{"id":"56454fc4-8785-478a-a7e9-35eb075197d0","arxiv_id":"2411.11539","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ADE-MI compresses each WiFi device's sensing spectrogram to a channel-capacity-sized feature vector and trains a server-side multi-view classifier, reporting 92% accuracy with roughly 10^4 lower upload latency than raw transmission.","lead":"This paper proposes ADE-MI, a two-stage compression and inference scheme for multi-view WiFi sensing, where each device encodes its Doppler spectrogram into a low-dimensional feature vector sized by the wireless channel capacity before uploading to an edge server. On the Widar3.0 dataset the scheme reports 92% gesture recognition accuracy with upload latency about 10,000 times smaller than sending raw spectrograms.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The two-stage decomposition does not solve P1: minimizing per-device H(Y|Z_k) only shrinks an upper bound on H(Y|Z_1..Z_K), and the conditional independence in Eq. (10) is unverified; complementary views break the surrogate entirely.","rationale":"The reader correctly identifies the conditional-independence assumption in Eq. (10) as fragile, but the deeper load-bearing flaw is the logical leap in Section III.C.1: minimizing each per-device H(Y|Zk) is not equivalent to minimizing the joint H(Y|Z1,...,ZK), regardless of whether Eq. (10) holds. The inequality H(Y|Z1,...,ZK) ≤ min_k H(Y|Zk) only gives a valid upper bound, not a solution. Conditional independence would make the per-device surrogate more defensible, which is why the reader's concern is partially aligned, but it is not sufficient for exactness either. The load-bearing concern is therefore that the paper's central theoretical claim is unsupported, and the paper should be reframed as a heuristic with empirical support. Since the empirical results and latency gains may still have value after such a reframing, the reader's conditional verdict remains appropriate; no change is needed.","tokens_in":9553,"tokens_out":12436,"duration_ms":126895,"concrete_test":"Train ADE-MI on a synthetic two-view dataset with Y = X1 XOR X2, where X1 and X2 are independent fair bits, using the paper's Gaussian encoder, quantization, and losses; compare against a jointly trained encoder with the same bit budget. If ADE-MI accuracy is near chance while joint training is near perfect, the decomposition is not a valid surrogate for P1. As a complementary check on the real data, jointly train the encoders and server on Widar3.0 with gradient sharing and compare accuracy at identical d_k and n_k; a substantial accuracy gap would confirm that the per-device surrogate is lossy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.C.1 claims that solving P1.1 is 'equivalent to minimizing an upper bound of P1' because H(Y|Z1,...,ZK) ≤ min_k H(Y|Zk). This conflates minimizing an upper bound with solving the original problem. The inequality is always true, but it does not imply that encoders minimizing individual conditional entropies are optimal, or even near-optimal, for the joint inference objective. The gap is not benign: with two views that are each marginally independent of Y but jointly determine Y (e.g., Y = X1 XOR X2 with independent uniform bits X1, X2), each P1.1 encoder discards all task-relevant information, and the server receives useless features. The factorization in Eq. (10) would rule out this failure by assuming X1, ..., XK are conditionally independent given Y, but this assumption is not stated in P1 and is implausible for WiFi devices observing the same target in a shared environment. Thus ADE-MI is a heuristic upper-bound minimization, not the claimed solution of P1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADE-MI, a two-stage framework for multi-view WiFi sensing and edge inference. The authors formulate an information-theoretic problem P1 that minimizes the conditional entropy H(Y|Z_1,...,Z_K) of a gesture label Y given compressed features Z_k from K devices, subject to per-device mutual information constraints I(X_k;Z_k) <= C_k. They then decompose P1 into per-device encoding subproblems P1.1 and a server-side multi-view inference subproblem P1.2. The device encoder uses a trainable feature extractor followed by uniform quantization, with the latent dimension d_k computed from channel capacity via Eq. (8). Experiments on the Widar3.0 dataset report 92% gesture recognition accuracy and a large latency reduction relative to raw data transmission.","tokens_in":9797,"tokens_out":4488,"duration_ms":49598,"significance":"The problem is relevant to task-oriented communication and integrated sensing and communication, and the paper connects an information-theoretic formulation with a practical WiFi sensing system. The experimental evaluation on a real dataset is a strength, as are the reported gains in upload latency. However, the central theoretical claim that ADE-MI solves P1 is not established: the reduction from the joint problem to per-device problems is a heuristic upper-bound minimization, the factorization in Eq. (10) relies on an unstated and likely violated conditional-independence assumption, and the capacity constraint is enforced by construction rather than by evaluating I(X_k;Z_k). If these gaps can be addressed or the claims appropriately weakened, the framework may still be a useful engineering contribution, but the manuscript as written overstates its theoretical contribution.","major_comments":[{"comment":"The statement that 'P1.1 is equivalent to minimizing an upper bound of P1' is not an equivalence. The inequality H(Y|Z_1,...,Z_K) <= min_k H(Y|Z_k) is always true, but minimizing each per-device conditional entropy only minimizes an upper bound of the joint objective; it does not imply optimality, or even near-optimality, for P1. A concrete failure mode is the two-view case with independent uniform bits X_1, X_2 and Y = X_1 XOR X_2: each view is individually uninformative about Y, so the per-device encoders can discard all task-relevant information while still minimizing H(Y|Z_k), leaving the server with useless features. The factorization in Eq. (10) would rule out this example by assuming conditional independence of X_1,...,X_K given Y, but that assumption is not stated in P1 and is not verified for WiFi devices observing the same target in a shared environment. This gap is load-bearing because it is the justification for the two-stage decomposition.","section":"Section III.C.1, Eq. (12)"},{"comment":"The channel capacity constraint in Eq. (5) is I(X_k;Z_k) <= C_k, where C_k is defined in Eq. (4) as a rate in bits per second. However, Eq. (8) sets d_k = floor(C_k/(n_k L)) and the paper claims the constraint is 'satisfied by adapting the dimensionality of z_k.' This enforces a bound on the number of bits per encoded vector, not on the mutual information I(X_k;Z_k). Even if a quantized d_k-dimensional vector carries at most d_k n_k bits, the relationship between that bit count and C_k is unclear because C_k is a rate and L is described as a number of samples per transmission; the units in Eq. (8) do not match unless an additional time-window normalization is supplied. Moreover, the Lagrange multiplier beta_k introduced in Eq. (7) never appears in the device loss of Eq. (17) or in Algorithm 1, so the stated information-bottleneck trade-off is not actually tuned. The capacity-aware mechanism therefore does not implement the constraint in P1 as formulated.","section":"Section III.B, Eq. (8) and constraint Eq. (5)"},{"comment":"The factorization p(y,z_1,...,z_K) = p(y) ∫ ∏_k p(X_k|y) p_θk(z_k|X_k) dX_1...dX_K assumes that the observations X_1,...,X_K are conditionally independent given the label y. This assumption is load-bearing for the Monte Carlo approximation in Eq. (11) and for the per-device training of P1.1, yet it is not stated as an assumption in the problem formulation and is not validated for WiFi sensing, where multiple devices observe the same target and share the same propagation environment. Correlations among the views beyond the label are likely and can break the surrogate objective. The authors should either justify the assumption for their setting, measure the degree of conditional dependence, or compare against a jointly trained multi-view encoder to assess the practical impact of the approximation.","section":"Section III.C, Eq. (10)"},{"comment":"The experimental section does not directly measure any of the information-theoretic quantities in the problem formulation, such as I(X_k;Z_k), H(Y|Z_1,...,Z_K), or the gap between the per-device bound and the joint objective, so the empirical results do not validate the claimed solution of P1. In addition, Table I reports ADE-MI latency that is constant at 1.4×10^-5 s across SNR values while d_k varies from 6 to 24; the paper should explain why the latency is independent of the encoded dimension and clarify whether this latency includes transmission time or only encoding/inference time. This is needed to support the headline latency-reduction claim.","section":"Section IV, Table I and Fig. 4"}],"minor_comments":[{"comment":"The phrase 'two-stage optimization two-stage optimization' is duplicated in the abstract and in the introduction; it should be corrected.","section":"Abstract and Section I"},{"comment":"The abstract reports 'over 10^4-fold reduction in latency' while the body text reports '104-fold'; the notation should be made consistent.","section":"Abstract"},{"comment":"In Eq. (1), the quantities α_{k,l}, φ_{k,l}, and f^D_{k,l} are said to be in R^d, but the dimension d is never defined; this should be clarified.","section":"Section II.A, Eq. (1)"},{"comment":"The model setting says 'the number of samples over time S = T/Δt' and later uses L both as batch size in Algorithm 1 and as number of samples uploaded per transmission in Eq. (8); the same symbol is used with different roles and should be disambiguated.","section":"Section IV.B"},{"comment":"The sentence describing the first baseline contains a grammatical error: 'single-view sensing utilizing DFS data from only one device without any encoding or compression at the device side and directly transmits it to server' should be rewritten for clarity.","section":"Section IV.A"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable engineering study with a real-dataset evaluation, but the theoretical claims in Sections III.B and III.C are overstated and the decomposition argument is not sound as stated. The issues are not purely cosmetic: the main algorithmic design is justified by the invalid equivalence between P1 and P1.1 and by an unverified conditional-independence assumption. However, these are fixable in principle by reframing the contribution as a heuristic and validating the approximation empirically. The fit with the journal's scope in cs.IT would be strengthened if the mutual information quantities in the formulation were actually measured or bounded in the evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things before you read this: the practical system is plausible and the latency numbers on Widar3.0 are probably real, but the information-theoretic framing does not hold up. The paper claims to solve constrained optimization P1 exactly, but the two-stage decomposition solves a heuristic surrogate.\n\nWhat is new is the specific application: channel-capacity-aware dimension selection for multi-view WiFi sensing, where each device picks d_k = floor(C_k/(n_k L)) based on its link capacity. That is a legitimate, if incremental, idea. The empirical latency reduction (10^4x) versus raw transmission is sensible because you are transmitting up to 24 floats rather than full spectrograms. The scheme as described is reproducible.\n\nThe soft spots are load-bearing. Section III.C.1 says P1.1 is \"equivalent to minimizing an upper bound of P1\" because H(Y|Z1..ZK) <= min_k H(Y|Zk). That inequality is true, but minimizing the upper bound is not the same as minimizing the joint conditional entropy. Two views that are each independent of Y but jointly determine it (e.g., Y = X1 XOR X2) break the surrogate completely. The paper tries to rule this out with the factorization in (10), which assumes X1..XK are conditionally independent given y. That assumption is never stated as a hypothesis, and it is implausible for WiFi devices observing the same target in the same room. If the views are correlated through the environment beyond the label, the per-device encoders have no reason to preserve complementary information.\n\nThere is also a conceptual issue with the capacity constraint. Equation (5) says I(Xk;Zk) <= Ck, but Ck is a rate in bits/sec. Mutual information is not a bit rate, and the system does not compute or enforce mutual information. What it actually enforces is a bit budget: d_k * n_k / L bits per transmission. That is a legitimate rate constraint, but it should be stated as such. Calling it a mutual information constraint is a category error.\n\nThe experiments compare against raw-transmission baselines only. There are no matched-bitrate compression baselines (e.g., PCA features, JPEG-style compression, or a standard autoencoder at the same dimensionality). So the claim that ADE-MI \"outperforms\" is not fully established. The communication-interval result (Figure 3) mostly shows a sampling-rate effect; more packets per second gives richer temporal data, which any compression method would also benefit from.\n\nWho is this for? People working on task-oriented communication and edge inference might find the application fresh. But the theory needs serious rework before the paper makes a strong claim. I would not cite the theoretical result as stated, but the empirical recipe might be useful.\n\nRecommendation: yes, send to peer review, with a referee who knows information theory. It deserves scrutiny, not a desk rejection, and the authors deserve a chance to fix the framing.","headline":"Real but incremental empirical result; the theory overclaims—the two-stage decomposition is a heuristic, not a solution to P1.","tokens_in":10319,"tokens_out":2726,"would_cite":false,"duration_ms":29283,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","94A17","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Channel-capacity-aware distributed encoding lets multi-view WiFi sensing reach 92% gesture accuracy with over a 10^4-fold cut in upload latency.","keywords":["integrated sensing and communication","task-oriented communication","WiFi sensing","multi-view edge inference","information bottleneck","channel capacity","distributed encoding","gesture recognition"],"falsifier":"Estimate the conditional mutual information $I(Z_i;Z_j|Y)$ on the Widar3.0 data after training ADE-MI. If it is clearly above zero, the factorization in Eq. (10) fails and the claim that P1.1 minimizes an upper bound of P1 does not hold for those views; alternatively, compare the achieved joint entropy $H(Y|Z_1,\\dots,Z_K)$ with $\\min_k H(Y|Z_k)$ to see whether independent per-device training sacrifices the complementary information that multi-view fusion is supposed to provide.","tokens_in":9333,"feed_emoji":"📡","tokens_out":7530,"duration_ms":64216,"temperature":0.7,"pith_summary":"The paper sets out to establish that multi-view WiFi sensing with edge inference can be made communication-efficient without giving up accuracy, by deciding what each device transmits based on the channel capacity available to it. It formulates the task as minimizing the conditional entropy of the gesture label given all devices' encoded features, subject to a per-device mutual information constraint, and decomposes that problem into independent per-device information bottleneck problems followed by a server-side multi-view classifier. The claim is that this two-stage scheme, ADE-MI, reaches 92% recognition accuracy on the Widar3.0 six-gesture benchmark while reducing upload latency by more than $10^4$-fold relative to raw data transmission. A fair reader would care because the scheme suggests a practical way to align communication and sensing in ISAC systems: constrained wireless links set how many bits each sensor sends, and the encoding keeps only the bits that are relevant to the inference task.","feed_headline":"WiFi sensing hits 92% accuracy at 10,000x lower latency","feed_subtitle":"Channel-aware on-device compression plus one-shot server fusion cuts upload cost without losing multi-view accuracy","key_machinery":"The load-bearing object is the two-stage decomposition of P1 into P1.1 and P1.2. Stage one is a per-device information bottleneck: each encoder $\\theta_k$ minimizes $H(Y|Z_k)$ subject to $I(X_k;Z_k)\\le C_k$, with the latent sampled via the reparameterization trick and quantized uniformly, and the constraint enforced by setting the latent dimension to $\\lfloor C_k/(n_k L)\\rfloor$. Stage two trains a server-side model $p_\\psi(y|z_1,\\dots,z_K)$ on the one-shot uploaded encodings. The decomposition is carried by the upper-bound inequality $H(Y|Z_1,\\dots,Z_K)\\le \\min_k H(Y|Z_k)$ together with Monte Carlo approximations of both stage objectives; these pieces convert a joint, communication-heavy training problem into two independent training problems.","core_discovery":"The central discovery is that the constrained information-theoretic problem P1, minimize $H(Y|Z_1,\\dots,Z_K)$ subject to $I(X_k;Z_k)\\le C_k$, can be solved approximately by two stages that avoid joint training across devices and server. The first stage, adaptive distributed encoding, lets each device minimize its own conditional entropy $H(Y|Z_k)$ under the same constraint, using an encoder with a reparameterized Gaussian latent and uniform quantization; the second stage, multi-view inference, trains the server classifier on the one-shot encoded vectors from all devices. The paper argues this is valid because, under the Markov chain $Y\\leftrightarrow X_k\\leftrightarrow Z_k$, the joint conditional entropy satisfies $H(Y|Z_1,\\dots,Z_K)\\le \\min_k H(Y|Z_k)$, so each local problem minimizes an upper bound of the original objective. The channel capacity directly determines the encoded feature dimension through $d_k=\\lfloor C_k/(n_k L)\\rfloor$, where $n_k$ is the bits per element and $L$ the samples per transmission. On the Widar3.0 dataset this two-stage scheme is reported to achieve 92% accuracy with over a $10^4$-fold latency reduction compared with raw data upload.","pith_inferences":["Beyond the paper's claims: the inequality $H(Y|Z_1,\\dots,Z_K)\\le \\min_k H(Y|Z_k)$ leaves room for multi-view gain, so when views carry complementary information a jointly trained encoder set should outperform the independent decomposition; the paper does not quantify this gap.","Beyond the paper's claims: the conditional-independence factorization in Eq. (10) can be checked empirically by estimating $I(Z_i;Z_j|Y)$ on the encoded features; a clearly positive value would mean the per-device objective is not a faithful surrogate for the joint objective.","Beyond the paper's claims: the same capacity-aware encoding recipe should transfer to other multi-sensor edge tasks, such as mmWave radar or microphone-array sensing, whenever observations are high-dimensional and share a single label.","Beyond the paper's claims: the reported accuracy gain from shorter communication intervals may partly reflect better temporal sampling of gestures rather than communication rate itself; an ablation that fixes the number of CSI frames per gesture would separate the two effects."],"forward_implications":["Multi-view edge inference can be deployed with a single upload per device after local training, so the training phase no longer requires repeated server-device exchange of encoded vectors and parameter updates.","Channel conditions directly set the encoded feature dimension, so devices can adapt to varying SNR and bandwidth by changing $d_k$ while the server-side model remains fixed.","More frequent communication between the WiFi access point and edge devices improves sensing accuracy, meaning communication rate and sensing quality can rise together rather than only trade off.","A roughly $10^4$-fold reduction in upload latency makes real-time gesture recognition feasible under tight bandwidth and latency constraints where raw data transmission would be too slow.","Because local encoders are trained independently, the number of participating devices can change without retraining the per-device encoders, and the server can fuse whatever views are present."],"supporting_citations":[{"why":"supplies the Widar3.0 six-gesture WiFi dataset and its CSI preprocessing, which all experiments use.","marker":"[8]"},{"why":"is the task-oriented multidevice cooperative edge inference scheme that this paper extends by adding channel-capacity awareness.","marker":"[16]"},{"why":"frames task-oriented communication and edge inference as the motivation for transmitting only task-relevant features.","marker":"[15]"},{"why":"establishes the ISAC paradigm and the spectrum-sharing context behind the claimed communication-sensing synergy.","marker":"[1]"},{"why":"provides the CSI phase-error removal method used to preprocess raw CSI in Eq. (1).","marker":"[17]"},{"why":"justifies the STFT-based Doppler spectrogram representation that serves as the encoder input.","marker":"[19]"}],"fun_headline_variants":["92% WiFi sensing accuracy with 10,000x less upload","WiFi sensing hits 92% accuracy while uploading 10,000x less data","Channel-aware WiFi sensing: 92% accuracy, 10,000x less upload","Edge inference from WiFi: 92% accuracy with 10,000x less data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each device's observations are conditionally independent given the gesture label, so the joint inference objective can be replaced by independent per-device encoders; if the devices see correlated views of the same scene, the decomposition no longer solves the original problem.","fun_headline_variants_meta":{"raw":{"variants":["92% WiFi sensing accuracy with 10,000x less upload","WiFi sensing hits 92% accuracy while uploading 10,000x less data","Channel-aware WiFi sensing: 92% accuracy, 10,000x less upload","Edge inference from WiFi: 92% accuracy with 10,000x less data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001162,"raw_usage":{"total_tokens":4867,"prompt_tokens":1061,"completion_tokens":3806,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":677,"completion_tokens_details":{"reasoning_tokens":3719}},"tokens_in":677,"tokens_out":3806,"duration_ms":27724,"temperature":1.0,"reasoning_tokens":3719,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:23:17.756802+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate the conditional mutual information $I(Z_i;Z_j|Y)$ on the Widar3.0 data after training ADE-MI. If it is clearly above zero, the factorization in Eq. (10) fails and the claim that P1.1 minimizes an upper bound of P1 does not hold for those views; alternatively, compare the achieved joint entropy $H(Y|Z_1,\\dots,Z_K)$ with $\\min_k H(Y|Z_k)$ to see whether independent per-device training sacrifices the complementary information that multi-view fusion is supposed to provide.","supporting_citations":[{"cited_title":"WiFi CSI-based de vice- free sensing: from Fresnel zone model to CSI-ratio model,","cited_arxiv_id":null,"evidence_quote":"provides the CSI phase-error removal method used to preprocess raw CSI in Eq. (1)."},{"cited_title":"Towa rd an Intelligent Edge: Wireless Communication Meets Machine Le arning,","cited_arxiv_id":null,"evidence_quote":"frames task-oriented communication and edge inference as the motivation for transmitting only task-relevant features."},{"cited_title":"Widar3.0: zero-effort cross-domain gesture recognition with Wi-Fi,","cited_arxiv_id":null,"evidence_quote":"supplies the Widar3.0 six-gesture WiFi dataset and its CSI preprocessing, which all experiments use."},{"cited_title":"Task-oriented communica tion for multidevice cooperative edge inference,","cited_arxiv_id":null,"evidence_quote":"is the task-oriented multidevice cooperative edge inference scheme that this paper extends by adding channel-capacity awareness."},{"cited_title":"WiFi sensing with channel st ate information: A survey,","cited_arxiv_id":null,"evidence_quote":"justifies the STFT-based Doppler spectrogram representation that serves as the encoder input."}],"review_version":1}