{"id":"24f4c02f-48b6-459c-a49b-837fbbf4cf23","arxiv_id":"2506.12785","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A family of frequency-adaptive convolutions improves CRNN sound event detection on DESED by up to 10.98% in PSDS1, with a lighter TFD variant matching the best score.","lead":"This dissertation develops a family of frequency-adaptive convolution modules that change their filters according to the frequency content of audio, and tests them on the DESED sound event detection benchmark. The best module improves a standard CRNN baseline by about 11% on the PSDS1 metric, and a lighter variant reaches the same score with fewer parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline gain is not capacity-controlled: MDFD and TFD use 18.2M/12.7M parameters versus a 4.4M baseline, and no matched-capacity static CRNN is evaluated.","rationale":"The reader's weakest assumption identifies attribution of the PSDS1 gain to frequency adaptivity versus capacity, attention, and validation selection. My stress-test agrees and isolates the most decisive untested confound: parameter count. MDFD and TFD are not controlled against a standard CRNN of comparable size, so the central empirical claim is underdetermined. The paper does include useful evidence (PFD at 5.041M reaches 0.442, and the Table 3.5 sweep shows a non-monotonic relationship between dynamic proportion and performance), which argues against a simple 'more dynamic parameters is better' story. However, none of these comparisons controls for static capacity. Two additional inconsistencies, such as baseline PSDS1 being 0.396 in Table 3.1 and 0.410 in Table 4.2, and Table 3.10 showing FDY baseline at 0.444 versus 0.441 elsewhere, reinforce the need for a single reproducible, matched-capacity experiment before the frequency-adaptivity claim is accepted. This is an addressable experimental gap, not a mathematical contradiction, so the appropriate verdict remains CONDITIONAL rather than REJECT. No ad hominem is intended; the critique is about the evidence base for the causal attribution.","tokens_in":47520,"tokens_out":6939,"duration_ms":83874,"concrete_test":"Train a standard 2D CRNN with channel widths increased to match the parameter counts of TFD-CRNN (12.703M) and MDFD-CRNN (18.157M), using the exact training recipe (mean-teacher, mixup/FilterAugment, same post-processing, PSDS1 evaluation, same validation split) and at least three seeds. If either matched-capacity static CRNN reaches PSDS1 >= 0.455, the headline improvement is not attributable to frequency adaptivity; if both stay near the 0.410 baseline, capacity is ruled out as the explanation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Table 4.2 compares MDFD-CRNN (18.157M params, PSDS1 0.455) and TFD-CRNN (12.703M, PSDS1 0.455) against a baseline CRNN with 4.428M params, a 4.1x and 2.9x capacity gap. The paper never trains a standard 2D CRNN with matched parameter count under the same mean-teacher, augmentation, and post-processing pipeline. Thus the 7.6%-10.98% PSDS1 improvements attributed to frequency adaptivity could be partly or wholly a capacity effect. PFD-CRNN (5.041M, 0.442) and the Table 3.5 proportion sweep provide partial evidence, but even a static CRNN at 5.0M params is untested. The issue is compounded by model selection on the same DESED real validation set across Tables 3.3-3.9 and 4.3-4.5, with no repeated-seed variance reported, so differences of 0.004-0.007 PSDS1 are not known to be above noise. The load-bearing condition for the central claim is that frequency adaptivity, not extra parameters or validation-set selection, causes the observed gain; this condition is currently untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The dissertation proposes a family of frequency-adaptive convolutions for sound event detection: Frequency Dynamic Convolution (FDY conv) and its extensions DFD, PFD, MDFD, and TFD. The core idea is to replace standard 2D convolutions with kernels that are dynamically combined using frequency-dependent attention weights, motivated by the argument that frequency-axis shift invariance is inappropriate for audio. The models are evaluated on DESED using PSDS1, with reported improvements over a CRNN baseline, parameter-efficiency claims for PFD and TFD, a class-wise analysis, and three engineering case studies. The paper claims MDFD is the best performer (PSDS1 0.455 vs 0.410 baseline) and that TFD matches it with 30% fewer parameters.","tokens_in":47776,"tokens_out":4461,"duration_ms":48285,"significance":"If the reported gains are real, the work is significant: it offers a systematic exploration of frequency-adaptive convolutions in a realistic SED benchmark and produces a parameter-efficient variant that matches the best accuracy. The paper also deserves credit for evaluating on the external DESED benchmark, using a standard mean-teacher pipeline, reporting parameter counts, and attempting class-wise and qualitative analyses. However, the central claim that frequency adaptivity causes the observed improvements is currently not established: the comparisons are not capacity-controlled, the headline model is selected from many configurations on the same validation set without variance estimates, and there are internal numeric inconsistencies in the baseline and parameter-efficiency numbers. These issues are load-bearing because the paper's contribution is precisely the attribution of the gains to the proposed mechanism.","major_comments":[{"comment":"The baseline score is inconsistent across the manuscript: Table 3.1 reports a baseline PSDS1 of 0.396, while Table 4.2 (and Table 3.2) reports 0.410 for the same CRNN with 4.428M parameters. Correspondingly, Section 3.3.3 states that FDY conv improves over the baseline by 8.95% (0.434 vs 0.396), whereas the abstract and Table 4.2 imply a 7.56% improvement (0.441 vs 0.410). The headline 10.98% gain for MDFD conv is computed from 0.455 vs 0.410, so the choice of baseline directly changes the central quantitative claim. The authors should identify which configuration and post-processing setting each table refers to and make the numbers consistent.","section":"§4.2, Table 4.2 vs §3.2.4, Table 3.1"},{"comment":"The main comparison is not capacity-controlled. MDFD conv uses 18.157M parameters and TFD conv uses 12.703M, versus 4.428M for the baseline CRNN, a 4.1x and 2.9x gap. No standard 2D CRNN with a matched parameter count is trained under the same mean-teacher, augmentation, and post-processing pipeline. The 7.6%-10.98% PSDS1 gains could therefore be partly or wholly due to extra model capacity rather than to frequency adaptivity. PFD conv at 5.041M/0.442 provides partial evidence, but even a static CRNN at 5M parameters is untested. The load-bearing condition for the paper's claim should be tested by training static CRNNs with matched parameter counts (for example, approximately 5M, 12M, and 18M parameters) under otherwise identical settings.","section":"Table 4.2"},{"comment":"All proposed modules are selected on the same DESED real validation set that is later used for the headline comparison. Tables 3.3-3.9 and 4.3-4.5 sweep dilation sets, branch proportions, channel widths, and pooling components, and the best configurations are then reported in Table 4.2. No repeated-seed experiments or variance estimates are provided, so differences of 0.004-0.007 PSDS1 (e.g., 0.448 vs 0.441, or 0.455 vs 0.451) are not shown to be above evaluation noise. The authors should report multiple runs (or confidence intervals) for at least the key comparisons, or otherwise demonstrate that the selected configurations are not artifacts of validation-set overfitting.","section":"§3.3-§3.7 and §4.2-§4.3"},{"comment":"The parameter-efficiency claim is internally inconsistent. Section 3.5.3 states that PFD-CRNN (1/8) has 5.401M parameters and reduces parameters by 51.9% relative to FDY conv (11.061M), while Section 3.7.4 and Table 4.2 list PFD conv as 5.041M, and the abstract reports a 54.4% reduction. These numbers imply different model configurations or an arithmetic error. Since parameter efficiency is a stated contribution, this needs to be resolved with a single consistent set of model sizes and a clear definition of which proportion and channel configuration is being reported.","section":"§3.5.3, §3.7.4, Table 4.2, Abstract"}],"minor_comments":[{"comment":"There are typos in the abstract and keywords: 'leadin to inconsistencies' should be 'leading to', 'translational equivariacne' should be 'translational equivariance', and 'temproal attention pooling' should be 'temporal attention pooling'.","section":"Abstract and Keywords"},{"comment":"The text calls the property in Eq. (3.1) 'shift invariance' but the equation T(F(x)) = F(T(x)) actually defines translation equivariance. The authors should either fix the mathematical statement or adjust the terminology, since the distinction matters for the paper's motivation.","section":"§3.1, Eq. (3.1)"},{"comment":"The text refers to 'Table 3.7' when presenting PFD conv results, but the results are in Table 3.5. The cross-reference should be corrected.","section":"§3.5.3"},{"comment":"In Eq. (3.11), the first two terms use xs,t but the third term uses xt; the subscript convention should be unified, and the dimensions of the attention weights should be stated explicitly.","section":"§3.7.2, Eqs. (3.11)-(3.13)"},{"comment":"The comparison with 1D CRNN is interesting but the 1D model is described as 'adapted from Wav2Vec2.0 and HuBERT' without a precise layer configuration; please provide the exact architecture (kernel sizes, strides, channels, pooling) so the result is reproducible.","section":"§4.5, Table 4.8"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is a PhD dissertation posted on arXiv and explicitly incorporates the author's prior publications (FDY, DFD, PFD, MDFD, TFD). The editorial question of whether the incremental novelty is sufficient for a journal submission is worth considering, but my recommendation is driven primarily by the missing capacity-controlled comparison and the absence of variance estimates, both of which are fixable within the manuscript's scope. The reviewer should also verify that the reported gains in Tables 4.6 and 4.7 come from the same reproducible pipeline as the rest of the paper, since several rows cite previous papers rather than newly run experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This dissertation packages the author's own prior work on frequency dynamic convolutions (FDY, DFD, PFD, MDFD, TFD) with a systematic set of ablations and three industrial case studies. The genuinely new pieces are the careful comparisons of dilation schedules, dynamic-branch proportions, and TAP pooling components on DESED, plus a 1D-vs-2D CRNN test showing the 1D encoder collapses on SED in this setting. Those ablations are useful for anyone designing audio front-ends.\n\nThe soft spot is the headline claim. Table 4.2 credits MDFD with a 10.98% PSDS1 gain over the baseline CRNN (0.455 vs 0.410), and TFD with the same score at 12.703M parameters. But the baseline CRNN has 4.428M parameters; MDFD has 18.157M, TFD has 12.703M. No matched-capacity static CRNN is trained, so the gain could be mostly capacity rather than frequency adaptivity. The PFD result (5.0-5.4M params, PSDS1 around 0.442) is encouraging, but even that uses 15-20% more parameters than the baseline, and a static CRNN at that size is untested. Hyperparameters for dilation, branch proportion, and pooling were also selected on the same DESED real validation set used for the final comparison, with no repeated-seed variance. The paper has internal numeric inconsistencies: baseline PSDS1 is 0.396 in Table 3.1 but 0.410 in Table 4.2, and PFD's parameter count appears as 5.401M in Table 3.5 but 5.041M in Table 4.2. These are addressable but make the central improvement not yet established.\n\nNothing here is fatal. The method family is coherent, the external benchmark grounds the evaluation, and the case studies indicate real-world motivation. The statistical analysis in Section 4.6 is truncated in the version I saw, so I cannot judge the class-wise claims. If the author adds matched-capacity static baselines, reports repeated-seed variance, and fixes the tables, this could be a solid archival contribution. As it stands, I'd treat the 10.98% as an upper bound, not a proven effect. I'd send it to peer review and ask for those changes before accepting.","headline":"Useful ablations around a known frequency-adaptive convolution family, but the headline PSDS1 gains are confounded with a 3-4x parameter increase and no matched-capacity static baseline.","tokens_in":48362,"tokens_out":3303,"would_cite":false,"duration_ms":32211,"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":"This paper claims that convolutions whose kernels adapt to the input's frequency content raise sound event detection on DESED by up to 10.98% over a baseline CRNN, and that a lighter variant matches the top score with 30% fewer parameters.","keywords":["sound event detection","frequency dynamic convolution","frequency-adaptive kernels","temporal attention pooling","dilated convolution","convolutional recurrent neural network","DESED dataset","PSDS"],"falsifier":"Train the final MDFD or TFD model with per-frequency attention weights replaced by one shared weight across all frequency bins, keeping parameter count and the training recipe identical, and compare PSDS1 on DESED's strongly labeled real test recordings; if the gain over the baseline CRNN does not shrink, the reported improvement is not due to frequency adaptivity. Replication of the 0.455 PSDS1 under the paper's exact augmentation and post-processing settings, with standard errors across seeds, would also settle whether the 10.98% figure lies outside evaluation noise.","tokens_in":47261,"feed_emoji":"🎧","tokens_out":11334,"duration_ms":103055,"temperature":0.7,"pith_summary":"Conventional 2D convolutions treat the frequency axis like the time axis, assuming that a pattern shifted upward or downward in pitch means the same thing. The paper argues this is wrong for audio, and sets out to show that letting convolutional kernels adapt to the input's frequency content improves sound event detection. It proposes Frequency Dynamic Convolution (FDY conv), in which the effective filter is a weighted combination of basis kernels with weights chosen per frequency bin from the input itself, plus variants: a dilated version (DFD), a hybrid static/dynamic version (PFD) that cuts parameters, a multi-branch dilated version (MDFD), and a temporal-attention-pooling version (TFD) that reaches the same top score with fewer parameters. On the DESED benchmark the best models raise the polyphonic sound detection score PSDS1 from 0.410 to 0.455, a 10.98% improvement, and class-wise analysis links each variant to a different event type (non-stationary, broad-spectrum, quasi-stationary, transient). If the gain is genuine, frequency-adaptive convolutions are a drop-in alternative to standard 2D convolutions for audio spectrogram processing.","feed_headline":"Adapting convolutions to frequency beats a standard CRNN by 10.98%","feed_subtitle":"Frequency-adaptive kernels raise the DESED score from 0.410 to 0.455; a lighter variant ties the top score.","key_machinery":"The load-bearing object is the frequency-adaptive attention weight vector $\\pi(f, x) \\in \\mathbb{R}^K$: a small learned subnetwork reads the input feature map, aggregates over time, and outputs, for every mel-frequency bin, $K$ weights that recombine a shared bank of $K$ basis kernels into a frequency-specific effective filter. Because the basis kernels are shared across frequencies, parameter growth stays modest; because the weights depend on the input, the same layer behaves differently for a broadband noise burst and a narrow tone. Each later variant changes how the basis kernels are organized (dilation per kernel, multiple dynamic branches, a static/dynamic channel split) or how time is aggregated (Temporal Attention Pooling combining time attention, velocity attention, and average pooling), but the input-dependent, per-frequency recombination of kernels is the mechanism that carries the whole argument.","core_discovery":"The central claim is that the shift invariance built into standard 2D convolution is the wrong inductive bias for the frequency axis of audio, and that replacing it with frequency-adaptive convolution measurably improves sound event detection. In FDY conv, the convolution output at frequency bin $f$ is the attention-weighted sum of $K$ basis-kernel responses, $$y(f) = \\sum_{i=1}^{K} \\pi_i(f, x)\\, (W_i * x + b_i),$$ where the weights $\\pi_i$ depend on both the frequency bin $f$ and the input $x$, so each frequency bin receives a different effective filter matched to the content there. Extending this idea, DFD conv diversifies basis kernels by dilation, PFD conv mixes a static branch with a small dynamic branch to cut parameters, MDFD conv combines several dilated dynamic branches, and TFD conv replaces temporal average pooling with attention-based pooling over time. The dissertation reports that MDFD achieves the highest PSDS1, 0.455 versus 0.410 for the baseline CRNN (a 10.98% gain), and that TFD matches 0.455 with 12.703M parameters versus 18.157M for MDFD.","pith_inferences":["A natural experiment not reported in the paper is to hold capacity constant while collapsing the per-frequency attention weights to a single shared weight; that ablation would isolate how much of the 10.98% gain is frequency adaptivity itself rather than extra parameters, the attention mechanism, or architecture selection.","The smooth, class-dependent clustering of attention weights along the frequency axis (the paper's PCA analysis) suggests FDY conv is learning an input-dependent filterbank, which connects this work to learnable front-end filter designs in speaker verification where the same mechanism could transfer.","The class-wise specialization across variants points toward a learned router or mixture-of-experts design in which the model picks the dilated, partial, or temporal-attention adapter per event type; the paper does not explore that combination.","Because the headline numbers come from a single validation set that also drove the architecture sweeps, multi-seed replication on a fresh test split is the prudent way to confirm the ordering of the variants before building on it."],"forward_implications":["CRNN-style sound event detection systems can be upgraded by swapping standard 2D convolutions for FDY-type layers, with reported PSDS1 gains of 7.56% for FDY conv and up to 10.98% for MDFD conv over the baseline on DESED.","The parameter-lean variants matter for deployment: PFD conv keeps roughly baseline-level accuracy while cutting parameters by 54.4% relative to FDY conv, and TFD conv matches the best PSDS1 at 12.703M parameters, about 30% fewer than MDFD conv.","Different event classes favor different designs: FDY helps non-stationary events, DFD helps broad-spectrum events, PFD helps quasi-stationary events, and TFD helps transient events, so architecture choice can be guided by the target sound inventory.","Combining the MDFD-CRNN with pretrained transformer encoders (ATST-frame plus BEATs) and change-detection-based event bounding raises the true PSDS1 to 0.577 without external pretraining data or ensembling.","The paper's controlled comparison of 2D and 1D front-ends (PSDS1 0.410 versus 0.192) indicates that preserving the frequency axis as a spatial dimension matters more than raw model capacity in this setting."],"supporting_citations":[{"why":"supplies the DESED dataset and the semi-supervised mean-teacher training recipe that all experiments share.","marker":"[2]"},{"why":"the CRNN baseline architecture whose convolution layers the proposed methods replace.","marker":"[3]"},{"why":"defines the PSDS evaluation metric used for every headline comparison.","marker":"[5]"},{"why":"the mean-teacher consistency method used to train on unlabeled DESED audio.","marker":"[42]"},{"why":"the original frequency dynamic convolution that this dissertation extends.","marker":"[107]"},{"why":"the dilated FDY variant that motivates DFD and MDFD conv.","marker":"[111]"},{"why":"the partial and multi-dilated FDY variants (PFD and MDFD) central to the main results.","marker":"[112]"},{"why":"the frequency-kernel and frequency-weighted convolutions used as earlier frequency-dependent baselines.","marker":"[115]"},{"why":"the temporal-attention-pooling variant (TFD conv) that matches the top score at lower parameter count.","marker":"[117]"}],"fun_headline_variants":["Frequency-adaptive conv nets lift sound detection by 10.98%","MDFD conv matches best SED score with 30% fewer parameters","Frequency-dynamic kernels beat baseline CRNN by ~11%","Sound event detection improved 10.98% with frequency-adaptive conv","Adaptive frequency filters boost SED: 10.98% gain, 30% lighter option"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes the performance gain is caused by frequency adaptivity, while its comparisons also add parameters, a learned attention mechanism, and hyperparameters selected on the same validation set used for the headline numbers.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-adaptive conv nets lift sound detection by 10.98%","MDFD conv matches best SED score with 30% fewer parameters","Frequency-dynamic kernels beat baseline CRNN by ~11%","Sound event detection improved 10.98% with frequency-adaptive conv","Adaptive frequency filters boost SED: 10.98% gain, 30% lighter option"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000557,"raw_usage":{"total_tokens":2938,"prompt_tokens":1523,"completion_tokens":1415,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":1139,"completion_tokens_details":{"reasoning_tokens":1314}},"tokens_in":1139,"tokens_out":1415,"duration_ms":11802,"temperature":1.0,"reasoning_tokens":1314,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:43:15.650607+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the final MDFD or TFD model with per-frequency attention weights replaced by one shared weight across all frequency bins, keeping parameter count and the training recipe identical, and compare PSDS1 on DESED's strongly labeled real test recordings; if the gain over the baseline CRNN does not shrink, the reported improvement is not due to frequency adaptivity. Replication of the 0.455 PSDS1 under the paper's exact augmentation and post-processing settings, with standard errors across seeds, would also settle whether the 10.98% figure lies outside evaluation noise.","supporting_citations":[{"cited_title":"Frequency Dynamic Convolution: Frequency- Adaptive Pattern Recognition for Sound Event Detection,","cited_arxiv_id":null,"evidence_quote":"the original frequency dynamic convolution that this dissertation extends."},{"cited_title":"Diversifying and Expanding Frequency- Adaptive Convolution Kernels for Sound Event Detection,","cited_arxiv_id":null,"evidence_quote":"the dilated FDY variant that motivates DFD and MDFD conv."}],"review_version":1}