{"id":"49135aa8-febc-402b-b19c-b06b8323fa5d","arxiv_id":"2509.16577","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A jointly learned codebook and unrolled AMP decoder improves low-SNR digital over-the-air aggregation for federated edge learning, extending the reliable SNR range by about 7 dB.","lead":"This paper trains a neural decoder and a shared codebook together to recover federated learning updates transmitted simultaneously over the air in noisy, low-power settings. The learned system reportedly extends the usable signal-to-noise range by about 7 dB compared with the prior digital over-the-air baseline, which could make federated edge learning practical in IoT conditions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"7 dB claim rests on untested assumption that BS-local popularity prior (Eq. 8/16) matches device codeword usage; a mismatch could erase the low-SNR gain.","rationale":"The central claim is a 7 dB extension of viable SNR at equal uplink overhead, supported by Table 1. For that claim to be more than a tuned-simulation artifact, the decoder must be robust to the one quantity the method estimates from the BS's local data: the popularity distribution π. Eq. (16) makes π part of the prior for every codeword, and the EM updates for K_a and π feed back into the decoder. In the experiments, the BS and devices are generated from the same dataset pipeline, so π is calibrated by construction; the paper's only evidence for transferability is an unquantified statement in Sec. 1. In a real FEEL deployment the BS may not have data at all, and if it does, that data can be label-skewed relative to devices. Under mismatch, the decoder's prior mass is placed on wrong codewords, the K_a estimate is biased, and the low-SNR recovery advantage—the entire basis of the headline—could collapse to the AMP-DA baseline. The proposed check directly tests this by mismatching the BS dataset. This is the same weak spot the reader identified, and it justifies the existing CONDITIONAL verdict.","tokens_in":8284,"tokens_out":10140,"duration_ms":96545,"concrete_test":"Repeat the Table 1 experiment but construct the BS's local dataset from a non-representative label distribution (e.g., only 2 of 10 classes, or a disjoint label shard), while keeping the devices' data split unchanged. Record final test accuracy and K_a estimation error at SNR 0, 3, and 5 dB. If accuracy at 3 dB stays within ~2% of the matched-prior case (0.788), the concern is resolved; if it drops toward the AMP-DA baseline (0.135), the 7 dB claim is conditional on BS representativeness and the paper must include a mismatch study.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decoder's Bayesian denoiser is built around the Poisson prior λ_j = K_a π_j (Eq. 16), where π is estimated by k-means clustering the BS's own local update (Sec. 3.1, Eq. 8). The central low-SNR gain in Table 1 (proposed 0.690/0.788 at 0/3 dB vs AMP-DA 0.139/0.135) depends on this prior being calibrated to actual device codeword usage. In the experiments, the BS and devices are generated from the same data pipeline, so π is calibrated by construction; the only support for transferability is a contribution bullet in Sec. 1 claiming BS local training on small non-IID subsets captures device-level dynamics, with no experimental quantification. If the BS's local data is label-skewed relative to devices, π is miscalibrated, the spike-and-slab prior puts mass on the wrong codewords, and the EM updates for K_a and π inherit the bias. The low-SNR recovery advantage—the entire basis of the headline—could then collapse toward the AMP-DA baseline. Thus the 7 dB extension is not shown to survive the heterogeneity the paper claims to address.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a learned digital over-the-air (OTA) aggregation scheme for federated edge learning (FEEL). The method, AMP-DA-Net, is an unrolled approximate message passing (AMP) decoder with per-layer damping, residual scaling, temperature-controlled Bayesian denoising, CNN refinement, and EM-based updates of active-device count, popularity, and noise variance. It is trained jointly with a URA codebook parameterized as a product of two matrices (C = DW). The central claim is a roughly 7 dB extension of the reliable SNR range over the state-of-the-art AMP-DA baseline at the same uplink overhead, together with generalization across models, activity levels, and heterogeneous data. Experiments on a non-IID FEEL setup with a ResNet global model show large improvements at low SNR (Table 1), and one additional experiment transfers the trained system to a simpler CNN at 5 dB SNR.","tokens_in":8606,"tokens_out":5290,"duration_ms":49816,"significance":"If the reported gains hold, this is a useful contribution to digital OTA aggregation: it demonstrates that end-to-end learning of the URA codebook and the AMP decoder can extend reliable operation into low-SNR IoT regimes without increasing uplink overhead. Strengths include the well-structured decoder design, the two-matrix codebook parameterization, and the release of code, which supports reproducibility. The paper does not fit constants to a target result; test accuracies are measured on held-out rounds after training on separate data. However, the central 7 dB claim currently rests on a critical assumption about the representativeness of the BS-local popularity prior and on a single simulation scenario, so the demonstrated scope is narrower than the abstract and conclusions suggest.","major_comments":[{"comment":"The low-SNR advantage in Table 1 (proposed 0.690/0.788 at 0/3 dB vs AMP-DA 0.139/0.135) depends on the Poisson prior lambda_j = K_a pi_j in Eq. (16), where pi is estimated from the BS's own local update (Eq. (8)). If the BS-local data is label-skewed relative to the device population, lambda_j is miscalibrated and the EM updates for K_a and pi inherit that bias. The paper claims generalization to heterogeneous data, but the only evidence is a contribution bullet in Sec. 1 without a quantitative experiment. Please either add a controlled mismatch study (e.g., varying label skew between BS-local and device data and re-running Table 1) or qualify the claims. As written, the 7 dB claim is not shown to survive the heterogeneity the paper advertises.","section":"Secs. 3.1 and 3.2, Eq. (16)"},{"comment":"The central 'more than 7 dB' claim is supported by a single simulation scenario: one global model (ResNet), one non-IID split, one activity range (K_a uniform in [7,13]), and one SNR grid. No standard deviations or seed counts are reported, and the AMP-DA baseline's hyperparameters (damping, number of iterations, prior settings) are not specified. The reader cannot assess whether the improvement is statistically robust or an artifact of baseline de-tuning. Please report multiple independent runs with error bars, specify the baseline tuning, and ideally include ablations over activity levels and data heterogeneity.","section":"Section 4, Table 1"},{"comment":"The abstract claims the method 'generalises across models, activity levels, and heterogeneous data', and the conclusions repeat this. The only direct test is one simpler CNN at 5 dB SNR; K_a is varied only in the sense that training and evaluation draw from the same [7,13] range, and no systematic variation of IID/non-IID composition is presented. This is a mismatch between claim and evidence. Either add a systematic generalization study (different architectures, heterogeneity levels, and activity-count distributions) or restrict the generalization statements to the tested regime.","section":"Section 4, generalization paragraph"}],"minor_comments":[{"comment":"The notation paragraph says '⊙ and circ denote element-wise multiplication and exponentiation, respectively,' but the exponentiation symbol is garbled. Please fix the rendering.","section":"Notation, Sec. 1"},{"comment":"The definitions of var1 and var2 are ambiguous. Specify dimensions and transposes explicitly (e.g., 'C⊙2' should likely be the squared matrix or its transpose).","section":"Eq. (15)"},{"comment":"The caption says 'Codebook setup evaluation accuracy' but reports 0.949 for the proposed setup at 5 dB, while Table 1 reports 0.794 at the same SNR for the same system. If Table 2 reports codeword-recovery accuracy rather than final test accuracy, say so explicitly; otherwise reconcile the two numbers.","section":"Table 2"},{"comment":"Reference [16] appears to have an incomplete author list ('... and Safi'). Please complete it.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the code release is a clear strength. The main risk is the popularity-prior assumption: the 7 dB claim may not survive realistic mismatch between the BS-local dataset and the device population. I would encourage a controlled mismatch experiment and tighter generalization claims. Also please check the Table 2/Table 1 discrepancy; it may indicate a metric mismatch rather than an error, but the captions need to be clearer."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, honest systems paper that combines two known ideas — unrolled AMP decoding and a learned URA codebook — into a new pipeline for digital OTA federated learning. It is the first, to my reading, to jointly train the codebook and decoder end-to-end, and the public code plus the codebook ablations give me confidence the authors built what they claim. The 7 dB low-SNR extension over MD-AirComp/AMP-DA is real in the tested setup, but it is not yet shown to be robust.\n\nWhat is good. The decoder is well-constructed: per-layer damping, residual scaling, temperature-controlled Bayesian denoiser, CNN refinement, and EM updates for K_a and pi are standard pieces assembled sensibly. The two-matrix codebook parameterization (SimVQ-style) with data-driven initialization is a reasonable way to avoid dead codewords. The ablations in Table 2 are useful and show popularity ordering and learning both matter. Table 1 is clear: at 0/3 dB the proposed method holds accuracy near the perfect-aggregation level while the baseline collapses.\n\nWhere it is soft. The headline depends on a single simulation scenario (ResNet, 80/20 non-IID, one SNR grid), with no error bars or repeated seeds. The AMP-DA baseline tuning is unspecified; a fair comparison needs the baseline's hyperparameters stated. Generalization is tested on one simpler CNN at 5 dB only. These are fixable with more experiments, but as it stands the 'generalises across models, activity levels, and heterogeneous data' claim is broader than the evidence.\n\nThe more interesting worry is the popularity prior. The Bayesian denoiser assumes lambda_j = K_a pi_j with pi estimated by k-means on the BS's own local update (Eqs. 8, 16). In the experiments, BS and devices come from the same data-generation pipeline, so the prior is calibrated by construction. The paper gives only a contribution bullet claiming BS local training on small non-IID subsets captures device dynamics; there is no experiment with a deliberately skewed BS distribution. If the BS's local data is label-skewed relative to devices, the spike-and-slab prior points at the wrong codewords and the low-SNR margin could shrink toward the baseline. This is not a fatal flaw — the architecture could still help via the learned codebook and CNN refinement — but it is a load-bearing assumption and it is unverified.\n\nBottom line: this is a genuine contribution to the wireless FL subfield, not a paradigm shift. It deserves peer review, but I would ask for seeds/error bars, baseline tuning, and a popularity-mismatch experiment before I would trust the 7 dB claim outside the exact setup tested. I would bring it to a reading group interested in learned physical-layer schemes.","headline":"Worth refereeing: a real architectural novelty in learned digital OTA aggregation, but the 7 dB claim needs more than one simulated pipeline and the BS-side popularity prior is doing more work than the paper verifies.","tokens_in":9116,"tokens_out":2572,"would_cite":true,"duration_ms":23897,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Learned digital over-the-air computation keeps federated learning accurate at 0 dB SNR.","keywords":["federated edge learning","over-the-air computation","unsourced random access","approximate message passing","learned decoder","codebook learning","compressed sensing","low-SNR aggregation"],"falsifier":"Deploy the same pipeline with a deliberate mismatch: give the base station a small IID dataset while devices draw from label-sharded non-IID shards, then measure final test accuracy at 0–3 dB SNR. If accuracy falls toward the AMP-DA baseline (about 0.14) rather than the reported 0.69–0.79, the representativeness of the local dataset is the load-bearing assumption.","tokens_in":8156,"feed_emoji":"📶","tokens_out":5318,"duration_ms":44344,"temperature":0.7,"pith_summary":"This paper proposes replacing the fixed encoder–decoder pair in digital over-the-air (OTA) computation with a learned one: an unsourced random access (URA) codebook and an unrolled approximate message passing decoder (AMP-DA-Net) trained jointly end to end. The authors show that at the same uplink overhead as the state-of-the-art AMP-DA baseline, the learned system extends reliable aggregation by roughly 7 dB, achieving 0.690 test accuracy at 0 dB SNR where the baseline collapses to 0.139. The design keeps the digital robustness that makes OTA aggregation practical while pushing usable operation into the low-SNR regime typical of IoT devices. This matters because OTA aggregation is a leading approach to make federated edge learning communication-efficient; if the reported margin holds, learned digital OTA could make it viable in noise regimes where current digital schemes fail to converge.","feed_headline":"Learned codebook and decoder gain 7 dB for low-SNR federated learning","feed_subtitle":"At 0 dB SNR the learned system reaches 0.690 test accuracy; the state-of-the-art baseline collapses to 0.139.","key_machinery":"AMP-DA-Net, an unrolled generalized approximate message passing decoder with a learned Bayesian denoiser (spike at zero, Poisson slab, temperature sharpening), per-layer learnable damping and residual scaling, a CNN refinement, and alternating EM updates of K_a, π, and σ². The URA codebook is jointly trained via the factorized parameterization C = DW with row normalization and popularity ordering; ordering makes the codeword usage predictable and supplies the prior for the denoiser. The decoder alternates an output block (measurement-domain residual update with learnable Onsager correction) and an input block (pseudo-channel denoising in codeword domain), then post-processes via non-negativi","core_discovery":"The paper's central claim is that the OTA decoding bottleneck is not the compressed-sensing formulation itself but the fixed codebook and the hand-tuned denoiser: both can be replaced by learned components whose parameters are optimized offline and then frozen at deployment. The URA codebook is factorized as C = DW (a data-driven base matrix times a learned shear/rotation), row-normalized, and trained jointly with AMP-DA-Net, a ten-layer unrolled AMP decoder that adds per-layer damping, residual scaling, a temperature-controlled spike-and-slab (spike at zero, Poisson slab) Bayesian denoiser, a CNN refinement, and EM updates for the number of active devices, the codeword popularity distributi","pith_inferences":["Because the codebook is learned jointly with the decoder, the same end-to-end training recipe could benefit unsourced random access beyond federated learning, wherever representative activity patterns are available for offline training.","The popularity-ordering trick suggests a general principle for non-stationary compressed sensing: steering the input distribution toward a predictable prior can be as valuable as improving the decoder's representation.","The temperature-controlled Bayesian denoiser can be read as a learned annealing schedule; adapting temperature online per round could extend the operating SNR range further without retraining.","The paper validates generalization to different models and data splits but not to mismatched popularity distributions; a direct test would reveal whether the base station's local dataset is a sufficient statistical proxy for device updates."],"forward_implications":["The same uplink overhead as the state-of-the-art baseline supports reliable aggregation down to roughly 0 dB SNR, extending the usable range by about 7 dB.","At 0 dB SNR, global model test accuracy is 0.690 with the proposed method versus 0.139 with AMP-DA; at 3 dB, 0.788 versus 0.135, approaching the perfect-aggregation ceiling of 0.804.","Jointly learned codebooks outperform fixed constructions (0.949 vs 0.696 test accuracy at 5 dB SNR for the best fixed baseline), and data-driven initialization beats Gaussian or Bernoulli initializations.","The method generalizes across models (trained on ResNet updates, deployed with a simpler CNN at 5 dB without noticeable convergence loss), varying numbers of active devices, and highly non-IID data.","Active-device estimation remains within the ±0.5 tolerance needed to correctly normalize the aggregated update, avoiding instability from underestimation and slowdown from overestimation."],"fun_headline_variants":["Learned OTA codebook and decoder gain 7 dB SNR","7 dB SNR extension from learned digital OTA for FEEL","Unrolled AMP plus learned codebook: 7 dB low-SNR gain","Federated edge learning over the air gets 7 dB boost"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The decoder's Bayesian prior is built from the popularity of codewords in the base station's own local update, and the whole design assumes that this local distribution matches what the devices actually transmit; when it does not, the Poisson prior and the EM updates that depend on it become miscalibrated.","fun_headline_variants_meta":{"raw":{"variants":["Learned OTA codebook and decoder gain 7 dB SNR","7 dB SNR extension from learned digital OTA for FEEL","Unrolled AMP plus learned codebook: 7 dB low-SNR gain","Federated edge learning over the air gets 7 dB boost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000266,"raw_usage":{"total_tokens":1460,"prompt_tokens":770,"completion_tokens":690,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":614}},"tokens_in":514,"tokens_out":690,"duration_ms":6711,"temperature":1.0,"reasoning_tokens":614,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:03:51.069215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy the same pipeline with a deliberate mismatch: give the base station a small IID dataset while devices draw from label-sharded non-IID shards, then measure final test accuracy at 0–3 dB SNR. If accuracy falls toward the AMP-DA baseline (about 0.14) rather than the reported 0.69–0.79, the representativeness of the local dataset is the load-bearing assumption.","supporting_citations":[],"review_version":1}