{"id":"2d0817fd-b6be-40c0-b8db-45f032a627cd","arxiv_id":"2502.07208","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Frequency-dependent augmentation and convolution improve sound event detection through complementary mechanisms, as shown by class-wise, Grad-CAM, and PCA analyses.","lead":"This paper dissects two frequency-aware tricks for sound event detection: FilterAugment, which randomly boosts or cuts frequency bands during training, and frequency dynamic convolution, which adapts its convolution kernel to each frequency bin. It finds that the two methods help through different routes and that combining them gives a larger gain than either alone, while also offering class-by-class explanations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that FDY conv's gain stems from relaxing frequency-axis translation equivariance is not isolated from the added attention capacity; the missing non-frequency-adaptive attention control is the load-bearing gap.","rationale":"I read the paper's central claim as the causal statement that frequency dependence is what makes FDY conv effective and that FilterAugment and FDY conv are complementary. The paper provides three convolution variants, but the comparison cannot separate 'frequency-adaptive' from 'extra content-adaptive attention capacity.' The FK conv ablation is informative because it shows that naive per-bin kernels hurt, and FW conv's modest gain is the only clean evidence for frequency dependence. FDY conv's larger gain is attributed to the attention mechanism without a control that removes frequency specificity while preserving attention capacity. The reader's weakest assumption is exactly this, and I agree. A concrete control, same attention module with weights shared across frequency, would settle it. I also note a secondary logical issue in the complementarity argument: the observed combined improvement of 11.6% is less than the multiplicative expectation of about 14.4%, which suggests overlap rather than 'mostly unrelated' or 'complementary.' That point is in Section IV.A and weakens the wording, but the missing control is more load-bearing because it threatens the core interpretation of why FDY conv works. The Grad-CAM and PCA analyses are qualitative and self-limited, so they do not rescue the causal claim. Since the reader already conditioned acceptance on this weakness, I leave the verdict unchanged.","tokens_in":17581,"tokens_out":3456,"duration_ms":33562,"concrete_test":"Train the baseline CRNN with an 'FDY-but-not-frequency-adaptive' variant: replace the per-frequency-bin attention weights with a single attention vector shared across all frequency bins, keeping the basis-kernel count K and the attention module parameter count identical to FDY conv. If this control matches FDY conv's PSDS1+PSDS2 within run-to-run variation, then frequency adaptation is not the cause of the gain; if it drops back to FW-conv-level performance, the frequency-specificity claim is supported. Report results over the same 12-seed protocol, preferably with error bars.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II.B motivates FDY conv as relaxing translation equivariance along the frequency axis, and Section IV.C interprets its 8.95% gain over baseline (compared to +3.72% for FW conv and -14.66% for FK conv) as evidence that the attention mechanism has fitted FW conv well. However, FDY conv differs from FW conv not only in making kernel weights input-dependent along frequency, but also in containing a learnable attention module with extra parameters (two 1D convolution layers). No control applies the same attention module in a frequency-agnostic manner, such as sharing one attention vector across all frequency bins or making the weights time-adaptive. Therefore, the observed gain could come from added attention capacity or content-adaptive weighting generally, not specifically from relaxing frequency-axis translation equivariance. If that is the case, the central claim that frequency dependence is the operative factor collapses, and the paper's own FK conv result shows that a purely frequency-dependent kernel can even hurt. This is the load-bearing assumption: the causal role of frequency adaptation is not isolated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an empirical analysis of two frequency-dependent techniques for sound event detection (SED) on the public DESED dataset with a DCASE 2022 Task 4-style CRNN baseline. The authors compare FilterAugment and frequency dynamic convolution (FDY conv) against a baseline, report class-wise CB-F1 box plots, use Grad-CAM to compare two FilterAugment variants with frequency masking, introduce two simpler frequency-dependent convolution variants (FK conv and FW conv), and apply PCA to the attention weights of FDY conv. The central claims are that frequency dependence plays a significant role in SED, that FilterAugment and FDY conv are effective and complementary, and that FDY conv works by relaxing translation equivariance along the frequency axis.","tokens_in":17733,"tokens_out":5038,"duration_ms":46551,"significance":"If the interpretational claims were supported, the paper would be a useful design guide for frequency-adaptive SED architectures, since FDY conv is already used in several strong SED systems and FilterAugment is a common augmentation. The manuscript is transparent about several limitations, including the qualitative nature of the Grad-CAM case studies and the controlled foreground-only setting of the PCA analysis, and it provides public code. The main analytical conclusions, however, currently rest on an uncontrolled comparison, an unsupported product-rule argument, and maximum-over-seeds reporting, so the significance is conditional on the additional experiments and reinterpretation described below.","major_comments":[{"comment":"The complementarity claim is load-bearing and rests on an unjustified calculation. The text states that if the two methods were 'completely unrelated,' the joint improvement would equal the product of their individual improvements (14%), and since the observed improvement is 11.6%, the methods are 'complementary rather than overlapping.' This product rule is not derived from any property of the metrics (PSDS1+PSDS2 is a sum, not a multiplicative score) or from a statistical model of interactions. Moreover, Table III reports only the maximum over 12 runs, so the difference between 11.6% and 14% is within the range one would expect from seed variability. Please replace this argument with an interaction analysis based on the actual metric definitions and with error bars, or remove the complementarity claim and report the 2x2 factorial results descriptively.","section":"Section IV.A"},{"comment":"The causal role of frequency adaptation in FDY conv is not isolated. FDY conv differs from FW conv not only in making kernel weights input-dependent per frequency bin but also in containing a learnable attention module with two additional 1D convolution layers. The FK/FW/FDY comparison varies frequency-dependence and attention capacity simultaneously, so the observed +8.95% over baseline and +5.23% over FW conv could be due to added capacity or content-adaptive weighting in general rather than to relaxing frequency-axis translation equivariance. A control that applies the same attention mechanism in a frequency-agnostic manner, such as sharing one attention vector across all frequency bins or making the attention time-adaptive, is needed. Without this control, the central interpretation in the abstract and conclusion—that frequency dependence itself drives the gain—is not established, especially because FK conv, a purely frequency-dependent kernel, degrades performance by 14.66%.","section":"Section II.B and Section IV.C"},{"comment":"All headline performance claims report the maximum score over 12 training runs, with no variance, mean, or significance test reported. Because the paper's analytical conclusions are comparative (FDY better than FW, FDY+FA better than FDY+FM, FDY+FA improves the baseline by 11.6%), reporting maxima makes the rankings and the magnitudes unstable and non-falsifiable. Please report the mean and standard deviation (or median and interquartile range) across the 12 runs, and where feasible a paired test across seeds. The class-wise box plots in Fig. 6 should also be accompanied by a statistical test or effect sizes, especially because the stationary/non-stationary classification in Section IV.A is developed after inspecting those box plots.","section":"Section III.H and Table III"},{"comment":"The class-wise conclusions are derived post hoc. The statements that FDY conv is 'especially strong on non-stationary sound events' and that FilterAugment is unsuitable for events with 'specific and localized frequency peaks' are based on visual inspection of box plots and on a categorization that is adjusted after seeing the results: electric shaver/toothbrush, frying, and running water are reclassified as 'partially non-stationary' precisely because FDY conv improved on them. This risks circularity. The event categories should be defined a priori using an objective acoustic measure (for example spectral flux or stationarity), and the class-wise hypotheses should then be tested with an appropriate multiple-comparison procedure.","section":"Section IV.A and Section V"},{"comment":"The PCA analysis is used to conclude that 'FDY conv applies frequency-dependent convolution kernels' and that the class-wise separability 'supports its effectiveness,' but the analysis is descriptive and has no null baseline. For example, the observed local clustering of similar colors is unsurprising because the attention weights are produced by 1D convolutions over frequency and are therefore locally smooth; a null model with randomized attention weights or a comparison with the fixed weights of FW conv would be needed to show that the PCA structure is specific to FDY conv's frequency-adaptive mechanism. Please temper the causal wording and, if possible, add a quantitative dispersion or separability measure.","section":"Section IV.D"}],"minor_comments":[{"comment":"Two different tables are both labeled 'Table IV': the Grad-CAM performance table in Section IV.B and the convolution-method comparison in Section IV.C. Renumber the tables and fix the in-text cross-references.","section":"Section IV.B and Section IV.C"},{"comment":"Please proofread for typographical errors, including 'equivament' (Section II.B), 'laeyrs' and 'tatget' (Section IV.B), 'ocassionally' (Section IV.B), 'resluts' (Section IV.A), and 'ths' (author biography).","section":"Throughout"},{"comment":"The notation in Eq. (1) is unclear: T(x) is used without defining its action on a spectrogram, and the terms 'shift-invariant' and 'shift-variant' are used interchangeably with 'translation equivariance.' Consider defining the operators and consistently using 'translation equivariance' for the convolution property.","section":"Section II.B"},{"comment":"The LM-GC metric is described as the squared difference between log-mel and Grad-CAM outputs averaged over active time-frequency regions, but the exact normalization and the criterion for 'active' regions are not specified. A formal definition would make the values in Table IV reproducible.","section":"Section IV.B"},{"comment":"The sentence 'This does not merely claim the excellence of attention mechanism which adapts to the contents of convolution input' is ungrammatical and should be rephrased for clarity.","section":"Section IV.C"},{"comment":"Please state how many audio clips per class were used in the PCA analysis and report the percentage of variance explained by the two principal components; otherwise the plot in Fig. 8 is difficult to interpret.","section":"Section IV.D"},{"comment":"Several references are incomplete or inconsistent, for example [12] and [14] lack venue and page information. The running header still contains the template placeholder '> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER'; remove it before submission.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper analyzes the authors' own prior methods, which is acceptable but creates a risk of over-claiming: the 'complementary' and 'frequency-dependence causes the gain' conclusions are presented more strongly than the experimental design supports. The missing non-frequency-adaptive attention control is the most important gap; if it cannot be added, the authors should substantially weaken the causal claims. The product-rule argument for independence should be removed or replaced regardless, since it is not derived from the metric definitions. The use of maximum-over-12-runs is also a major reproducibility concern for an analysis paper whose conclusions are comparative."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a mostly solid, narrowly scoped analysis of the authors' own FilterAugment and FDY conv, and it earns its keep by adding two simple baselines (FK and FW conv) plus class-wise, Grad-CAM, and PCA breakdowns. It does not settle a big question, but it gives SED architecture designers something concrete. The main flaw is the one the stress-test flags: FDY conv's advantage over FW conv is attributed to relaxing translation equivariance, but FDY conv also adds a learnable attention module. No control applies the same attention mechanism in a frequency-agnostic way, so you cannot tell whether the 8.95% gain comes from content-adaptive weighting generally or from frequency-specific adaptation. That is a real gap, but not fatal: the FK vs FW comparison already shows that bare frequency-dependent kernels help only when a shared basis is used, and FilterAugment's effect supports frequency dependence through a different route.\n\nWhat is new: FK conv and FW conv are simple, clean probes. The class-wise box plots are post hoc but suggestive, and the PCA showing structured per-class frequency patterns adds a concrete illustration. The Grad-CAM section is honest—the authors explicitly label it qualitative, which is rare and welcome. They also disclose that the performance tables report only the maximum over 12 runs and that their numbers are not comparable to their earlier papers because they used a simpler median filter. That disclosure is a point in their favor.\n\nSoft spots, in order of severity: (1) no variance or median on the headline 11.6% joint gain, so the complementarity claim (11.6% vs the expected 14%) is weaker than it looks; that product heuristic also assumes independent mechanisms, which is dubious. (2) Class-wise pros and cons are read off box plots without statistical tests; labels like 'non-stationary' are plausible but fitted to the data. (3) The missing attention control is the load-bearing gap for the FDY interpretation—a real confound. (4) Minor: the computational complexity analysis is fine but not central.\n\nWho it's for: SED researchers thinking about frequency-adaptive CNNs, and anyone doing interpretability analyses of augmentation or architecture choices. A fair referee would ask for a control experiment and some error bars, but the core probes deserve publication.\n\nRecommendation: I'd engage with it—send it to peer review, conditional on addressing the attention confound.","headline":"A workmanlike analysis of the authors' own frequency-dependent SED methods, with useful new probes and honest limitations, but the FDY conv gain is never isolated from added attention capacity.","tokens_in":18340,"tokens_out":2244,"would_cite":true,"duration_ms":21347,"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":"Frequency dependence is a real and exploitable structure in sound event detection: two frequency-aware methods improve SED for distinct reasons and combine for an 11.6% gain over the baseline.","keywords":["sound event detection","frequency dependence","FilterAugment","frequency dynamic convolution","translation equivariance","Grad-CAM","data augmentation","time-frequency analysis"],"falsifier":"Train a model identical to the FDY conv model except that the attention weights are shared across all frequency bins, or replaced by a fixed per-bin mask with the same number of parameters. If this frequency-independent attention control matches the 8.95 percent gain over baseline, then the frequency-adaptive kernel hypothesis is not what explains the improvement; if it does not, the paper's interpretation is supported.","tokens_in":17306,"feed_emoji":"🔊","tokens_out":5976,"duration_ms":52194,"temperature":0.7,"pith_summary":"This paper tries to establish that sound event detection should treat the frequency axis differently from the time axis, and that two previously proposed frequency-aware techniques work well and work for complementary reasons: FilterAugment, which randomly weights frequency bands during training, and frequency dynamic convolution, which applies kernels that adapt to each frequency bin. The authors analyze class-level performance, visualize where models attend using Grad-CAM, build simpler frequency-dependent convolutions as controls, and run PCA on the attention weights inside the adaptive kernel mechanism. Their combined use improves the baseline by 11.6 percent on the DESED domestic-environment benchmark, a gain close to what independent effects would predict. If the analysis is right, frequency-adaptive data augmentation and frequency-adaptive convolution should both be part of SED systems, because they target different failure modes: FilterAugment helps wideband events but can hurt sounds with localized spectral peaks, while frequency dynamic convolution helps non-stationary events.","feed_headline":"Pairing frequency augmentation with dynamic kernels lifts SED 11.6%","feed_subtitle":"The two approaches are complementary: one broadens spectral evidence, the other adapts kernels per frequency bin.","key_machinery":"The central mechanism is the relaxation of translation equivariance along the frequency axis. Ordinary 2D convolution assumes that shifting input content along either axis changes only position, not meaning; this paper argues that is false for spectrograms because shifting a sound in frequency changes the sound itself. FDY conv implements frequency-dependent kernels by squeezing a time-pooled input through two 1D convolution layers to produce, for each frequency bin, K attention weights, then forming that bin's kernel as a weighted sum of K trainable basis kernels. FilterAugment is the complementary data-side mechanism: during training it multiplies random frequency bands by random gains, either as step-like bands or linearly interpolated weights, so the model must learn sound-event evidence across a range of spectral conditions. Two simpler variants, FK conv with a separate trained kernel per frequency bin and FW conv with fixed per-bin weighted sums of shared basis kernels, act as controls that isolate what the attention mechanism adds.","core_discovery":"On a CRNN sound event detection model trained on the DESED dataset with the mean-teacher semi-supervised recipe, the paper claims that FilterAugment and frequency dynamic convolution are each helpful and mostly independent: FilterAugment raises CB-F1 from 0.478 to 0.486, adding FDY conv to the baseline raises it to 0.505, and using both reaches 0.517, an 11.6 percent relative gain over baseline that is close to the 14 percent predicted by multiplying the individual gains. Class-wise F1 shows FDY conv is especially strong on non-stationary events such as alarm/bell, dishes, and running water, and weaker on stationary machinery sounds like blender and vacuum cleaner, while FilterAugment consistently helps wideband events but hurts cat, dog, and electric shaver/toothbrush, whose sounds have localized spectral peaks. Grad-CAM visualizations show that FilterAugment draws attention into time-frequency regions tied to the true event, with step-type weighting focusing on narrow, sharp regions and linear-type weighting on broader, continuous regions. Replacing FDY conv with fixed per-frequency kernels (FK conv) hurts performance by 14.66 percent, while fixed frequency-wise weighted sums of basis kernels (FW conv) helps by 3.72 percent, and FDY conv's input-dependent frequency attention helps by 8.95 percent; PCA of the attention weights shows neighboring frequency bins receive similar kernels, and the kernel trajectories differ by sound class and become more distinct in later layers.","pith_inferences":["Beyond the paper: the same frequency-is-shift-variant argument should transfer to audio tagging, acoustic scene classification, and other spectrogram-based tasks, predicting that combining global spectral perturbation with local kernel adaptation will help there too, but this is an extrapolation rather than a result of the paper.","Beyond the paper: the smooth frequency trajectories seen in the PCA plots suggest the attention weights could be parameterized more cheaply, for example as a low-rank or smoothly interpolated function of frequency, reducing FDY conv's overhead without losing most of its benefit.","Beyond the paper: a matched control that keeps FDY conv's attention module and basis-kernel count but makes the attention weights frequency-independent would settle whether the gain is really due to frequency adaptation or to added model capacity; the paper does not run this control."],"forward_implications":["Combining FilterAugment with FDY conv gives an 11.6 percent relative gain over the baseline, close to the 14 percent expected if the two methods are independent, so a strong SED system should treat them as complements rather than substitutes.","Vanilla 2D convolution over spectrograms carries an inappropriate inductive bias for the frequency axis; replacing it with kernels that vary along frequency improves detection, but only when the kernels share basis across neighboring bins, since separate per-bin kernels hurt performance.","FilterAugment is not universally beneficial: it degrades classes with narrow, fixed spectral peaks, so augmentation schedules may need to be class-aware or gentler for those sounds.","FDY conv's benefit concentrates on non-stationary events, so frequency-adaptive convolution is most valuable in layers where transient, spectrally evolving patterns must be captured."],"supporting_citations":[{"why":"Supplies the FDY conv method, its frequency-shift-variance motivation, and the setting used to train the FDY model.","marker":"[26]"},{"why":"Supplies FilterAugment, its step and linear variants, and the optimal hyperparameters used throughout the experiments.","marker":"[35]"},{"why":"Supplies the DESED dataset and the mean-teacher semi-supervised training framework that produces the strong, weak, and consistency losses.","marker":"[49]"},{"why":"Supplies Grad-CAM, adapted here to time-wise SED predictions to produce the attention maps.","marker":"[47]"},{"why":"Supplies the frequency masking augmentation used as the comparison condition in the Grad-CAM study.","marker":"[13]"},{"why":"Supplies the mean-teacher consistency training algorithm used for semi-supervised learning on unlabeled audio.","marker":"[51]"},{"why":"Supplies the PSDS metric used to evaluate detection performance alongside CB-F1.","marker":"[55]"}],"fun_headline_variants":["Frequency augmentation plus dynamic kernels yields 11.6% SED boost","FilterAugment + adaptive frequency kernels: 11.6% better SED","Two frequency-dependent SED methods combine for an 11.6% gain","Pairing FilterAugment with dynamic kernels lifts SED 11.6%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that FDY conv's gains come from letting kernels change along frequency rather than from its extra parameters and attention machinery; the paper interprets the gain through frequency adaptation but does not run a control that removes frequency adaptation while keeping the added capacity.","fun_headline_variants_meta":{"raw":{"variants":["Frequency augmentation plus dynamic kernels yields 11.6% SED boost","FilterAugment + adaptive frequency kernels: 11.6% better SED","Two frequency-dependent SED methods combine for an 11.6% gain","Pairing FilterAugment with dynamic kernels lifts SED 11.6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000981,"raw_usage":{"total_tokens":4251,"prompt_tokens":1120,"completion_tokens":3131,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":736,"completion_tokens_details":{"reasoning_tokens":3047}},"tokens_in":736,"tokens_out":3131,"duration_ms":21003,"temperature":1.0,"reasoning_tokens":3047,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:27:16.143678+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a model identical to the FDY conv model except that the attention weights are shared across all frequency bins, or replaced by a fixed per-bin mask with the same number of parameters. If this frequency-independent attention control matches the 8.95 percent gain over baseline, then the frequency-adaptive kernel hypothesis is not what explains the improvement; if it does not, the paper's interpretation is supported.","supporting_citations":[{"cited_title":"Frequency Dynamic Convolution: Frequency-Adaptive Pattern Recognition for Sound Event Detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the FDY conv method, its frequency-shift-variance motivation, and the setting used to train the FDY model."},{"cited_title":"FilterAugment: An Acoustic Environmental Data Augmentation Method,","cited_arxiv_id":null,"evidence_quote":"Supplies FilterAugment, its step and linear variants, and the optimal hyperparameters used throughout the experiments."},{"cited_title":"Sound Event Detection in Domestic Environments with Weakly Labeled Data and Soundscape Synthesis,","cited_arxiv_id":null,"evidence_quote":"Supplies the DESED dataset and the mean-teacher semi-supervised training framework that produces the strong, weak, and consistency losses."},{"cited_title":"Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization,","cited_arxiv_id":null,"evidence_quote":"Supplies Grad-CAM, adapted here to time-wise SED predictions to produce the attention maps."},{"cited_title":"Specaugment: A simple data augmentation method for automatic speech recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the frequency masking augmentation used as the comparison condition in the Grad-CAM study."},{"cited_title":"Mean teachers are better role models: Weight-averaged consistency targets improve semi- supervised deep learning results,","cited_arxiv_id":null,"evidence_quote":"Supplies the mean-teacher consistency training algorithm used for semi-supervised learning on unlabeled audio."},{"cited_title":"A Framework for the robust evaluation of Sound Event Detction,","cited_arxiv_id":null,"evidence_quote":"Supplies the PSDS metric used to evaluate detection performance alongside CB-F1."}],"review_version":1}