{"id":"cea1c1e3-ea93-4698-aab9-1f643e924f52","arxiv_id":"2411.12830","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"An incremental learning method with MSE distillation lets a SELD model add four new sound classes after eight while roughly matching the performance of a model trained on all twelve at once.","lead":"This paper tests whether a sound-event localization and detection (SELD) system can learn four new sound classes after an initial eight, using a distillation loss to keep the old classes from being forgotten. The authors report that their incremental method matches a model trained on all twelve classes at once, on the TAU-NIGENS 2021 spatial sound scenes dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim relies on an unmatched baseline: DCASE 2022 baseline vs DCASE 2021 architecture of CIL model, so 'maintains baseline performance' is unverified.","rationale":"The reader's weakest_assumption identifies the most load-bearing concern: the baseline used for the central claim is not matched to the CIL model's architecture. The CIL model is built on the DCASE 2021 baseline, while the reference baseline is from DCASE 2022. Without a same-architecture joint-trained baseline, the quantitative equivalence claimed in the abstract and Table I cannot be attributed to the incremental method. This is not a minor methodological detail; it determines whether the paper demonstrates what it claims. The concern is concrete: a matched baseline would either support the claim (if CIL performs comparably) or refute it (if the matched baseline performs better). The paper's internal evidence that distillation prevents catastrophic forgetting (the contrast with FT and IndL, and the class-wise F1 in Table II) is convincing qualitatively, but the central quantitative claim of 'maintaining baseline performance' is unverified. I agree with the reader's CONDITIONAL verdict: the method shows promise, but the evidence as presented is insufficient for full acceptance. The proposed test is the minimal additional experiment needed to settle the concern.","tokens_in":7427,"tokens_out":8397,"duration_ms":76366,"concrete_test":"Train the exact DCASE 2021 CRNN architecture used for the CIL model (F0/H0) on all 12 classes from scratch, using the same 400/100/100 split, optimizer, epochs, and evaluation metrics. Compare its overall and class-wise F1, ER, LE, LR (especially old classes 1–8) to the CIL-MSE row in Tables I–II. If the matched baseline's metrics are within a few points of CIL-MSE, the claim holds; if the matched baseline is substantially better (e.g., F1 > 28 or old-class F1 > 25), the comparison to the DCASE 2022 baseline is misleading and the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that CIL-MSE (λ=0.5) 'successfully maintains baseline performance across all metrics' (Table I: F1 23.1, ER 0.71, LE 26.6, LR 40.9 vs baseline 23.1, 0.70, 39.4, 41.2). This comparison is not controlled: Section II states the CIL model uses a CRNN based on the DCASE 2021 baseline [24], while Section III.C defines the baseline as the DCASE 2022 Challenge baseline [25]. No matched baseline with the same architecture trained on all 12 classes is reported. Therefore, the observed similarity—and the large LE improvement of 12.8°—could be due to architectural differences (feature extraction, network size, training scheme) rather than the incremental method. The claim 'maintains baseline performance' is only meaningful relative to a baseline that would be obtained by the same model trained on all classes jointly. Additionally, Table II reports old-class F1 for CIL-MSE (19.7) but no corresponding baseline old-class F1, so the central claim about retaining old knowledge cannot be evaluated. This is the load-bearing weakness because it directly affects the interpretation of every quantitative result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage class-incremental learning (CIL) framework for sound event localization and detection (SELD). In Stage 0, a CRNN based on the DCASE 2021 SELD baseline is trained on 8 sound classes using the ACCDOA representation. In Stage 1, the classifier head is expanded to 12 classes and the model is trained on data for the 4 new classes while an output distillation loss (MSE) penalizes discrepancies between the old and updated models' outputs for the original 8 classes. Experiments on TAU-NIGENS Spatial Sound Events 2021 report that the proposed CIL-MSE (lambda=0.5) approximately matches a DCASE 2022 Challenge SELD baseline on overall F1, ER, LE, and LR, and outperforms fine-tuning and independent learning, with additional ablations on distillation loss type and lambda.","tokens_in":7693,"tokens_out":2445,"duration_ms":26228,"significance":"If the results are substantiated, this is the first demonstration of class-incremental learning for the joint detection and localization of sound events, which is a meaningful step for adaptive SELD systems that need to add classes without full retraining. The paper also includes useful ablations (MSE vs KL distillation, lambda sweep, class-wise analysis, and analysis of regression weight norms) that go beyond a single accuracy number. However, the central claim that CIL 'maintains baseline performance' is currently supported only by an unmatched baseline comparison, and the choice of lambda=0.5 is not adequately justified. These issues are load-bearing and must be addressed before the claim can be accepted.","major_comments":[{"comment":"The comparison that supports the central claim is not controlled. The proposed CIL model is built on the DCASE 2021 baseline CRNN architecture [24], while the 'Baseline' in Table I is the DCASE 2022 Challenge SELD baseline [25]. These are different architectures, feature pipelines, and training configurations. The paper never trains the same architecture used for CIL on all 12 classes jointly and reports its performance. Therefore the observation that CIL-MSE is close to the DCASE 2022 baseline (and even improves LE by 12.8 degrees) could be due to architectural differences rather than the incremental method. To support the claim that CIL 'maintains baseline performance', the authors must supply a matched baseline: the same CRNN/feature extractor trained jointly on all 12 classes, with the same optimizer, epochs, and evaluation protocol.","section":"Section II vs Section III.C, Table I"},{"comment":"The choice of lambda=0.5 is not justified with respect to data leakage. Figure 3 shows average F1 over all 12 classes as a function of lambda, and the paper selects lambda=0.5 for the reported CIL-MSE results. It is not stated whether this sweep was performed on the validation split or on the same test split used in Tables I and II. If the sweep and the final evaluation use the same test partition, the reported numbers are the result of tuning on the test set, which overstates the method's expected performance. Please clarify the data split used for Figure 3 and, preferably, report results for a fixed lambda chosen on validation only. In addition, no error bars or multiple-seed runs are reported; given the small metric differences involved, this is needed to establish that the reported ordering is not noise.","section":"Section IV.B, Figure 3"},{"comment":"The paper's claim that CIL preserves old-class knowledge cannot be evaluated from the presented tables. Table II reports old-class (1-8) and new-class (9-12) F1 for FT, IndL, CIL-MSE, and CIL-KLD, but it does not include the corresponding old-class and new-class breakdown for the baseline model. Without the baseline's old-class F1, the reader cannot tell whether CIL-MSE's old-class F1 of 19.7 represents retention comparable to joint training or a substantial drop. Please add the same-architecture baseline's old-class and new-class F1 scores (and ideally the per-class values) to Table II.","section":"Table II"}],"minor_comments":[{"comment":"Reference [25] is cited as the source of the DCASE 2022 SELD baseline, but [25] is the STARSS22 dataset paper; the actual DCASE 2022 baseline system should be cited explicitly, or the reference should be corrected.","section":"Section III.C"},{"comment":"The distillation loss is denoted inconsistently: the text and equation use LOD with a subscript, while the figure uses what appears to be a calligraphic L_OD. Please unify the notation throughout.","section":"Equation (1) and Figure 1"},{"comment":"The statement that LE bars are omitted for classes with zero F1 'as LE for these cases is 180 degrees' is confusing, since localization error is normally undefined or set to a penalty value when no true positive exists. Please specify the exact convention used in the evaluation code.","section":"Section IV.A"},{"comment":"The phrase 'probably due to its MSE-based distillation loss which directly minimizes discrepancies in logits' is speculative; either remove 'probably' or provide supporting evidence (e.g., an analysis of output distributions or per-class errors).","section":"Section IV, paragraph after Table I"},{"comment":"The heading 'I. I N TRODUCTION' contains an extra space in 'INTRODUCTION'; this is a minor typographical issue.","section":"Section I heading"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and underexplored problem (CIL for joint SED and DOA), and the central idea is sensible. However, the load-bearing comparison uses an unmatched baseline, and the lambda selection procedure risks test-set tuning. Both are fixable within the manuscript's scope: the authors should add a matched same-architecture baseline and clarify the validation protocol. I recommend major revision rather than rejection, provided these issues are resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first class-incremental learning result for sound event localization and detection, and the core qualitative effect—output distillation stops catastrophic forgetting—is visible in the numbers. But the headline claim that CIL-MSE 'maintains baseline performance' is not actually supported, because the baseline is the DCASE 2022 system [25] while the incremental model uses the DCASE 2021 CRNN [24]. No matched baseline trained on all 12 classes with the same architecture is reported, so the comparison in Table I confounds method with architecture. That is the load-bearing weakness.\n\nWhat is new: applying the authors' own distillation-based CIL approach [19] to ACCDOA-based SELD, replacing classification loss with MSE, and ablating MSE vs KLD distillation. The finding that CIL-MSE retains old-class F1 (19.7) while FT/IndL drop to ~0 is convincing and consistent with the weight-norm analysis in Figure 4. The single-stage extension is clearly described, and the evaluation on TAU-NIGENS 2021 with standard SELD metrics is appropriate. Credit where due: this is a reasonable first step for a practical problem.\n\nSoft spots, in order. (1) Unmatched baseline, as above. The 12.8° LE improvement over the DCASE 2022 baseline is more likely an architecture artifact than a distillation benefit; without a same-architecture joint-training baseline, 'maintains' is unverified. (2) No error bars or repeated-seed variance anywhere; with F1 around 20 and one run per condition, the differences between CIL-MSE and CIL-KLD (23.1 vs 20.0) could easily be noise. (3) The choice of λ=0.5 is justified by Figure 3, which appears to be the same test partition used for reporting; if λ was selected on the test set, the reported numbers are optimistic. (4) Table II gives no baseline old-class F1, so the retention claim cannot be compared to the joint model.\n\nThe math and citation pattern are fine—self-citation to [19],[20] is fair since the method genuinely builds on those. No circularity. The paper ships no code, but the dataset and metrics are standard.\n\nWho this is for: researchers in audio continual learning and SELD; the paper is a solid workshop-to-ICASSP-level contribution. It deserves a serious referee, but the referee should ask for a matched baseline and variance estimates. My recommendation: send to peer review, require the matched baseline and at least repeated runs for the main conditions.","headline":"First CIL-SELD paper with a real effect, but the headline comparison to 'baseline' is not controlled and needs a matched baseline before the central claim holds.","tokens_in":8204,"tokens_out":1866,"would_cite":true,"duration_ms":17830,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An MSE distillation loss lets a sound-event localization and detection model add four new sound classes while keeping baseline-level detection and improving localization.","keywords":["class-incremental learning","sound event localization and detection","catastrophic forgetting","output distillation","ACCDOA","TAU-NIGENS","continual learning","polyphonic audio"],"falsifier":"Train the same architecture that the incremental model uses (the DCASE 2021 CRNN) jointly on all 12 classes with the same hyperparameters and evaluate it on the same test split; if that model clearly outperforms CIL-MSE in F1 or error rate, the paper's claim of matching baseline performance would no longer stand.","tokens_in":7224,"feed_emoji":"🎧","tokens_out":6513,"duration_ms":56421,"temperature":0.7,"pith_summary":"This paper asks whether a sound event localization and detection (SELD) model can be taught four new sound classes after deployment without forgetting the eight classes it already knows, and without retraining on old-class recordings. The proposed two-stage method, CIL-MSE, keeps the first-stage model's outputs for the original classes as a distillation target while the expanded model learns the new classes from new data. On the TAU-NIGENS Spatial Sound Events 2021 dataset, the method matches the reference baseline's overall F1-score and error rate, and even improves localization error. The result matters because real-world audio systems are often deployed with a fixed class set and later need to grow; if this holds, such growth can be done cheaply without catastrophic forgetting.","feed_headline":"Distillation lets sound-event models add 4 classes and match baseline","feed_subtitle":"An 8-to-12 class incremental step on spatial audio keeps old-class F1 and cuts localization error.","key_machinery":"The load-bearing mechanism is the MSE-based output distillation loss, adapted from an earlier class-incremental learning method for multi-label audio classification [19]. At stage 1, the frozen stage-0 model's predictions for the original 8 classes serve as soft targets; the active model is trained so that its outputs for those classes remain close to the frozen outputs, while a standard MSE loss supervises the 4 new classes. The two terms are combined as $L = (1-\\lambda) L_{\\mathrm{MSE}} + \\lambda L_{\\mathrm{OD}}$, with $\\lambda = 0.5$ in the main result. Because the regression head uses ACCDOA vectors, the distillation penalty acts directly on the joint detection-and-localization output space rather than on class probabilities.","core_discovery":"The paper's central claim is that output-distillation-based class-incremental learning can add four new sound classes to an eight-class SELD model while preserving the original classes' detection and localization performance. Using the ACCDOA representation, in which each active sound event is encoded as a three-dimensional Cartesian direction-of-arrival vector, the stage-1 model expands the final regression layer from $8\\times3$ to $12\\times3$ neurons. It is trained with a total loss that blends the mean-square-error loss for the new classes with a mean-square-error distillation loss on the old classes' outputs, $L = (1-\\lambda)L_{\\mathrm{MSE}} + \\lambda L_{\\mathrm{OD}}$. With the mixing weight $\\lambda=0.5$, the resulting CIL-MSE model achieves overall F1 of 23.1, error rate 0.71, localization error 26.6°, and localization recall 40.9%, compared to the reference baseline's 23.1, 0.70, 39.4°, and 41.2%. The authors conclude that CIL-SELD performs similarly to the baseline, showing improved localization for both existing and newly introduced classes without training on the entire dataset.","pith_inferences":["The comparison baseline is the DCASE 2022 SELD baseline, while the incremental model uses the DCASE 2021 backbone; a same-architecture jointly-trained 12-class model would make the 'maintains baseline' claim directly testable.","The observed localization-error improvement (26.6° vs 39.4°) may stem partly from the architecture difference rather than from incremental learning; a matched baseline would isolate the effect of distillation.","If the method is extended to many small incremental steps, distillation targets would be refreshed at each stage, but the accumulation of distillation error over multiple stages is an open question the paper does not address.","Because the distillation loss operates on continuous regression outputs, the same idea could transfer to other multi-output audio regression tasks such as direction-of-arrival estimation alone."],"forward_implications":["A deployed SELD system can absorb new sound classes in an incremental stage without access to labelled recordings of the old classes, cutting the data and compute cost of model updates.","Because the classifier head expands by a fixed block of neurons, the same two-stage recipe extends to multiple incremental steps, which the paper names as future work.","MSE distillation beats KL-divergence distillation on overall F1 in this setup, so the choice of distillation loss is a deciding factor in whether old-class knowledge survives.","The balanced L2 norms of the final-layer regression weights under CIL-MSE, in contrast to fine-tuning's skewed norms, indicate the model distributes capacity evenly across old and new classes."],"supporting_citations":[{"why":"Supplies the TAU-NIGENS dataset and the DCASE 2021 CRNN backbone that the incremental model is built on.","marker":"[24]"},{"why":"Provides the DCASE 2022 SELD baseline trained on all 12 classes that serves as the comparison point.","marker":"[25]"},{"why":"Defines the ACCDOA regression format that both detection and localization outputs are encoded in.","marker":"[6]"},{"why":"Introduces the output-distillation formulation for class-incremental audio classification that the paper adapts to MSE.","marker":"[19]"},{"why":"Defines the joint SELD evaluation metrics and the 20-degree threshold used in the tables.","marker":"[7]"},{"why":"Details the ER, F1, LE, and LR metrics and their computation over one-second frames.","marker":"[26]"}],"fun_headline_variants":["Incremental SELD adds 4 classes, keeps F1, cuts localization error by 12.8°","CIL-MSE lets SELD learn new sounds without forgetting old ones","Spatial audio model learns 4 new classes, still beats baseline on localization","Distillation-based CIL: SELD gains classes, keeps performance on TAU-NIGENS","Class-incremental sound event detection: add 4 classes, match baseline, better localizatio"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that CIL-MSE maintains baseline performance rests on comparing against a DCASE 2022 baseline built on a different architecture than the DCASE 2021 backbone of the incremental model, with no same-architecture 12-class model trained as a matched control.","fun_headline_variants_meta":{"raw":{"variants":["Incremental SELD adds 4 classes, keeps F1, cuts localization error by 12.8°","CIL-MSE lets SELD learn new sounds without forgetting old ones","Spatial audio model learns 4 new classes, still beats baseline on localization","Distillation-based CIL: SELD gains classes, keeps performance on TAU-NIGENS","Class-incremental sound event detection: add 4 classes, match baseline, better localization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000939,"raw_usage":{"total_tokens":4005,"prompt_tokens":925,"completion_tokens":3080,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":2964}},"tokens_in":541,"tokens_out":3080,"duration_ms":20553,"temperature":1.0,"reasoning_tokens":2964,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:08:37.505171+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same architecture that the incremental model uses (the DCASE 2021 CRNN) jointly on all 12 classes with the same hyperparameters and evaluate it on the same test split; if that model clearly outperforms CIL-MSE in F1 or error rate, the paper's claim of matching baseline performance would no longer stand.","supporting_citations":[{"cited_title":"Accdoa: Activity-coupled cartesian direction of arrival representation for sound event localization and detection,","cited_arxiv_id":null,"evidence_quote":"Defines the ACCDOA regression format that both detection and localization outputs are encoded in."},{"cited_title":"Class-incremental learning for multi- label audio classification,","cited_arxiv_id":null,"evidence_quote":"Introduces the output-distillation formulation for class-incremental audio classification that the paper adapts to MSE."},{"cited_title":"Joint measurement of localization and detection of sound events,","cited_arxiv_id":null,"evidence_quote":"Defines the joint SELD evaluation metrics and the 20-degree threshold used in the tables."}],"review_version":1}