{"id":"9417f392-3a4c-4a65-bceb-bba1f8e4a6ff","arxiv_id":"2504.15171","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A prototype-based audio-visual continual learning method, HAIL-FFIA, reaches 75.92% average accuracy on a new six-species fish-feeding dataset while storing only 0.1% of raw data.","lead":"This paper presents a new dataset of over 80,000 audio and video clips of six fish species during feeding, and a method that lets a fish-feeding monitoring system learn new species without forgetting old ones. It matters because aquaculture operators need systems that adapt to new species without storing large amounts of old data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exemplar-based comparisons lack a defined memory budget; without byte/sample-level specification the 2.66-point accuracy margin and 0.1%-storage advantage over AV-CIL are not yet established.","rationale":"I focused on the undefined memory budget because it attacks both components of the central claim simultaneously: the accuracy advantage over exemplar-based methods and the storage-efficiency advantage. The absence of error bars is a real secondary concern, but it only affects whether the 2.66-point margin is noise; the memory-budget issue affects whether the comparison is meaningful at all. If the budget were defined and matched, the paper's conclusion could still hold, which is why this is a conditional concern rather than a rejection. The dataset, the hierarchical representation, and the ablations are constructive, and each component appears to add value, so the work is not undermined in its entirety. The reader's weakest assumption already identified the same issue, and the requested clarification and re-analysis are consistent with the existing conditional verdict; therefore no change to the reader's verdict is needed.","tokens_in":18471,"tokens_out":7630,"duration_ms":76807,"concrete_test":"Re-run AV-CIL, iCaRL-FC, and AFC-LSC with three explicitly defined budgets: (i) byte-matched to HAIL's prototype storage, stating bytes as number of stored prototypes times feature dimension times 4 bytes; (ii) sample-count matched to HAIL's 120 stored prototypes; and (iii) standard exemplar budgets used in the original papers, for example 20 or 50 exemplars per class, with the resulting percentage of raw data reported. Record average accuracy and forgetting for each budget, ideally with per-run values. If HAIL's 75.92% is not the best in all three settings, or if the margin falls within run-to-run variation, the central superiority claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (75.92% vs 73.26% for AV-CIL; 0.10% vs 5-10% raw-data storage) depends on a comparison that the paper does not specify. Section V-D says \"all methods use the same memory budget for storing exemplars\" but neither defines the budget unit nor reports its size. Section VI-C then states HAIL stores about 0.10% of the raw data while exemplar-based methods \"typically require 5-10%\" of the original dataset. These two statements are in tension: if exemplar baselines were run under the same 0.1% budget, they are not at their typical 5-10% regime; if they were run at 5-10%, the \"same memory budget\" statement is false. Exemplar count is the primary accuracy/forgetting knob for iCaRL, SS-IL, AFC, and AV-CIL, so an arbitrarily small exemplar budget can manufacture HAIL's margin. A budget defined in sample count versus bytes also changes the comparison because raw audio-visual clips are orders of magnitude larger than feature prototypes. Without resolving this, the claimed superiority over exemplar-based methods is not supported, and the storage-efficiency headline is not an apples-to-apples comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces AV-CIL-FFIA, a new dataset of 81,932 labeled audio-visual feeding clips from six fish species, and proposes HAIL-FFIA, an exemplar-free prototype-based audio-visual class-incremental learning method. The method separates general feeding-intensity knowledge from species-specific knowledge, maintains compact prototypes via k-means clustering with an EMA update, and uses a dynamic audio-visual balancing mechanism. On the new benchmark, the authors report that HAIL-FFIA achieves 75.92% average accuracy and 9.36% average forgetting in the audio-visual setting, outperforming the strongest baseline AV-CIL (73.26% accuracy, 12.48% forgetting) while storing roughly 0.10% of the raw data volume. The paper also benchmarks single-modality and exemplar-based methods and reports ablations isolating the hierarchical representation, prototype management, and dynamic modality balancing components.","tokens_in":18649,"tokens_out":3311,"duration_ms":33367,"significance":"If the reported results hold, the paper makes a useful contribution: it provides the first audio-visual continual-learning benchmark for fish feeding intensity assessment, with a sizable real-world dataset and public code/data commitments. The prototype-based approach is a genuinely interesting middle ground between exemplar-based replay and fully exemplar-free regularization, and the ablation study gives some evidence that each proposed component contributes. The dataset alone is likely to be valuable to the aquaculture and multimodal continual-learning communities. However, the central empirical claims of superiority over exemplar-based methods and of a large storage advantage rest on an underspecified comparison, and the reported accuracy margin is not accompanied by any measure of variability. These issues are load-bearing and need to be resolved before the results can be taken as established.","major_comments":[{"comment":"The comparison with exemplar-based methods is not reproducible because the memory budget is never defined. Section V-D states that 'all methods use the same memory budget for storing exemplars' but does not specify whether the budget is measured in bytes, in number of raw clips, in number of feature vectors, or in some other unit, nor does it report the budget size. Section VI-C simultaneously claims that HAIL-FFIA stores about 0.10% of the raw dataset while exemplar-based methods 'typically require 5-10%' of the original dataset. These two statements are in tension: if the exemplar-based baselines were run under the same 0.10% budget, they are not operating in their typical 5-10% regime; if they were run in that regime, then the 'same memory budget' statement is false. Since the exemplar count is the primary accuracy/forgetting knob for iCaRL, SS-IL, AFC, and AV-CIL, an arbitrarily small exemplar budget could manufacture the observed margin. The authors should report, for every exemplar-based method, the exact exemplar budget in samples and in bytes, and ideally include a sensitivity sweep showing accuracy as a function of exemplar budget. The 0.10% storage claim should be computed in the same units used for the baselines.","section":"Section V-D and Section VI-C"},{"comment":"The central accuracy claim is not statistically supported. The paper states that all results are averaged over 3 independent runs (Section V-C), but Table I and Fig. 4 report only point estimates with no standard deviations, confidence intervals, or significance tests. The reported advantage over AV-CIL in the audio-visual setting is 2.66 percentage points (75.92% vs. 73.26%), which may well be within run-to-run variability for this kind of continual-learning experiment. The authors should report per-run results or error bars for at least the primary comparison, and ideally a paired significance test across the shared random seeds. Without this, the headline claim of consistent superiority over AV-CIL is not fully established.","section":"Section VI-A, Table I"},{"comment":"The hyperparameter choices that materially affect the method are reported as empirical findings on the same dataset without a sensitivity analysis or a principled selection procedure. Specifically, the prototype count m=5, the EMA coefficient alpha=0.7, the regularization eta=1.0, the modality-balancing bounds gamma_max=0.8 and gamma_min=0.3, and the similarity floor 0.2 in lambda_p all appear to have been chosen on the validation portion of AV-CIL-FFIA. The only controlled sensitivity study is the prototype-count sweep in Fig. 5. This raises a mild overfitting-to-benchmark concern and makes it harder to assess how much of the margin over baselines comes from the proposed architecture versus from per-dataset tuning. The authors should either justify these values with a more systematic analysis or clearly state which values were fixed a priori and which were tuned on the validation split.","section":"Section V-C and Section III-C"}],"minor_comments":[{"comment":"The 'Exemplar-free?' column uses the symbols '%' and '!' without a legend, and the assignment is unclear for Fine-tuning and for the Oracle row; please add an explicit legend or replace the symbols with 'Yes'/'No'/'N/A'.","section":"Table I"},{"comment":"The dataset section states that the camera resolution is 1920x1080, while the video processing section says the original resolution is 2560x1440; these numbers should be reconciled.","section":"Section IV-A vs. Section V-B"},{"comment":"The feature expansion layer is described as a '2-layer linear feed-forward network' and Wav_up is said to be randomly initialized and fixed, but the expanded feature dimension dup is never specified; please state the exact output dimension used in the experiments.","section":"Section III-B.1 and Eq. (7)"},{"comment":"The text gives URLs for some baselines but not for iCaRL, SS-IL, AFC, and AV-CIL in a uniform way; please provide complete citations and code links for all baselines.","section":"Section V-D"},{"comment":"Figure 4 is described as showing accuracy at each incremental step, but the caption does not say whether these are single-run or averaged values; please state this explicitly and, if averaged, include variability information.","section":"Section VI-A, Fig. 4"},{"comment":"The dataset description does not report the per-species and per-intensity class distribution; since the paper uses a fixed label space with species-specific tasks, class imbalance across species could affect the reported accuracy and forgetting numbers, and a brief class-distribution table would improve the benchmark's utility.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: the dataset is the real asset, and the method is a coherent, if incremental, assembly of known pieces. The headline experimental claim, however, rests on a comparison that isn't specified well enough to accept as is.\n\nWhat's actually new: AV-CIL-FFIA, 81,932 synchronised audio-visual clips of feeding intensity for six fish species in real tanks, is a useful benchmark for the aquaculture community. The hierarchical prototype design—general intensity prototypes shared across species plus species-specific prototypes—is an appropriate way to handle the structure of the problem. The ablation study is genuinely informative: each component (hierarchical representation, prototype management, dynamic modality balancing) adds accuracy, and the full model reaches 75.92% average accuracy and 9.36% forgetting. The paper also honestly notes the final species (Sunfish) has lower-quality data.\n\nThe soft spots, in order of importance. First, the memory budget for the exemplar-based baselines is undefined. Section V-D says all methods use the same memory budget but never states the unit or the size. Then Section VI-C says exemplar methods typically need 5–10% of the raw data while HAIL uses 0.10%. These cannot both be true in the way reported. If the baselines were given only a 0.1% budget, they are being run in a regime where they are uncompetitive; if they were given 5–10%, then the 'same memory budget' statement is false. The reviewer needs the budget in bytes or in number of exemplars per class, and the storage comparison made apples-to-apples. Second, the results are averaged over three runs with no error bars, so the 2.66-point margin over AV-CIL (75.92 vs 73.26) might be within run-to-run noise. Third, the task is domain-incremental rather than class-incremental—the labels stay fixed across species—and the 'pioneering CIL' framing oversells this, though the authors do state the label space is constant. Fourth, the GitHub link for code and data appears in the abstract; I haven't verified it resolves, but the paper does claim availability.\n\nOverall, this is a plausible applied engineering paper. The central idea makes sense, and none of the flaws are load-bearing; they are fixable with reporting changes. The dataset alone justifies sending this to review. I'd encourage a serious referee to ask for the memory budget specification, error bars, and a toned-down framing.\n\nRecommendation: engage with it; accept for peer review with a request for major revision.","headline":"Useful dataset plus a plausible prototype-based CIL method, but the exemplar-baseline memory budget is undefined and the 2.66-point margin over AV-CIL needs error bars.","tokens_in":19299,"tokens_out":5340,"would_cite":false,"duration_ms":42289,"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":"A prototype-based hierarchical audio-visual continual learning method reports 75.92% average accuracy and 9.36% forgetting while storing only about 0.10% of the raw data as feature prototypes.","keywords":["fish feeding intensity assessment","audio-visual learning","class-incremental learning","catastrophic forgetting","prototype-based learning","multimodal fusion","aquaculture","continual learning"],"falsifier":"Re-run the AV-CIL-FFIA benchmark giving each exemplar-based baseline a byte budget equal to the size of HAIL-FFIA's prototype store (about 0.10% of the raw data) and compare average accuracy and forgetting; if iCaRL or AV-CIL then matches or exceeds 75.92% accuracy, the superiority claim collapses. A second check is to reproduce the 75.92% figure from the released code and data using the stated species order and prototype count.","tokens_in":18159,"feed_emoji":"🐟","tokens_out":6470,"duration_ms":54245,"temperature":0.7,"pith_summary":"Fish feeding intensity assessment in aquaculture usually trains a model on one fish species and fails when a new species arrives. This paper asks whether a model can keep learning new species without forgetting old ones and without storing raw audio-visual clips. The authors introduce HAIL-FFIA, a prototype-based audio-visual class-incremental learning framework, together with AV-CIL-FFIA, a dataset of 81,932 labelled clips from six fish species in real tanks. They report 75.92% average accuracy with 9.36% forgetting, ahead of exemplar-free and exemplar-based baselines, while storing about 0.10% of the raw data as feature prototypes. If the result holds, aquaculture systems could scale to new species on modest hardware.","feed_headline":"Fish-feeding AI learns new species with 75.9% accuracy","feed_subtitle":"Prototype memories cut storage to 0.1% of raw clips while limiting forgetting to 9.4%.","key_machinery":"The central object is HAIL-FFIA, a hierarchical audio-visual class-incremental learning framework. Pre-trained S3D and PANNs encoders extract visual and audio features, which are fused through bidirectional cross-modal attention. A two-level representation then separates a general intensity layer, solved by closed-form ridge regression, from per-species audio and visual classifiers computed the same way. A prototype management system stores k-means cluster centroids, five per intensity level per species, as the only retained memory; prototype-enhanced updates concatenate those prototypes with new features in the closed-form weight update, and an exponential-moving-average step with stability coefficient $\\alpha=0.7$ updates the prototypes themselves. Dynamic modality balancing computes per-species, per-intensity audio and visual importance weights and mixes general and species-specific predictions with a confidence-based coefficient.","core_discovery":"The paper's central claim is that separating general feeding-intensity knowledge from fish-specific characteristics, stored as compact feature prototypes, lets an audio-visual model add new fish species without catastrophic forgetting. On the new AV-CIL-FFIA benchmark, HAIL-FFIA achieves the highest average accuracy (75.92%) and the lowest forgetting (9.36%) among all compared methods, with a prototype library that stores only about 0.10% of the raw dataset size. The authors attribute the result to three interacting components: a hierarchical representation that isolates shared intensity patterns from species-specific signatures, prototype-enhanced closed-form updates that carry old knowledge into new classifiers without raw data, and a dynamic audio-visual weighting that adapts to feeding conditions.","pith_inferences":["Editorial inference: if the memory budget ambiguity is resolved in favor of bytes, the storage advantage of HAIL-FFIA is real, but the accuracy advantage over exemplar replay should be re-tested under byte-equivalent budgets before generalizing.","Editorial inference: because the label space stays fixed while only the fish species changes, the setup is closer to domain-incremental learning than classic class-incremental learning; framing it that way could invite stronger baselines not used here.","Editorial inference: the same hierarchical prototype recipe could transfer to other multimodal continual-learning tasks where categories are shared across domains, such as acoustic scene or animal behaviour recognition.","Editorial inference: a direct test of the dynamic modality balancing would be to degrade one modality (e.g., turbid water or added noise) and measure whether accuracy drops less than with fixed fusion weights."],"forward_implications":["An aquaculture operator could add a new fish species to a monitoring system by storing only a few prototype vectors per feeding intensity, making multi-species deployment feasible on resource-constrained hardware.","Audio-visual fusion consistently beats single-modality inputs in the incremental setting: HAIL-FFIA gains 5.65% accuracy over audio-only and 3.34% over visual-only, and every baseline shows a similar gap.","Exemplar-free prototype learning can match or beat exemplar replay on this benchmark (75.92% vs 73.26% for AV-CIL) while avoiding raw clip storage, which also reduces privacy and data-transfer concerns.","The Oracle upper bound of 86.78% shows a remaining 10.86-point gap between continual learning and full joint training, indicating room for stronger knowledge-preservation mechanisms."],"supporting_citations":[{"why":"Defines the audio-visual FFIA task and reports multimodal fusion gains that this work extends to the incremental setting.","marker":"[1]"},{"why":"Provides the AV-CIL audio-visual class-incremental learning method and serves as the strongest baseline the paper must outperform.","marker":"[32]"},{"why":"Supplies MMAL, the exemplar-free multimodal analytic learning baseline closest to HAIL-FFIA's closed-form approach.","marker":"[34]"},{"why":"Provides LwF knowledge-distillation baseline representing exemplar-free continual learning.","marker":"[19]"},{"why":"Provides iCaRL, the exemplar-replay baseline whose storage overhead HAIL-FFIA aims to avoid.","marker":"[42]"},{"why":"Provides ACIL, the analytic learning method whose closed-form weight updates HAIL-FFIA adapts with prototypes.","marker":"[51]"},{"why":"Supplies the pre-trained S3D visual encoder that extracts spatio-temporal features from video clips.","marker":"[57]"},{"why":"Supplies the pre-trained PANNs audio encoder that extracts spectral features from feeding sounds.","marker":"[58]"}],"fun_headline_variants":["Fish feeding AI: 75.9% accuracy, 0.1% storage","Audio-visual fish AI adds species, forgets only 9.4%","Prototype-based model keeps fish feeding knowledge, cuts storage","New fish species? Fish AI learns without forgetting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison against exemplar-based methods assumes the phrase 'same memory budget' has a clear meaning, but the paper never states whether the budget is measured in bytes or in number of stored samples, so the claimed storage-efficiency and accuracy advantage could be an artifact of the comparison.","fun_headline_variants_meta":{"raw":{"variants":["Fish feeding AI: 75.9% accuracy, 0.1% storage","Audio-visual fish AI adds species, forgets only 9.4%","Prototype-based model keeps fish feeding knowledge, cuts storage","New fish species? Fish AI learns without forgetting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0008,"raw_usage":{"total_tokens":3538,"prompt_tokens":988,"completion_tokens":2550,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":2474}},"tokens_in":604,"tokens_out":2550,"duration_ms":19479,"temperature":1.0,"reasoning_tokens":2474,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:31:25.338848+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the AV-CIL-FFIA benchmark giving each exemplar-based baseline a byte budget equal to the size of HAIL-FFIA's prototype store (about 0.10% of the raw data) and compare average accuracy and forgetting; if iCaRL or AV-CIL then matches or exceeds 75.92% accuracy, the superiority claim collapses. A second check is to reproduce the 75.92% figure from the released code and data using the stated species order and prototype count.","supporting_citations":[{"cited_title":"Multimodal fish feeding intensity assessment in aquaculture,","cited_arxiv_id":null,"evidence_quote":"Defines the audio-visual FFIA task and reports multimodal fusion gains that this work extends to the incremental setting."},{"cited_title":"Audio-visual class-incremental learning,","cited_arxiv_id":null,"evidence_quote":"Provides the AV-CIL audio-visual class-incremental learning method and serves as the strongest baseline the paper must outperform."},{"cited_title":"Mmal: Multi-modal analytic learning for exemplar-free audio-visual class incremental tasks,","cited_arxiv_id":null,"evidence_quote":"Supplies MMAL, the exemplar-free multimodal analytic learning baseline closest to HAIL-FFIA's closed-form approach."},{"cited_title":"iCaRL: Incremental classifier and representation learning,","cited_arxiv_id":null,"evidence_quote":"Provides iCaRL, the exemplar-replay baseline whose storage overhead HAIL-FFIA aims to avoid."},{"cited_title":"ACIL: Analytic class-incremental learning with absolute memorization and privacy protection,","cited_arxiv_id":null,"evidence_quote":"Provides ACIL, the analytic learning method whose closed-form weight updates HAIL-FFIA adapts with prototypes."},{"cited_title":"Rethinking spa- tiotemporal feature learning: Speed-accuracy trade-offs in video clas- sification,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained S3D visual encoder that extracts spatio-temporal features from video clips."},{"cited_title":"PANNs: Large-scale pretrained audio neural networks for audio pattern recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained PANNs audio encoder that extracts spectral features from feeding sounds."}],"review_version":1}