{"id":"1b9e9cf3-63fc-4442-9d37-9d8a97f65ad1","arxiv_id":"2412.19471","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A meta-learning-based delayless subband adaptive filter using a self-attention RNN learns the adaptive filter update rule for active noise control and outperforms NFxLMS and DSNFxLMS in simulations.","lead":"This paper trains a neural network to act as the adaptive filter update rule in active noise control, replacing the hand-designed FxLMS algorithm with a learned update that handles nonlinear loudspeaker effects and nonstationary noise. It reports 1 to 5 dB lower residual noise than classical NFxLMS and delayless-subband baselines in simulated rooms, while running in real time on a desktop CPU.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed superiority over 'traditional methods' is not established because only two linear baselines (NFxLMS and DSNFxLMS) are compared; nonlinear adaptive filters cited in the paper's own introduction are omitted from the experiments.","rationale":"The reader's weakest_assumption focuses on simulation-to-real generalization. That is a valid limitation, but the paper explicitly scopes its results as 'simulation results' in the abstract, so the absence of real-world validation is a limitation rather than a direct contradiction of the stated claim. The more load-bearing issue is internal to the presented evidence: the comparison set is too narrow to support the broad 'superior to traditional methods' claim. The paper motivates nonlinear ANC and cites numerous nonlinear adaptive filters, yet evaluates only two linear baselines with a fixed step size. This directly affects whether Tables 2 and 3 substantiate the central claim even within the simulation domain. I partially agree with the reader because the rationale does mention omitted nonlinear baselines, but the reader's selected weakest assumption is the generalization gap. My concern does not change the verdict: CONDITIONAL remains appropriate, since the paper has a plausible meta-learning formulation, a delayless subband architecture with complexity analysis, and a secondary-path-robust variant. The condition should require a fair comparison against representative nonlinear adaptive filters (and ideally a cross-environment test) before the superiority claim is accepted.","tokens_in":16138,"tokens_out":7564,"duration_ms":74497,"concrete_test":"Implement and tune a second-order Volterra FxLMS filter and/or the tanh-based FxLMS from [22] on the same NOISEX-92 test conditions used in Tables 2 and 3, optimizing their step sizes per condition, and compare the resulting NMSE values against the reported MDSAF and MDSAF-MD numbers; if a nonlinear classical filter matches or beats the learned update rule, the superiority claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a learned update rule achieves superior noise reduction relative to traditional methods. However, the only adaptive-filter baselines used are NFxLMS and DSNFxLMS, both linear filters (Section 4.3). The paper's own motivation is the failure of linear updating rules under nonlinear loudspeaker saturation, and Section 1 surveys nonlinear traditional alternatives, including Volterra FxLMS [18-20], tanh-based FxLMS [21-22], and functional-link ANN filters [24-26]. None of these are evaluated. Moreover, the two included baselines are run with a single fixed step size (0.01) across all noise types and nonlinearity levels, which may unnecessarily handicap them. Consequently, the 1-5 dB improvements in Tables 2 and 3 demonstrate superiority over linear FxLMS with a fixed step size, but not over the broader class of traditional nonlinear adaptive filters referenced in the introduction. This gap is load-bearing because the paper's stated contribution is 'superior noise reduction performance compared to traditional methods,' not merely better-than-linear performance. Without a nonlinear adaptive-filter baseline, the empirical evidence does not rule out the possibility that a well-tuned classical nonlinear filter achieves comparable or better NMSE.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a meta-learning-based delayless subband adaptive filter (MDSAF) for single-channel feedforward active noise control. The adaptive filter weights are updated by a recurrent neural network with a complex self-attention module, and the update rule is trained end-to-end from noisy error observations without oracle labels. A modified delayless subband architecture and a skip-updating strategy reduce the required update rate, and a variant (MDSAF-MD) is trained using only the main delay of the secondary path. The experiments compare MDSAF and MDSAF-MD with NFxLMS and DSNFxLMS on four NOISEX-92 noise types at SNRs of 5, 15, and 25 dB and nonlinearity levels eta^2 = 0.5, 2, and infinity, with a mid-test primary path change. The proposed methods consistently achieve lower NMSE than the two linear baselines by roughly 1-5 dB.","tokens_in":16387,"tokens_out":9714,"duration_ms":80866,"significance":"If the reported results hold, the idea of learning an adaptive-filter update rule for ANC from noisy observations is a valuable contribution, particularly for nonlinear loudspeaker saturation and nonstationary noise. The evaluation has genuine strengths: the update rule is trained on separate data (ESC-50 and Nonspeech) and tested on held-out NOISEX-92 noises and unseen primary-path changes; the method does not require true labels; and the paper includes a real-time complexity analysis with measured inference time. The main weakness is that the empirical claim of superiority over 'traditional methods' is broader than the baseline set actually tested. With additional baselines and a correction to the gradient post-processing formula, the central idea would be well supported.","major_comments":[{"comment":"The claimed superiority over 'traditional methods' is not supported by the baseline set. The only adaptive-filter competitors are NFxLMS and DSNFxLMS, both linear filters, and their step sizes are fixed at 0.01 for all noise types, SNRs, and nonlinearity levels. Since the paper's motivation is the failure of linear updating rules under loudspeaker saturation, and Section 1 surveys nonlinear traditional alternatives (Volterra FxLMS [18-20], tanh-based FxLMS [21-22], functional-link ANN [24-26]), the experiments need at least one nonlinear classical baseline with a tuned or per-condition step size. As it stands, Tables 2 and 3 demonstrate superiority over linear FxLMS with a fixed step size, not over the broader class of traditional nonlinear adaptive filters invoked in the abstract.","section":"Section 4.3, Tables 2 and 3"},{"comment":"Equation (20) does not implement the stated amplitude clamp and appears to be independent of g(n). For every |g| < exp(10), the expression max[min(|g|, exp(-10)), exp(10)] evaluates to exp(10), and for |g| >= exp(10) it also evaluates to exp(10); hence |~g(n)| is the constant ln(exp(10)/10)+1, approximately 8.7, not a value in [0,2] as claimed in the surrounding text. This makes the exact gradient post-processing used in training and testing ambiguous. Please correct the formula or the description and verify that the reported experiments use the corrected version.","section":"Section 3.3, Eq. (20)"},{"comment":"The closest learned-update-rule baselines, the meta-learning adaptive filters of [48] and [49] from which the present architecture is directly derived, are not evaluated. Because the contribution is positioned as a meta-learning-based adaptive filter, a comparison with Meta-AF or an equivalent learned optimizer is necessary to show that the proposed self-attention and delayless subband modifications improve over the prior learned update rule, rather than only over classical linear FxLMS. Without this comparison, the novelty claim relative to [48,49] remains unquantified.","section":"Sections 3.1-3.3 and 4.3"}],"minor_comments":[{"comment":"The tables report averaged NMSE over 50 independent runs but do not report standard deviations or confidence intervals; for differences that are sometimes around 1 dB, please add variability measures or a significance test.","section":"Tables 2 and 3"},{"comment":"The abstract and contribution list claim robustness to 'various environments,' but all simulations use the same 5 m x 4 m x 3 m room and the same fNSE loudspeaker model; please temper the claim or add a second room geometry or a real-world recording.","section":"Section 4.2"},{"comment":"The text states that the compression in (17) 'does not improve performance' but is still applied; either remove it or provide an ablation, since as written this is contradictory.","section":"Section 3.3"},{"comment":"The contribution bullet says the source code is available, but no URL appears in the manuscript; please include the link for reproducibility.","section":"Section 1, contribution list"}],"recommendation":"major_revision","confidential_remarks":"The core idea is plausible and the held-out evaluation is a positive feature. The main blockers are the narrow baseline set (no nonlinear classical ANC filters and no prior meta-learning adaptive filters) and the unambiguous error in Eq. (20). If the authors add the missing comparisons, fix Eq. (20), and report variance, I would be willing to accept the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is a meta-learned update rule for ANC, wrapped in a delayless subband architecture and trainable with only the main delay of the secondary path. That is a real extension of Meta-AF, and the empirical story is mostly solid: consistent 1–5 dB NMSE gains over NFxLMS and DSNFxLMS, held-out noise types, a mid-test primary path change, and measured inference times that clear the real-time threshold. The fact that the filter is trained without true labels is a genuine plus.\n\nWhere the paper wobbles is in its central claim of superiority over “traditional methods.” The stress-test note is right: the only baselines are two linear FxLMS variants, and the introduction itself surveys nonlinear classical alternatives (Volterra, tanh FxLMS, functional-link ANN). You cannot claim to beat traditional methods without letting a few of those compete, especially when the nonlinearity model in the paper is exactly the saturation type that tanh-FxLMS was designed for. The fixed step size of 0.01 for both baselines also likely handicaps them; a tuning sweep would make the comparison fairer.\n\nAlso softer, but worth noting: the skip updating strategy is presented as a contribution and never evaluated. An ablation with B>0 is missing. The learnable feature embedding is not ablated either. The paper says “the source code is available at:” and then provides nothing, which is no help. No error bars, so I cannot tell whether the 1–5 dB gaps are statistically meaningful.\n\nThe core mechanism, though, appears sound. The reported numbers are held-out measurements, not fitted values, and the training protocol is clearly described. The omission of stronger baselines and ablations is fixable, not fatal. This deserves peer review—a serious referee should push for the missing comparisons and the code. I would not cite it as evidence of superiority over nonlinear ANC yet, but I would read a revision that addresses these points.","headline":"A promising learned-update-rule ANC paper whose headline claim outpaces its evidence because the baselines are only two linear FxLMS variants.","tokens_in":16922,"tokens_out":1814,"would_cite":false,"duration_ms":18193,"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":"A neural network trained as a meta-learning update rule outperforms classical adaptive filters in active noise control.","keywords":["active noise control","adaptive filter","meta-learning","delayless subband","complex self-attention","recurrent neural network","nonlinear noise control","skip updating"],"falsifier":"Run the trained MDSAF model on a real single-channel ANC setup with a physical loudspeaker, reference microphone, and error microphone, and compare its noise reduction against NFxLMS with a properly tuned step size; if the learned rule does not beat or match the baseline on real recorded noise, the simulation-to-real transfer claim fails. Alternatively, measure whether the model remains stable when the secondary path is replaced by a measured impulse response with a different delay than in training.","tokens_in":15901,"feed_emoji":"🎧","tokens_out":3825,"duration_ms":32060,"temperature":0.7,"pith_summary":"The paper claims that active noise control can be improved by replacing the hand-designed linear update rule of an adaptive filter with a neural network that learns the update rule itself. This neural network is trained using only noisy error observations, not clean target signals, and it adapts the filter weights to suppress noise in nonlinear and nonstationary environments. The authors report that their method, called MDSAF, reduces the normalized mean squared error by about 1 to 5 dB compared with the normalized filtered-x LMS and its delayless subband variant across four noise types, three signal-to-noise ratios, and several loudspeaker nonlinearity levels. If this holds, a data-driven update rule could make active noise control systems effective in conditions where classical linear updates struggle, such as loudspeaker saturation and sudden path changes.","feed_headline":"Learned ANC update rule beats classic filters by up to 5 dB","feed_subtitle":"A neural network trained on noisy observations alone adapts faster and cuts more noise in nonlinear settings.","key_machinery":"The central object is a neural network used as the adaptive filter's update rule. The network is a single-headed attention recurrent network (SHA-RNN) whose attention block uses learnable feature embeddings to weight the elements of the input vector, and its output is a complex-valued gradient that is amplitude-limited to act like a normalized step. This update rule is embedded in a modified delayless subband architecture: the filtered reference and error signals are split into subbands by an analysis filter bank, each subband weight is updated in the frequency domain, and the fullband filter is reconstructed by weight stacking and IFFT. The architecture also includes a skip-updating factor that reduces how often the network runs, relaxing the real-time constraint.","core_discovery":"In the paper's own terms, the authors establish that a meta-learning-based delayless subband adaptive filter, using a single-headed attention recurrent network with learnable feature embedding as the update rule, can adapt an ANC filter from noisy observations alone. The network predicts a frequency-domain gradient that updates the adaptive filter weights, and the delayless subband architecture plus a skip-updating strategy let the update happen less frequently without adding signal delay. Simulations with the fNSE loudspeaker saturation model show the learned rule achieves lower NMSE than NFxLMS and DSNFxLMS in all tested conditions, including when the primary path changes suddenly mid-test. A variant trained with only the main delay of the secondary path still outperforms the classical baselines, indicating the rule does not need exact secondary-path knowledge.","pith_inferences":["A natural next step is to test the trained update rule on real measured impulse responses and a physical loudspeaker; the biggest risk is that the simulation's room geometry and saturation model do not cover real acoustic variability.","The same meta-learning formulation could be applied to other adaptive filtering tasks, such as echo cancellation or feedback control, wherever a linear update rule is the bottleneck.","The reported 1–5 dB gain is over fixed-step-size classical baselines; an even more direct comparison would tune the baselines' step sizes per condition."],"forward_implications":["ANC controllers can be trained end-to-end from noisy data without clean reference signals, removing a major obstacle to deep learning in practical noise control.","The delayless subband and skip-updating design mean the learned update can run in real time on moderate hardware, since the update frequency drops by the downsampling factor.","The model's ability to adapt to sudden primary-path changes suggests learned update rules generalize to nonstationary acoustic environments better than fixed-linear algorithms.","Training with only the main delay of the secondary path indicates the approach can work when exact secondary-path identification is unavailable, easing deployment."],"supporting_citations":[{"why":"Supplies the meta-learning formulation of learning an update rule for adaptive filters, which the paper adapts to ANC.","marker":"[49]"},{"why":"Provides the original delayless subband adaptive filter architecture that the paper modifies.","marker":"[58]"},{"why":"Defines the DSNFxLMS algorithm used as a baseline for comparison.","marker":"[77]"},{"why":"Supplies the fNSE loudspeaker saturation nonlinearity model used in the simulations.","marker":"[52]"},{"why":"Provides the spatial room arrangement and reverberation settings used in the simulated ANC environment.","marker":"[42]"},{"why":"Describes the FFT-based weight stacking method used to reconstruct fullband filter weights.","marker":"[59]"}],"fun_headline_variants":["Meta-learning ANC rule beats classic filters without true labels","Neural ANC updater learns from noisy observations alone","Self-attention subband ANC beats classics without clean labels","Meta-learned ANC filter: no secondary path knowledge required"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned update rule transfers from the specific simulated room, noise set, and loudspeaker saturation model used in training to real acoustic environments and hardware.","fun_headline_variants_meta":{"raw":{"variants":["Meta-learning ANC rule beats classic filters without true labels","Neural ANC updater learns from noisy observations alone","Self-attention subband ANC beats classics without clean labels","Meta-learned ANC filter: no secondary path knowledge required"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000473,"raw_usage":{"total_tokens":2362,"prompt_tokens":968,"completion_tokens":1394,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1329}},"tokens_in":584,"tokens_out":1394,"duration_ms":11223,"temperature":1.0,"reasoning_tokens":1329,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:34:03.404392+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained MDSAF model on a real single-channel ANC setup with a physical loudspeaker, reference microphone, and error microphone, and compare its noise reduction against NFxLMS with a properly tuned step size; if the learned rule does not beat or match the baseline on real recorded noise, the simulation-to-real transfer claim fails. Alternatively, measure whether the model remains stable when the secondary path is replaced by a measured impulse response with a different delay than in training.","supporting_citations":[{"cited_title":"Casebeer, N","cited_arxiv_id":null,"evidence_quote":"Supplies the meta-learning formulation of learning an update rule for adaptive filters, which the paper adapts to ANC."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the original delayless subband adaptive filter architecture that the paper modifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the DSNFxLMS algorithm used as a baseline for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the fNSE loudspeaker saturation nonlinearity model used in the simulations."},{"cited_title":"Zhang, D","cited_arxiv_id":null,"evidence_quote":"Provides the spatial room arrangement and reverberation settings used in the simulated ANC environment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the FFT-based weight stacking method used to reconstruct fullband filter weights."}],"review_version":1}