplies the standard smoothness and weak-convexity bounds used in the proof of Lemma 4.","marker":"[41]"}],"fun_headline_variants":["Dual-threshold offload lifts rare-event accuracy, cuts traffic","Two thresholds, one smart offload: rare events get edge AI","Event-triggered offloading: only uncertain rare events go to edge","Dual-threshold scheme: local for easy rare events, offload for hard ones","Smart edge AI: offload only uncertain events, keep the rest local"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantees depend on the threshold decision functions being smooth with small bounded first and second derivatives, yet the paper defines them as infinitely sharp step functions, which have no such bounds.","fun_headline_variants_meta":{"raw":{"variants":["Dual-threshold offload lifts rare-event accuracy, cuts traffic","Two thresholds, one smart offload: rare events get edge AI","Event-triggered offloading: only uncertain rare events go to edge","Dual-threshold scheme: local for easy rare events, offload for hard ones","Smart edge AI: offload only uncertain events, keep the rest local"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000661,"raw_usage":{"total_tokens":3057,"prompt_tokens":1018,"completion_tokens":2039,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":1941}},"tokens_in":634,"tokens_out":2039,"duration_ms":13507,"temperature":1.0,"reasoning_tokens":1941,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:40:01.435138+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the Hessian of the indicator functions in Eqs. (5) and (8) using the finite logistic function with $\\alpha \\to \\infty$: if the second derivative exceeds the constant $k^2$ used in Lemma 2, the Lipschitz, weak-convexity, and strong-convexity claims do not follow. A simpler check: run Algorithm 1 on a fixed dataset and compare its optimized thresholds to an exhaustive grid search over $(\\beta_\\ell, \\beta_u)$ under the same constraints; any significant gap in objective value would show the algorithm is not finding the claimed optimum.","supporting_citations":[{"cited_title":"Branchynet: Fast inference via early exiting from deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the early-exiting architecture (BranchyNet) that the dual-threshold detector extends."},{"cited_title":"Bayesian ann classifier for ecg arrhythmia diagnostic system: A comparison study,","cited_arxiv_id":null,"evidence_quote":"Source of the dual-threshold idea for head/tail detection in an ANN classifier."},{"cited_title":"Complexity of an inexact proximal- point penalty method for constrained smooth non-convex optimization,","cited_arxiv_id":null,"evidence_quote":"Provides the inexact proximal-point penalty method used to turn the non-convex threshold problem into a strongly convex one."},{"cited_title":"Shufflenet v2: Practical guidelines for efficient cnn architecture design,","cited_arxiv_id":null,"evidence_quote":"ShuffleNetV2 is one of the two local CNN backbones used in the experiments."},{"cited_title":"Mo- bilenetv2: Inverted residuals and linear bottlenecks,","cited_arxiv_id":null,"evidence_quote":"MobileNetV2 is the other local CNN backbone used in the experiments."},{"cited_title":"Code-free deep learning for multi-modality medical image classification,","cited_arxiv_id":null,"evidence_quote":"Provides the retinal image dataset used to train local and server models and evaluate rare-event accuracy."},{"cited_title":"Dynamic computation offloading for mobile-edge computing with energy harvesting devices,","cited_arxiv_id":null,"evidence_quote":"Gives the energy consumption model for local computation and offloading used in the problem formulation."},{"cited_title":"Optimal task offloading and resource allocation in mobile-edge computing with inter-user task dependency,","cited_arxiv_id":null,"evidence_quote":"Supplies the rate/SNR model and task-offloading resource allocation that underpin the feasibility condition."}],"review_version":1}