{"id":"58d3bfae-9c19-468d-9a44-d80d955860bd","arxiv_id":"2411.12541","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Depthwise separable U-Net variants for blind co-channel interference cancellation achieve higher MSE scores than WaveNet and ConvTasNet at a fraction of the computational cost.","lead":"Two compact U-Net neural networks are designed to remove unknown interfering wireless signals from a desired QPSK signal, and the models are trimmed with depthwise separable convolutions and 8-bit quantization. On the ICASSP 2024 radio-signal dataset, the authors report higher signal-reconstruction scores than WaveNet and ConvTasNet while using far fewer multiply-accumulate operations, which matters for running interference cancellation on resource-limited edge devices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is conditional on the MSE score being a valid proxy for BER; without a BER/SINR evaluation, higher MSE scores do not establish better blind CCI for edge devices.","rationale":"The reader's conditional verdict already identifies the same load-bearing gap: the practical communication benefit rests on treating the MSE score as a proxy for BER, and the paper explicitly defers BER analysis. My read of the paper confirms this is the weakest point in the central argument. The MAC/parameter reductions reported in Table I are arithmetically consistent, and the depthwise-separable deltas (58.66% and 61.10%) match the stated MAC counts, so the efficiency side of the claim is credible. The performance side, however, is only a regression metric. Since QPSK decision regions are robust, two models with MSE scores of 27 and 44 dB could have nearly identical BER, or the higher-MSE model could decode worse if residual errors are correlated with symbol timing. The absence of a BER/SINR curve means the conclusion that these models are strong candidates for interference cancellation in communication systems is not yet established, exactly as the reader stated. I also note the pruning results in Table II are surprising — a 0.01 pruning ratio appears to drop every model's MSE score by roughly 12-15 dB relative to Table I — and would be worth checking, but the MSE-to-BER gap is the more load-bearing concern for the paper's central claim. No change to the reader's conditional verdict is needed.","tokens_in":16003,"tokens_out":6366,"duration_ms":68963,"concrete_test":"Run the 50 reserved test superframes through M1, M1(Dw), M2, M2(Dw), WaveNet, and ConvTasNet at SINR values of -30, -20, -10, -5, and 0 dB. Pass each estimated signal through the same RRC matched filter and symbol sampler used to generate Eq. (1), with F=16 and tau0=8, and compute bit-error rate versus SINR for QPSK. If M1/M2 do not achieve lower BER than the baselines at the same SINR, or if all models already achieve zero BER, the MSE-score advantage does not substantiate the claimed communication benefit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim — that M1/M2 deliver superior blind CCI with far fewer MACs — is a claim about communication performance, yet the only performance evidence is the truncated negative-log-MSE score defined in Section II-B. The paper explicitly defers BER analysis in Section II-B ('a BER analysis that includes bit decoding is left open for future research') and again in Section V-C ('extending the evaluation to include other relevant metrics, such as BER analysis'). The reported symbol rates are also raw model inference throughput, not decoded symbols, so '800,000 symbols per second' overstates communication throughput. A higher MSE score has not been shown to lower BER for QPSK after matched filtering: residual errors below the decision threshold can leave BER unchanged, while correlated or timing-dependent residuals can make similar-MSE reconstructions decode very differently. Thus the higher MSE scores in Fig. 4 and Table I support an improvement in waveform regression, but not the paper's practical conclusion about interference cancellation and edge-device communication. The central claim therefore holds only conditionally on an untested MSE-to-BER link.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two U-Net-based models, M1 (with an LSTM bottleneck) and M2 (fully convolutional), for blind co-channel interference cancellation on the ICASSP 2024 RF signal separation challenge dataset. It compares these models against WaveNet and ConvTasNet using a truncated negative-log MSE score, MAC counts, parameter counts, and CPU/GPU inference throughput, and it applies depthwise separable convolutions, quantization-aware training, and pruning to reduce computational cost. The headline results are that M1 and M2 achieve higher MSE scores than the baselines with far fewer MACs, and that depthwise separable variants reduce MACs by 58.66% (M1) and 61.10% (M2) with 0.72% degradation and 0.63% improvement in MSE score, respectively. The paper also reports batch-size scaling on GPUs and pruning sensitivity.","tokens_in":16246,"tokens_out":5435,"duration_ms":51404,"significance":"If the comparisons are reliable, the paper offers a genuinely practical efficiency/performance trade-off for learned blind CCI, with the depthwise separable variants being the strongest contribution: the MAC reductions are arithmetically consistent, and M2(Dw) even improves the MSE score. The paper is honest about several limitations, explicitly deferring BER analysis and reporting steep pruning degradation. However, the practical communication-level claim is not yet established because the evaluation metric is a waveform-regression score, not a decoding metric, and the reported symbol rates are raw inference throughputs rather than decoded symbols. No code or seeded training details are provided, which limits independent verification. The central engineering result, if confirmed with communication metrics and repeated-seed statistics, would be a useful baseline for edge-device blind CCI.","major_comments":[{"comment":"The central claim that M1 and M2 provide superior blind CCI for edge-device communication is conditional on the MSE score being a valid proxy for bit-error rate, but the paper explicitly defers BER analysis to future work in Section II-B ('a BER analysis that includes bit decoding is left open for future research') and again in Section V-C. Since the MSE score is a truncated negative-log waveform error, a higher MSE score does not imply a lower BER for QPSK with RRC pulse shaping after matched filtering: residual errors below the decision threshold can leave BER unchanged, while correlated or timing-dependent residuals can change BER at the same MSE. I therefore regard the headline practical claim as unsupported until the authors evaluate BER/SINR after an actual receive filter and symbol decision on the challenge test set, or otherwise calibrate the MSE score against BER.","section":"Section II-B / Section V-C"},{"comment":"The 'symbol rate' metric is described in Section IV-A2 as 'the number of communication symbols that can be decoded when performing interference cancellation in one second,' and Section V-B repeats 'up to 800,000 symbols per second on GPUs.' Since the models output an estimated waveform and no bit decoding or demapping is performed, these figures are raw waveform-estimation throughputs, not decoded communication symbols. This overstates the end-to-end communication throughput. The authors should relabel the metric as symbols processed by the cancellation model and, if the intended claim is communication throughput, provide an end-to-end measurement that includes matched filtering and demapping.","section":"Section IV-A2 / Section V-B"},{"comment":"All MSE scores in Table I and Table II are single-run values with no confidence intervals or repeated-seed statistics. This matters because the key depthwise claims are small relative differences (0.72% degradation for M1, 0.63% improvement for M2), which may be within training noise. Moreover, Table II shows that 1% pruning already reduces the MSE score from 44.07 to 31.67 for M1 and from 42.92 to 29.62 for M2, a loss far larger than the text's 'general trend of degradation' suggests. The pruning section should report repeated-seed variability and explicitly reconcile the 1% results with the stated goal of evaluating compression.","section":"Section IV-B / Table I / Section IV-D / Table II"}],"minor_comments":[{"comment":"The component list says LeakyReLU, but the encoder/decoder code equations and Figure 2 consistently use ReLU; please clarify which activation is actually used.","section":"Section III-A1"},{"comment":"The text describes 'unstructured pruning' that removes individual weights, but the parenthetical says 'pruning by removing filters with the lowest l1-norm'; please decide whether the pruning is weight-level or filter-level and state it consistently.","section":"Section IV-D"},{"comment":"Please specify the input length and inference configuration for which MACs and parameter counts are reported, since M2 is described as input-size independent and the MAC numbers must correspond to a particular segment length.","section":"Table I"},{"comment":"The dataset description reserves 50 superframes for testing but does not state whether these are balanced across the four interference types; please specify the test-set composition.","section":"Section III-C"},{"comment":"There are several typos, including 'Neural Networkss' in Section I-A and 'disurbed' near the end of Section I; a proofreading pass is needed.","section":"Throughout"},{"comment":"The bubble sizes are said to be proportional to parameter counts, but there is no legend or numeric label; please add a legend or labels so the visual comparison is readable.","section":"Figures 4-6"}],"recommendation":"major_revision","confidential_remarks":"The paper is a straightforward empirical comparison on an accepted challenge dataset; the main risk is not circularity but missing validation of the proxy metric. If the authors add BER/SINR evaluation and repeated-seed statistics, the paper could become a solid systems contribution. I would not block on the pruning section, but it should be reframed as a negative result, and the current 1% pruning drop is more severe than the text acknowledges."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nQuick take: this is a solid engineering paper, not a breakthrough. The new bit is the specific combination of depthwise separable convolutions, quantization-aware training, and pruning applied to U-Net-style models for blind co-channel interference cancellation, with careful MAC/MSE trade-offs measured on the ICASSP 2024 data-driven challenge dataset. The two proposed models (M1 with LSTM bottleneck, M2 fully convolutional) beat WaveNet and ConvTasNet on MSE score while needing far fewer MACs. The paper is honest about trade-offs: depthwise convolutions slightly hurt M1 but slightly help M2, quantization causes a large MSE drop (from 44 to 33 for M1), and pruning degrades performance sharply even at 1%. That transparency earns credit.\n\nWhere it gets soft: the central claim that these models deliver better blind CCI for edge devices is conditional on the MSE score being a valid proxy for BER. The paper explicitly defers BER analysis to future work, and the stress-test note is right: higher MSE score does not guarantee better decoding after matched filtering. So the headline overreaches, but the paper does not hide the gap. Also, there are no error bars or repeated-seed results, no code, and the 'edge' measurements are on a server CPU/GPU, not actual edge hardware. The missing comparison to the authors' own earlier U-Net [9] is a minor gap, though it would help to see the full delta from that baseline.\n\nNone of this is fatal. The paper says what it did, reports numbers that are internally consistent, and clearly states what is not shown. That is a reasonable contribution for a venue-level experimental paper. I would cite the MSE/MAC trade-off table as a benchmark, but not the edge-device claim without a BER follow-up.\n\nRecommendation: send it out. It deserves a serious referee, and the main referee question should be the MSE-to-BER link. Also ask for error bars and, ideally, a check on actual edge hardware.\n\nBest,\n[Your name]","headline":"A competent engineering study of efficient U-Nets for blind CCI, with honest trade-off reporting, but the central edge-device claim rests on an untested MSE-to-BER link.","tokens_in":16762,"tokens_out":3037,"would_cite":true,"duration_ms":30063,"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":"Two compact U-Net models achieve higher MSE scores than WaveNet and ConvTasNet on blind co-channel interference cancellation while requiring far fewer MACs, and depthwise separable versions cut computation by roughly 60% with little MSE…","keywords":["blind co-channel interference cancellation","U-Net","depthwise separable convolution","quantization-aware training","edge devices","signal separation","LSTM","MSE score"],"falsifier":"Decode the reconstructed QPSK signals from M1, M2, and the baselines on the same test data and compare bit-error rates; if M1 or M2 do not achieve lower BER than WaveNet and ConvTasNet, the central performance advantage is not confirmed for real communication.","tokens_in":15825,"feed_emoji":"📡","tokens_out":4447,"duration_ms":42580,"temperature":0.7,"pith_summary":"This paper tries to show that blind co-channel interference cancellation, which removes unknown interfering signals from a desired transmission, can be done accurately enough by small neural networks that fit on edge devices. It proposes two U-Net variants: one shallow model with an LSTM bottleneck (M1) and one deeper fully convolutional model (M2). The paper reports that both achieve higher MSE scores than the WaveNet and ConvTasNet baselines while needing significantly fewer multiply-accumulate operations (MACs). It further shows that replacing standard convolutions with depthwise separable convolutions cuts MACs by 58.66% for M1 and 61.10% for M2, with only a 0.72% MSE degradation for M1 and a 0.63% MSE improvement for M2. The paper explicitly defers bit-error-rate analysis to future work, so the practical communication benefit rests on the untested assumption that the MSE score translates to better decoding.","feed_headline":"Blind interference cancellation gets 60% cheaper with slim U-Nets","feed_subtitle":"Depthwise separable U-Nets beat WaveNet and ConvTasNet in MSE while cutting MACs by up to 61% for edge deployment.","key_machinery":"The central mechanism is the depthwise separable convolution, which replaces each standard convolution with a per-channel depthwise convolution followed by a pointwise 1x1 convolution, decoupling spatial filtering from channel mixing to cut MACs and parameters. This is combined with quantization-aware training (QAT) that simulates 8-bit precision during fine-tuning of convolutional layers, reducing memory footprint while preserving most of the full-precision performance. The U-Net encoder-decoder structure with skip connections and GroupNorm provides the base architecture; M1 adds an LSTM bottleneck to capture temporal dependencies, while M2 is fully convolutional for input-size flexibility and better GPU batch scaling.","core_discovery":"The central claim is that two U-Net convolutional architectures, one with an LSTM at the bottleneck (M1) and one fully convolutional (M2), outperform the established WaveNet and ConvTasNet baselines on the ICASSP 2024 blind co-channel interference cancellation task, achieving MSE scores of 44.07 and 42.92 respectively versus 25.45 for WaveNet and 27.66 for ConvTasNet, while requiring substantially fewer MACs and parameters. The paper also establishes that depthwise separable convolutions reduce computational cost to about half (51.25M MACs for M1(Dw) and 51.23M for M2(Dw)) with negligible MSE change, and that quantization-aware training to 8-bit shrinks model size to under 2.1 MB at a larger MSE cost. These results are presented as evidence that optimized U-Nets can bring blind CCI cancellation to resource-constrained edge devices without sacrificing signal reconstruction quality.","pith_inferences":["A natural next test would be end-to-end BER evaluation with a real decoder on the ICASSP 2024 test data; if the higher MSE scores do not translate into lower bit-error rates, the practical communication benefit is unproven.","The efficiency measurements are taken on server CPU and GPU hardware; actual edge processors such as mobile CPUs or microcontrollers may not realize the full MAC reduction as latency or power savings, especially for depthwise convolutions that are not always well optimized.","Because the model is trained across all four interference types without using the type labels, a strong robustness test would be to hold out an entire interference type during training and measure blind generalization.","The significant MSE drop from quantization (e.g., 44.07 to 33.10 for M1) might still be acceptable if the absolute signal quality remains above the threshold needed for decoding; BER testing would settle whether the trade-off is worthwhile."],"forward_implications":["Deploying blind CCI cancellation on edge devices becomes feasible: the depthwise models are under 2.7 MB and require roughly 51M MACs, far below the billion-scale MAC counts of the baselines.","The fully convolutional M2 model scales to the largest batch sizes on a GPU, reaching up to 800,000 symbols per second, making it the better choice for high-throughput applications.","Depthwise separable convolutions can be applied to U-Net signal separation models with minimal MSE change, although models with LSTM bottlenecks may see slight degradation because the LSTM relies on rich feature maps.","Quantized models M1(Q) at 1.90 MB and M2(Q) at 2.02 MB fit in the memory of low-precision processors, supporting deployment on 8-bit capable edge hardware.","Unstructured pruning is not an effective compression strategy for these tightly tuned models, as even a 25% pruning ratio drops MSE scores below 20 for all proposed variants."],"supporting_citations":[{"why":"Defines the signal model, the four interference datasets, the -50 dB truncated MSE score, and the WaveNet baseline configuration used throughout the evaluation.","marker":"[14]"},{"why":"Supplies the depthwise separable convolution technique that the paper uses to cut MACs and parameters in the proposed U-Net models.","marker":"[18]"},{"why":"Provides the ConvTasNet baseline architecture and its parameter configuration (N=512, L=16, B=128, etc.) used as a comparison point.","marker":"[19]"},{"why":"Establishes the blind CCI scenario and shows that WaveNet degrades on unknown interference, motivating the need for blind-capable architectures.","marker":"[15]"},{"why":"Links the number of MACs to energy consumption, supporting the paper's claim that reducing MACs translates to lower power usage on edge devices.","marker":"[20]"}],"fun_headline_variants":["U-Nets cut blind interference MACs by 61% with minimal MSE change","Slim U-Nets beat WaveNet in MSE with 61% fewer MACs","Quantized U-Nets: 61% fewer MACs, <2.1 MB for blind CCI","Depthwise separable U-Nets cut compute by 61% for edge CCI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The practical value of the reported MSE scores rests on the assumption that a higher truncated negative log-MSE score automatically means better bit decoding after interference cancellation, but the paper does not test bit-error rate.","fun_headline_variants_meta":{"raw":{"variants":["U-Nets cut blind interference MACs by 61% with minimal MSE change","Slim U-Nets beat WaveNet in MSE with 61% fewer MACs","Quantized U-Nets: 61% fewer MACs, <2.1 MB for blind CCI","Depthwise separable U-Nets cut compute by 61% for edge CCI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001169,"raw_usage":{"total_tokens":4887,"prompt_tokens":1046,"completion_tokens":3841,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":3747}},"tokens_in":662,"tokens_out":3841,"duration_ms":23092,"temperature":1.0,"reasoning_tokens":3747,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:23:34.811147+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Decode the reconstructed QPSK signals from M1, M2, and the baselines on the same test data and compare bit-error rates; if M1 or M2 do not achieve lower BER than WaveNet and ConvTasNet, the central performance advantage is not confirmed for real communication.","supporting_citations":[{"cited_title":"The Data-Driven Radio Frequency Signal Separation Challenge,","cited_arxiv_id":null,"evidence_quote":"Defines the signal model, the four interference datasets, the -50 dB truncated MSE score, and the WaveNet baseline configuration used throughout the evaluation."},{"cited_title":"Conv-Tasnet: Surpassing Ideal Time– Frequency Magnitude Masking for Speech Separation,","cited_arxiv_id":null,"evidence_quote":"Provides the ConvTasNet baseline architecture and its parameter configuration (N=512, L=16, B=128, etc.) used as a comparison point."},{"cited_title":"Blind Co-Channel Interference Cancellation Using Fast Fourier Convolutions,","cited_arxiv_id":null,"evidence_quote":"Establishes the blind CCI scenario and shows that WaveNet degrades on unknown interference, motivating the need for blind-capable architectures."},{"cited_title":"A method to estimate the energy consumption of deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Links the number of MACs to energy consumption, supporting the paper's claim that reducing MACs translates to lower power usage on edge devices."}],"review_version":1}