{"id":"3874812c-6b4e-47c3-910d-b52e5e8cbe16","arxiv_id":"2506.18406","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A fully few-shot class-incremental audio classifier using a frozen multi-level AST embedding extractor and an incrementally updated ridge regression classifier outperforms prior methods on three datasets.","lead":"The authors introduce a system for classifying sounds from new classes using only a handful of training examples, combining an Audio Spectrogram Transformer with a ridge regression classifier. It reports higher average accuracy and less forgetting than earlier few-shot incremental audio methods on three public datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncontrolled baseline comparison: prior-method numbers may come from different class splits and evaluation protocols, so the claimed AA/PD superiority is not yet demonstrated.","rationale":"The reader's weakest_assumption lists both the transferability of frozen AST embeddings and the consistency of baseline protocols. I agree with the second as the more load-bearing concern. Transferability is a design assumption, but it is directly probed by the paper's own reported results: if the method achieves high base-session accuracy and maintains reasonable incremental accuracy on three datasets, then the embeddings do transfer for those splits (assuming the numbers are correct). The uncontrolled baseline comparison, however, is a threat to the central claim independent of the method's internal validity: even a perfect method cannot be said to exceed prior state of the art if the prior numbers come from different class subsets, different orders, or different evaluation rules. The paper provides no evidence that baselines were run under identical conditions, and the internal inconsistency in Section 3.1/Table 1 plus the anomalous session-9 jump in Table 4 make the protocol ambiguity concrete rather than hypothetical. The lack of error bars compounds this: the FSC-89 AA advantage over AISP is only 0.34 points, which could easily be split-level noise. I do not see a fatal flaw in the ridge-regression formulation itself; the G/C accumulation in Eqs. 6-9 is a standard closed-form update, and the ablation in Table 3 supports the fusion module and RRC contributions. The code availability is a positive, but it does not by itself verify the baseline comparison. The reader's CONDITIONAL verdict is appropriate; my concern is the same one that motivated that conditionality, so I recommend no change. The concrete test above would either substantiate the claimed improvements or force a re-evaluation of the headline.","tokens_in":10904,"tokens_out":4241,"duration_ms":47552,"concrete_test":"Obtain or reconstruct the exact 50-class selection, session order, and 5-shot training samples used by the authors, then re-run at least one baseline (e.g., AISP [39] or a simple frozen-AST plus ridge-regression variant) under the Section 3.2 protocol. If the reproduced AA/PD differs from Tables 4-6 by more than expected sampling variability (roughly >1 point given 100-test averaging), the superiority claim is not established. Additionally, verify whether the LS-100 session-9 value of 57.36 is reproducible; if it is a typo or arises from a different test set, recompute AA and PD and compare again.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: Tables 4-6 report that MEE+RRC beats nine prior methods in AA and PD on LS-100, NSynth-100, and FSC-89. But the comparison is not controlled. Section 3.1 states 'We randomly select samples of fifty classes from each one of the three datasets' yet Table 1 describes the full datasets (100/100/89 classes), and no random seed, class ordering, or session assignment is given. Section 3.4 quotes baseline results without stating whether those numbers were re-computed under this paper's exact split and evaluation protocol or copied from prior publications. Since the closest competitor AISP [39] is the authors' own prior work, it may have used a different 50-class subset or different incremental order; the reported AA differences (0.3-7.3 points) could then be artifacts of the protocol rather than the method. The absence of any error bars, despite Section 3.2 saying results are averaged over 100 tests, makes it impossible to assess whether even the within-paper differences are significant. A concrete anomaly supporting this concern is Table 4: LS-100 session-9 accuracy jumps from 46.93 (session 8) to 57.36, while AISP rises only from 47.90 to 48.60; a 10-point jump when adding one 5-class session is atypical and suggests either an evaluation-set inconsistency or a typo. Because the headline 'exceeds current methods' rests entirely on Tables 4-6, this uncontrolled comparison is the single most load-bearing risk.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses Fully Few-shot Class-incremental Audio Classification (FFCAC), where both base and incremental sessions provide only a few training samples per class. The proposed method decouples the model into a multi-level embedding extractor (MEE), consisting of an audio spectrogram Transformer encoder plus a trained fusion module, and a ridge regression classifier (RRC) updated by a closed-form least-squares solution. The MEE is finetuned in the base session and frozen in incremental sessions; the RRC accumulates Gram-matrix-like terms G and C and recomputes classifier weights after each session. Experiments on LS-100, NSynth-100, and FSC-89 are reported with average accuracy (AA) and performance degradation (PD) metrics. An ablation on NSynth-100 shows that both the fusion module and the RRC improve the results. The main claimed contribution is that the method exceeds nine prior methods in AA and PD on the three datasets while having middling complexity. The manuscript includes code and describes the training configuration in detail.","tokens_in":11159,"tokens_out":8027,"duration_ms":86467,"significance":"The paper's recipe — a frozen pretrained transformer with multi-level feature fusion and a non-parametric ridge-regression classifier — is simple, coherent, and, if the numbers hold, would be a strong baseline for FFCAC. The ablation study in Table 3 is informative and supports the contribution of each component, and the closed-form classifier update is a practical way to avoid repeated optimization in incremental sessions. The release of code is helpful. However, the empirical claim is currently not demonstrated: the comparison against prior methods is not shown to be protocol-controlled, no run-to-run variability is reported, and one session's accuracy in Table 4 is anomalous. These issues are load-bearing because the paper's headline is an empirical performance claim.","major_comments":[{"comment":"Table 1 lists the full datasets as 100/100/89 classes, while §3.1 states that fifty classes are randomly selected from each dataset; the paper never specifies the 50-class subset, the random seed, the class ordering, or the session assignment. Because the incremental order is known to affect few-shot continual learning results, the reported comparisons in Tables 4-6 cannot be reproduced or checked for protocol match, and this affects the paper's main claim.","section":"§3.1 and Table 1"},{"comment":"The baseline numbers are quoted without any statement that they were rerun under this paper's exact 50-class split, incremental order, and evaluation protocol. The closest baseline, AISP [39], is the authors' own prior work, and the earlier FFCAC method [38] is omitted from the comparison tables entirely. Without a controlled reimplementation or a detailed protocol-match statement, the reported AA advantages (0.34–7.32 points) may reflect split or ordering differences rather than method superiority.","section":"§3.4 and Tables 4–6"},{"comment":"Section 3.2 says that the final result is the average of 100 tests, but the paper reports no standard deviation, confidence interval, or per-seed values. Given the random 50-class selection and random session order, the FSC-89 AA gap of 0.34 points between Ours and AISP (33.27 vs. 32.93) is within plausible run-to-run variation, so the claim that the method 'exceeds all prior methods' is not statistically supported.","section":"§3.2 and Tables 4–6"},{"comment":"On LS-100, the proposed method's accuracy jumps from 46.93 in session 8 to 57.36 in session 9, while AISP changes only from 47.13 to 48.60; adding one 5-class session should not normally cause a 10.43-point jump in overall accuracy. This anomaly suggests an evaluation-set inconsistency or a typo, and because session 9 contributes about one percentage point to the reported AA, it needs to be explained or corrected.","section":"Table 4"},{"comment":"The AST encoder is initialized from a pretrained model and then frozen after the base session, but the paper does not state what data the pretrained model was trained on or whether its training classes overlap with the evaluation datasets. If the pretrained model has already seen the incremental classes, the 'fully few-shot' protocol is compromised; at minimum, the pretraining corpus and any class overlap should be reported before the few-shot claim can be interpreted.","section":"§2.2 and §2.3"}],"minor_comments":[{"comment":"Section 2.3 contains typos: 'characterisitcs' and 'incoporate' should be corrected.","section":"§2.3"},{"comment":"Equation (1) uses W_{y_k} before defining it; the text should clarify that W_{y_k} is the classifier weight vector for the class of the kth sample.","section":"Eq. (1)"},{"comment":"Section 2.4 says λ is obtained by cross-validation using training samples; the paper should state whether λ is tuned on the base session only or on all sessions and whether the same value is used throughout.","section":"§2.4"},{"comment":"The inference step uses cosine similarity between embeddings and classifier weights, but Eq. (5) computes weights with unnormalized embeddings; the paper should state whether embeddings and weights are normalized before the cosine operation.","section":"§2.4 and inference"},{"comment":"Table 7 reports different MACs and NP values for Ours across datasets, but the text does not explain that the input LMS sizes differ across datasets; a sentence noting this would help.","section":"Table 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is a candidate for acceptance if the authors supply protocol-matched baseline numbers, per-seed statistics or error bars, and a resolution of the LS-100 session-9 anomaly. Given that the closest baseline AISP and the omitted prior work [38] come from the same group, the editor should specifically ask for raw per-session logs or a shared evaluation harness."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper combines two established ingredients—multi-level feature fusion from a frozen AST and a ridge regression classifier—into a new system for fully few-shot class-incremental audio classification. The ablation is the real strength: it shows cleanly that both the fusion module and the ridge classifier contribute, and the improvement from case ① to case ④ on NSynth-100 is large enough to take seriously. The authors also release code and give a closed-form update rule that is simple and practical. That is a fair amount of substance for a conference paper.\n\nThe soft spots are the usual ones, but they land on the headline claim. The comparison tables list nine prior methods, but nothing indicates whether those numbers were recomputed under the same 50-class split and session order. The text says fifty classes are randomly selected; Table 1 then describes the full 100/100/89-class datasets. No seed, no class list, no session assignment. The closest competitor, AISP, is the authors' own prior work, and the reported AA advantages of 0.3 to 7.3 points could easily be protocol artifacts. The absence of error bars matters here because the text says results are averaged over 100 tests; that number should be accompanied by a standard deviation. Also, the LS-100 session-9 jump from 46.93 to 57.36 is anomalous and needs a check. Finally, RanPAC [44] is the obvious reference for ridge regression on pretrained embeddings in continual learning, and it is cited but never compared; the paper should at least explain the difference in this few-shot audio setting.\n\nNone of this is fatal to the method. The internal evidence is consistent, the transferability premise is reasonable, and the complexity numbers are plausible. But the central claim, that this method exceeds all prior methods, is not yet demonstrated to the standard the prose implies. A serious referee would need to ask for a unified baseline rerun, error bars, a corrected class-count statement, and a RanPAC comparison. That is a fixable checklist, not a dead end.\n\nI would send this to peer review. It is a competent, useful contribution for a niche but active community, and the flaws are empirical rigor rather than conceptual. A clean revision with controlled baselines would make it a solid paper. I would not cite it myself in the next year unless I were working directly on FFCAC, but I would bring it to a reading group as a cautionary example of why baseline protocol reporting matters.","headline":"A sensible incremental FFCAC system, but the state-of-the-art claim is not yet demonstrated because baseline numbers and splits are uncontrolled.","tokens_in":11728,"tokens_out":1214,"would_cite":false,"duration_ms":14164,"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":"Frozen multi-level embeddings and ridge regression outperform prior methods in fully few-shot continual audio classification.","keywords":["fully few-shot class-incremental learning","audio classification","Audio Spectrogram Transformer","multi-level embedding fusion","ridge regression classifier","continual learning","few-shot learning"],"falsifier":"Rerun the same 5-way, 5-shot, ten-session protocol on a fresh random selection of 50 classes from LS-100, NSynth-100, or FSC-89, with the same frozen-extractor and ridge-regression code; if average accuracy does not reproduce the reported margins over AISP (roughly +3.7, +7.3, and +0.3 points), the result depends on the particular class draw or evaluation protocol.","tokens_in":10668,"feed_emoji":"🔊","tokens_out":6432,"duration_ms":69032,"temperature":0.7,"pith_summary":"Fully few-shot class-incremental audio classification (FFCAC) asks a model to keep learning new sound classes when every class, including the base classes, has only a handful of training samples. The paper argues that this can be handled by decoupling the model into a multi-level embedding extractor and a ridge-regression classifier. The extractor, built on an Audio Spectrogram Transformer, is finetuned in the base session and then frozen; the classifier is updated in each incremental session by a closed-form ridge-regression solution. On LS-100, NSynth-100, and FSC-89, the method reports higher average accuracy and lower performance degradation than all compared prior methods, with the largest gain over the closest baseline on NSynth-100. If the result holds, it would show that a simple frozen-extractor recipe can replace more complex dynamically expanded classifiers for few-shot continual audio learning.","feed_headline":"Frozen audio embeddings beat prior few-shot continual methods","feed_subtitle":"On three public benchmarks, the ridge-regression classifier gains up to 7.3 points over the strongest baseline.","key_machinery":"The central objects are the Multi-level Embedding Extractor (MEE) and the Ridge Regression Classifier (RRC). MEE is an Audio Spectrogram Transformer whose block outputs are concatenated and reweighted by an MLP-plus-Softmax fusion module; this fuses local, concrete features from shallow blocks with global, abstract features from deep blocks into a single embedding. RRC computes classifier weights as $W=(E^T E+\\lambda I)^{-1}E^T Y$, maintaining the Gram matrix $G=E^T E$ and $C=E^T Y$ across sessions and updating them additively, which prevents catastrophic forgetting without retraining and reduces correlations between embedding dimensions to counter overfitting under few samples.","core_discovery":"On the paper's own terms, the central discovery is that a feature extractor pretrained on large-scale audio, lightly finetuned on only 25 base samples, and then frozen for all incremental sessions is enough to make a fully few-shot continual audio classifier state of the art. The embeddings come from fusing the features of several transformer blocks, and the classifier is a ridge-regression head whose weights are solved analytically, so each new session only adds to two accumulated matrices rather than retraining the network. Reported average accuracy is 58.95% on LS-100, 68.65% on NSynth-100, and 33.27% on FSC-89, exceeding the closest prior method AISP by about 3.7, 7.3, and 0.3 points, with lower performance degradation in every case.","pith_inferences":["Editorial inference: the recipe is not tied to AST specifically; if another pretrained audio encoder yields embeddings with similar or better transferability, the same ridge-regression update should carry the gain, which is directly testable by swapping the encoder.","Editorial inference: the success on only 25 base samples suggests the bottleneck in FFCAC is classifier calibration rather than feature learning, so pretrained encoders with stronger general-purpose audio representations could shift results further without architectural changes.","Editorial inference: the additive G and C updates assume class-disjoint sessions; a natural extension outside the paper is to test whether the same update remains stable when sessions overlap in label space or when classes reappear."],"forward_implications":["New sound classes can be added by updating two small matrices, so each incremental session is cheap in computation and memory relative to attention-based expansion modules.","The method's performance degradation over ten sessions is lower than that of every compared method, so old-class recognition degrades less as new classes arrive.","Fusion of multi-block features is a measurable ingredient: on NSynth-100, adding the fusion module to the ridge-regression classifier raises average accuracy from 65.31% to 68.65% in the paper's ablation.","Because the extractor is frozen after the base session, the method's complexity sits below several prior fully few-shot and few-shot continual methods while exceeding them in accuracy."],"supporting_citations":[{"why":"Establishes the FFCAC benchmark protocol, datasets, and the strongest baseline that the reported gains are measured against.","marker":"[39]"},{"why":"Supplies the pretrained Audio Spectrogram Transformer encoder that the MEE is initialized from and then freezes.","marker":"[40]"},{"why":"Supplies the closed-form regularized least-squares solution used to derive the classifier update.","marker":"[46]"},{"why":"Provides a prior fully few-shot continual audio method with an expandable dual-embedding extractor that this work builds on and compares with.","marker":"[38]"},{"why":"Contributes the prototype-based classifier used in the ablation and one of the compared prior methods.","marker":"[33]"},{"why":"Supplies a compared few-shot class-incremental audio method with adaptive mitigation of forgetting and overfitting.","marker":"[37]"}],"fun_headline_variants":["Frozen transformer beats prior few-shot continual audio","Ridge regression head tops few-shot continual audio","25 base samples enough to top few-shot continual audio","Analytical classifier solves fully few-shot continual audio","Frozen embeddings from transformer fusion win few-shot continual"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's key assumption is that a pretrained AST encoder, finetuned on only 25 base samples and then frozen, keeps producing discriminative embeddings for all future incremental classes.","fun_headline_variants_meta":{"raw":{"variants":["Frozen transformer beats prior few-shot continual audio","Ridge regression head tops few-shot continual audio","25 base samples enough to top few-shot continual audio","Analytical classifier solves fully few-shot continual audio","Frozen embeddings from transformer fusion win few-shot continual"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000763,"raw_usage":{"total_tokens":3358,"prompt_tokens":888,"completion_tokens":2470,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":2397}},"tokens_in":504,"tokens_out":2470,"duration_ms":21559,"temperature":1.0,"reasoning_tokens":2397,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:15:51.445206+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the same 5-way, 5-shot, ten-session protocol on a fresh random selection of 50 classes from LS-100, NSynth-100, or FSC-89, with the same frozen-extractor and ridge-regression code; if average accuracy does not reproduce the reported margins over AISP (roughly +3.7, +7.3, and +0.3 points), the result depends on the particular class draw or evaluation protocol.","supporting_citations":[{"cited_title":"Class-incremental grouping network for continual audio-visual learning,","cited_arxiv_id":null,"evidence_quote":"Establishes the FFCAC benchmark protocol, datasets, and the strongest baseline that the reported gains are measured against."},{"cited_title":"Few shot continual learning for audio classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained Audio Spectrogram Transformer encoder that the MEE is initialized from and then freezes."},{"cited_title":"Fully few-shot class-incremental audio classification using expandable dual-embedding extractor,","cited_arxiv_id":null,"evidence_quote":"Supplies the closed-form regularized least-squares solution used to derive the classifier update."},{"cited_title":"An efficient incremental learning algorithm for sound classification,","cited_arxiv_id":null,"evidence_quote":"Provides a prior fully few-shot continual audio method with an expandable dual-embedding extractor that this work builds on and compares with."},{"cited_title":"On the transferability of large-scale self-supervision to few-shot audio classification,","cited_arxiv_id":null,"evidence_quote":"Contributes the prototype-based classifier used in the ablation and one of the compared prior methods."},{"cited_title":"Class-incremental learning for multi- label audio classification,","cited_arxiv_id":null,"evidence_quote":"Supplies a compared few-shot class-incremental audio method with adaptive mitigation of forgetting and overfitting."}],"review_version":1}