{"id":"34d18f01-33e3-4f36-8c31-7f6ac61a14f8","arxiv_id":"2504.12670","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Temporal attention pooling, combining time attention, velocity attention, and average pooling, improves frequency dynamic convolution for sound event detection by about 3% average PSDS1 and reaches a maximum PSDS1 of 0.459 when combined with multi-dilated frequency dynamic convolution.","lead":"This paper replaces the temporal average pooling inside frequency dynamic convolution with an attention-based pooling that weights time frames, including a velocity branch for transients, and reports small but consistent gains on the DESED sound event detection benchmark. The main reasons to read it are the proposed three-branch pooling mechanism and the caution that the headline state-of-the-art number is a maximum over runs, not the average.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Velocity attention branch as written does not use temporal differences: Eq. (7) computes β from x, not from Δx defined in Eq. (6).","rationale":"The reader's conditional verdict is already appropriate, but the strongest reason is more specific than the reader's stated weakest assumption. The reader flagged the boundary case and the unnormalized sum in Eq. (3), yet the more serious problem is that Eq. (7) contradicts the textual claim of velocity attention: it uses x, not Δx. This is an internal inconsistency, not merely a missing detail. If the released code matches Eq. (7), then the paper's mechanistic narrative—that VA captures transients through temporal differences—is not supported by the implemented model, and the ablation attributes a gain to a mechanism that is not present. The empirical improvement from the extra attention branch may still be real, but the explanation would need revision. Because this can be settled by reading the code or by a targeted ablation, the paper should remain conditional rather than be accepted as-is. The reader's other concerns (no standard deviations, max-over-runs in Table VIII, missing external baselines) also support conditional status but are less sharply diagnostic than the Eq. (6)/Eq. (7) mismatch. I therefore agree with the conditional verdict while differing on the precise location of the load-bearing weakness.","tokens_in":16082,"tokens_out":3954,"duration_ms":41907,"concrete_test":"Inspect the official GitHub implementation's TAP module. Check the forward pass of the VA branch: is the input to the first 2D convolution (Wva1) the original tensor x, the difference Δx, or a concatenation? If it is x (or if Δx is computed but never used), then Eq. (6) is not part of the model. As a stronger test, modify the VA branch to actually take Δx (with zero padding at t=1) and retrain the 'FDY w/ VA' configuration; if PSDS1 does not change or degrades, the paper's stated velocity mechanism is not what drives the reported gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is TAP, and its transient-capture claim rests specifically on velocity attention pooling (VA) 'based on temporal differences.' However, Eq. (6) defines Δx_t = x_t − x_{t−1}, but Eq. (7) computes β = softmax(Wva2∗ReLU(BN(Wva1∗x + bva1)) + bva2), i.e., the input is the original feature x, not Δx. No redefinition of x appears between Eq. (1) and Eq. (7). Thus, as written, the velocity branch does not use temporal differences at all; Eq. (6) is inert. The boundary issue for t=1 is a symptom of this: if Δx were actually used, a boundary convention would be needed, but the text never supplies one. Consequently, the ablation result 'FDY w/ VA' (0.440 in Table II) demonstrates only that an additional attention branch with independently learned parameters helps, not that difference-based transient attention helps. This undermines the mechanistic interpretation of the 3.02% PSDS1 improvement and the claim that TAP specifically improves transient-event detection via velocity cues. The issue is directly checkable because the authors release code (footnote 1).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes temporal attention pooling frequency dynamic convolution (TFD conv), which replaces the temporal average pooling inside frequency dynamic convolution (FDY conv) with a temporal attention pooling (TAP) module. TAP consists of three branches: time attention pooling (TA), velocity attention pooling (VA), and conventional average pooling. Experiments on the DESED dataset with a CRNN backbone and mean-teacher semi-supervised training report an average PSDS1 improvement from 0.431 (FDY) to 0.444 (TFD), and a maximum PSDS1 of 0.459 for a combined TAP+MDFD configuration, which the authors describe as surpassing prior state-of-the-art SED systems. Classwise ANOVA and Tukey HSD analyses are used to support the claim that the gains concentrate on transient-heavy event classes.","tokens_in":16394,"tokens_out":8256,"duration_ms":80354,"significance":"If the claims hold, TAP would be a simple, drop-in modification to FDY-style SED models with modest parameter overhead and released code, which is valuable for reproducibility and for the DCASE-oriented community. However, the manuscript currently contains a load-bearing inconsistency in the definition of the velocity attention branch, an internal mismatch between the equations and the model that was actually ablated, no variance information for the reported average scores, and a state-of-the-art claim based on maximum scores over runs and configurations without external comparison. These issues must be resolved before the contribution and its mechanistic interpretation can be accepted with confidence.","major_comments":[{"comment":"Equation (6) defines Δx_t = x_t − x_{t−1}, but Equation (7) computes β from x, not from Δx, and Δx never appears again. As written, the velocity attention branch does not use temporal differences at all; Eq. (6) is inert. Consequently, the Table II result 'FDY w/ VA' (0.440) demonstrates only that an additional attention branch with independently learned parameters helps, not that difference-based transient attention helps. This undermines the central mechanistic claim that TAP improves transient-event detection through velocity cues. Please either amend Eq. (7) to consume Δx, with an explicit boundary convention for t=1, or revise the description, figure, and interpretation of the VA branch to match what is actually computed.","section":"Section III-B2, Eqs. (6)-(8)"},{"comment":"Equations (4) and (7) compute α and β from the raw input x, but Section V-A states that 'both time attention and velocity attention branches use xs as input to compute attention weights,' and Table I shows that the reported model uses xs (0.439 vs. 0.434 for TA with x). The equations therefore do not describe the model whose results are reported. Please align the equations with the actual implementation, including the path through xs, so that the method is reproducible from the text.","section":"Section III-B, Eqs. (4) and (7); Section V-A"},{"comment":"All average PSDS1 values are means over 12 runs with no standard deviations or confidence intervals, so the 3.02% improvement (0.444 vs. 0.431) cannot be judged for statistical significance. The 'state-of-the-art' claim (Abstract and Section V-H) is based on the maximum over runs and over the TAP+MDFD configurations in Table VII, and Table VIII compares only the authors' own FDY variants, not published SOTA systems on DESED (e.g., transformer-based systems cited in [44]-[49]). Please report variance, use a consistent statistic for performance claims, and add an external comparison on the same evaluation protocol.","section":"Section IV-G and Tables III, VIII"},{"comment":"The softmax axis in Eqs. (4) and (7) is not specified. Because x_TAP is formed by summing α_t⊙x_s,t and β_t⊙x_s,t over t and adding a time average, the normalization of α and β determines the scale of the pooled feature and the relative weight of the three branches. Please state the softmax dimension and any normalization applied to the sum of the three terms; otherwise the equations are ambiguous and the architecture cannot be reproduced exactly.","section":"Section III-B, Eqs. (3)-(5)"},{"comment":"The ANOVA/Tukey HSD results are reported only as qualitative inequality chains (e.g., 'DFD < TFD'). No test statistics, p-values, degrees of freedom, or the number of runs entering the analysis are given. As presented, the claim that TFD conv 'significantly enhances' transient-heavy classes is not substantiated. Please include the full statistical output or remove the significance language.","section":"Section V-D, Table IV"}],"minor_comments":[{"comment":"Typo: 'thaft' in 'indicating thaft while TA is useful' should be 'that'.","section":"Section V-B"},{"comment":"Inconsistent spacing appears in 'ANOV A', 'V A', and 'TAP + MDFD convs' (Section IV-A); please use consistent notation.","section":"Abstract and throughout"},{"comment":"The caption reads '(b)velocity'; a space is missing after '(b)'.","section":"Fig. 1 caption"},{"comment":"Section IV-G says all PSDS1 scores in the tables are averages over twelve runs, but Table VIII reports 'Max PSDS1'. Please clarify which tables report averages and which report maxima.","section":"Section IV-G and Table VIII"},{"comment":"Reference [56] appears to duplicate reference [47] (same title and venue); please merge or differentiate them.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper builds almost entirely on the authors' own prior FDY conv variants (FDY, DFD, PFD, MDFD). This is legitimate for controlled comparisons, but it makes the absence of external SED baselines more conspicuous: the 'state-of-the-art' claim in the abstract is not supported by the provided tables. I would ask the authors to either add a comparison with published DESED systems under equivalent evaluation conditions or to restrict the claim to 'state-of-the-art among FDY-based models'."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a real, if incremental, empirical contribution—replacing the temporal average pooling in FDY conv with an attention-weighted sum of three pooling branches gives a consistent gain on DESED PSDS1—but the velocity-attention mechanism, which is the paper's main hook, is mis-specified in the equations, and the state-of-the-art claim is built on best-of-run and best-of-configuration numbers rather than the reported averages.\n\nWhat's new: TAP combines time attention, a second attention branch labelled 'velocity', and plain average pooling, then feeds the sum into FDY conv's kernel generation. The ablations are reasonably thorough, the integration with DFD, PFD, and MDFD is explored, and the code is public. The average improvement from 0.431 to 0.444 over twelve runs is internally consistent, and the ANOVA/Tukey analysis for transient-heavy classes is a nice addition.\n\nThe problems are real. First, Eq. (6) defines Δx_t = x_t − x_{t-1}, but Eq. (7) computes the velocity attention weights β from x, not from Δx. No redefinition of x appears between the two equations. As written, the velocity branch is just a second attention branch over the same features; it does not use temporal differences. The paper's interpretation of the VA ablation as evidence for difference-based transient capture is therefore unsupported by the manuscript. This is checkable in the released code, but as it stands the central mechanism is mis-specified, and the boundary case t=1 is a symptom that Δx is never actually consumed.\n\nSecond, the reporting. Section IV.G says all PSDS1 scores are averages over twelve independent runs, but Table VIII switches to maximum scores, and the abstract highlights the maximums (0.456 and 0.459) as state-of-the-art. No standard deviations are given anywhere, and the best TAP+MDFD number comes from a search over 14 configurations. The gain over MDFD's average 0.444 is within what you'd expect from selection noise. Also, the comparison table only includes the authors' own FDY-family models; no external non-FDY systems are listed, so 'surpassing previous state-of-the-art SED systems' is not demonstrated.\n\nThese are fixable. If the code actually implements Δx, the equations need to be corrected. If not, the VA branch should be renamed and re-motivated, or the paper should show that the difference input is what causes the gain. Either way, the SOTA claim should be based on the average with variance, not the max.\n\nBottom line: the paper is worth refereeing because it ships code and the pooling idea is reusable, but the current version overstates what is shown. I'd send it out, with a request for a careful check of the equations and the reporting.","headline":"Useful incremental extension of FDY conv with a likely real gain on DESED, but the velocity-attention branch is mis-specified in the equations and the SOTA claim rests on best-of-run/best-of-configuration numbers.","tokens_in":16840,"tokens_out":3860,"would_cite":false,"duration_ms":38169,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Replacing temporal average pooling with time-plus-velocity attention improves sound event detection by 3% and sets a new state of the art.","keywords":["sound event detection","temporal attention pooling","frequency dynamic convolution","velocity attention pooling","PSDS1","DESED dataset","transient sound events","semi-supervised learning"],"falsifier":"Train TFD conv with the velocity branch replaced by a boundary-handled first difference (e.g., $\\Delta x_1 = x_2 - x_1$) and the three pooling outputs renormalized; if average PSDS1 on transient-heavy classes stays at or above 0.444, the specific unnormalized three-branch design is not what carries the gain. If the score instead drops below the FDY conv baseline of 0.431, the velocity branch and its scaling are load-bearing.","tokens_in":15892,"feed_emoji":"🔊","tokens_out":6075,"duration_ms":38027,"temperature":0.7,"pith_summary":"The paper claims that the temporal average pooling used inside frequency dynamic convolution (FDY conv) for sound event detection is a bottleneck: it weights every frame equally and dilutes short, transient sounds such as alarms, door knocks, and speech plosives. To fix this, the authors propose temporal attention pooling (TAP), which sums three pooling branches: a saliency-weighted time attention branch, a velocity attention branch that weights frames by their temporal differences, and the original average pooling for stationary sounds. Replacing only the pooling stage raises the average polyphonic sound detection score (PSDS1) from 0.431 to 0.444 over FDY conv with a 14.8% parameter increase, and the best TAP-plus-multi-dilated configuration reaches a maximum PSDS1 of 0.459, which the paper reports as surpassing previous state-of-the-art systems on the DESED benchmark. A reader should care because the change is a drop-in swap of the pooling operation, not a new architecture, and it specifically targets the transient events that average pooling systematically under-detects.","feed_headline":"A smarter pooling step lifts sound event detection past prior best","feed_subtitle":"Swapping average pooling for time-plus-velocity attention in FDY conv raises PSDS1 from 0.431 to 0.444.","key_machinery":"The machinery is temporal attention pooling (TAP), the three-branch pooling operator defined by $x_{\\mathrm{TAP}} = \\sum_{t=1}^{T}\\alpha_t \\odot x_{s,t} + \\sum_{t=1}^{T}\\beta_t \\odot x_{s,t} + \\frac{1}{T}\\sum_{t=1}^{T}x_t$. The first term is time attention pooling: softmax attention weights $\\alpha_t$ computed from 2D convolutions over the input, applied to a saliency-enhanced map $x_s$ obtained by two 2D convolutions with a sigmoid gate. The second term is velocity attention pooling: the same structure but with weights $\\beta_t$ computed from the first-order temporal difference $\\Delta x_t = x_t - x_{t-1}$, so frames with rapid spectral change receive higher weight. The third term is the original temporal average pooling on the raw feature, kept for robustness to stationary signals. The three outputs are summed without normalization, and the result replaces the average-pooled feature that FDY conv uses to generate frequency-adaptive kernels. The additional parameters come from the two saliency convolutions and the attention-weight convolutions, about 14.8% over FDY conv.","core_discovery":"The central claim is that the temporal aggregation step inside FDY conv—not the frequency-adaptive kernels themselves—limits detection of transient events, and that replacing temporal average pooling with TAP resolves this. TAP computes the pooled feature as the sum of three terms: time attention pooling, where softmax weights $\\alpha_t$ are applied to a saliency-enhanced representation $x_s$ obtained from two 2D convolutions; velocity attention pooling, where softmax weights $\\beta_t$ are derived from first-order temporal differences $\\Delta x_t = x_t - x_{t-1}$; and plain average pooling of the raw input. In ablation, all three branches together give the best result (0.444 average PSDS1), with velocity attention alone slightly outperforming time attention alone (0.440 vs 0.439), and the three-way sum behaving better than any pair. The paper further claims that the same pooling replacement transfers across FDY variants—dilated, partial, and multi-dilated—with the best configuration, TAP plus MDFD conv, reaching 0.459 maximum PSDS1 and outperforming the baseline CRNN by a wide margin.","pith_inferences":["The unnormalized sum of three pooling branches leaves the overall scale of $x_{\\mathrm{TAP}}$ dependent on attention-weight magnitudes; one testable extension is whether normalizing the sum (e.g., dividing by the sum of $\\alpha$ and $\\beta$ plus 1) changes stability across configurations or datasets.","Velocity attention is a first-order finite difference; second-order or multi-scale difference operators could capture acceleration of spectral change (e.g., frequency sweeps) and might extend the transient benefit beyond the classes tested here.","Because TAP is a generic pooling layer, the same recipe could be dropped into other average-pooling-dependent modules in audio tasks where transient cues matter, such as bioacoustic detection or audio tagging.","The paper reports only PSDS1; since PSDS2 emphasizes presence/absence rather than temporal boundaries, TAP's effect on audio-tagging-style evaluation remains untested and could differ."],"forward_implications":["A drop-in swap of the pooling stage in FDY conv yields a 3.02% average PSDS1 improvement (0.444 vs 0.431) with only 14.8% more parameters, so temporal pooling is a cheap place to gain sound event detection performance.","TAP generalizes across FDY variants: it improves or matches PFD conv at reduced dynamic-channel ratios, and its best combination with MDFD conv reaches 0.459 maximum PSDS1, the highest reported in the paper.","Classwise statistical analysis (ANOVA with Tukey HSD) shows TFD conv significantly beats other models on transient-heavy classes such as alarm/bell, cat, dog, and electric shaver, while keeping up on quasi-stationary classes—except vacuum cleaner, where the plain CRNN remains best.","Dilation in the frequency dimension does not help PSDS1 when combined with TAP; all dilated TAP+DFD variants score below the non-dilated TFD conv, suggesting that aggressive spectral receptive-field expansion can smooth away the transient cues TAP preserves."],"supporting_citations":[{"why":"Introduces frequency dynamic convolution, the base architecture whose temporal average pooling this paper replaces.","marker":"[37]"},{"why":"Provides the DESED dataset and the DCASE Task 4 training and evaluation framework used in all experiments.","marker":"[2]"},{"why":"Defines the PSDS metrics, particularly PSDS1, on which all headline numbers are reported.","marker":"[5]"},{"why":"Proposes DFD conv, the dilated FDY variant that TAP is integrated with and compared against.","marker":"[42]"},{"why":"Proposes PFD and MDFD conv, the partial and multi-dilated variants used for TAP integration and state-of-the-art comparison.","marker":"[43]"},{"why":"Supplies the mean-teacher semi-supervised training method that the model uses.","marker":"[57]"}],"fun_headline_variants":["Temporal attention pooling sharpens sound event detection","Attention pooling beats average pooling for transient sounds","New pooling method improves sound event detection","Temporal attention pooling lifts sound event detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that first-order temporal differences, with no boundary handling specified for the first frame, capture the transient salience that average pooling misses, and that simply summing the three pooling outputs yields a properly scaled feature for kernel generation.","fun_headline_variants_meta":{"raw":{"variants":["Temporal attention pooling sharpens sound event detection","Attention pooling beats average pooling for transient sounds","New pooling method improves sound event detection","Temporal attention pooling lifts sound event detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000548,"raw_usage":{"total_tokens":2704,"prompt_tokens":1117,"completion_tokens":1587,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":733,"completion_tokens_details":{"reasoning_tokens":1532}},"tokens_in":733,"tokens_out":1587,"duration_ms":12253,"temperature":1.0,"reasoning_tokens":1532,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:25:23.442448+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train TFD conv with the velocity branch replaced by a boundary-handled first difference (e.g., $\\Delta x_1 = x_2 - x_1$) and the three pooling outputs renormalized; if average PSDS1 on transient-heavy classes stays at or above 0.444, the specific unnormalized three-branch design is not what carries the gain. If the score instead drops below the FDY conv baseline of 0.431, the velocity branch and its scaling are load-bearing.","supporting_citations":[{"cited_title":"Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results,","cited_arxiv_id":null,"evidence_quote":"Supplies the mean-teacher semi-supervised training method that the model uses."}],"review_version":1}