{"id":"de93c550-7ec8-43a1-8a2a-8c48174ffdb9","arxiv_id":"2505.12904","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Unsupervised contrastive learning on unlabeled hydrophone audio produces embeddings that transfer to ship and marine mammal classification tasks with accuracy comparable to supervised contrastive baselines.","lead":"This paper trains an unsupervised contrastive learning model on a large amount of unlabeled underwater audio from one hydrophone, then checks whether the learned audio embeddings work for recognizing ship types and marine mammal calls on three labeled benchmark datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing random-initialization / wrong-domain control: reported transfer may reflect Conformer inductive bias plus spectrogram statistics rather than VICReg-learned generalized embeddings.","rationale":"I read the paper as a proof of concept. The central inference is from 'a model trained on unlabeled ONC data performs well on three benchmarks' to 'unsupervised CL generates generalized embeddings.' The missing random/domain control is the weakest link because it is the only comparison that would isolate the contribution of the unlabeled data. The reader's distribution-shift concern is real but partially answered by the cross-dataset results themselves: if the ONC distribution were wholly unrepresentative, ShipsEar and Watkins would likely fail. The more dangerous possibility is that even a poorly matched or untrained encoder would get similar numbers, meaning the paper's contribution is the Conformer backbone rather than the learning procedure. The supervised baselines do not close this gap: supervised CL is trained only on Deepship with limited labels and can underperform for reasons unrelated to whether unsupervised VICReg helps. I would keep the current CONDITIONAL verdict but make the random-initialization/wrong-domain control an explicit condition for acceptance. No claim of misconduct; this is a standard missing control.","tokens_in":12367,"tokens_out":5433,"duration_ms":58733,"concrete_test":"Run the Section 3.1 linear-probe protocol with (a) the same Conformer and ResNet18 architectures with random untrained weights, and (b) VICReg trained under identical settings but on white noise or on shuffled spectrograms (preserving marginal statistics). Compare Deepship/ShipsEar/Watkins accuracy and weighted F1 to Table 1. If random-init or noise-pretrained accuracy is within a few points of the unsupervised Conformer's 54.87/57.42/86.10, then the central claim that unlabeled ONC data produces generalized embeddings is not supported. Report at least 3 seeds to assess variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Abstract; Section 1 contribution 3) is that VICReg contrastive training on unlabeled ONC hydrophone data produces embeddings that generalize across Deepship, ShipsEar, and Watkins. The evidence in Section 3 (Tables 1, 3, 4) only compares VICReg-trained encoders with supervised-CL-trained encoders. There is no baseline in which the same Conformer/ResNet18 encoder is evaluated with random weights, or in which VICReg is trained on a non-representative unlabeled corpus (e.g., white noise or a different hydrophone). Without these controls, the observed linear-probe accuracies—especially the unsupervised Conformer's 54.87%/57.42%/86.10% in Table 1—cannot be attributed to the contrastive learning objective or to the ONC data. The Table 3 augmentation ablation shows LowPass alone essentially matches the full family (87.00 vs 87.00 on Watkins, 55.93 vs 55.61 on Deepship), suggesting the class signal may be largely present in the input spectrogram itself. If a randomly initialized Conformer attains similar accuracy, the contribution collapses to architecture choice, not generalized embeddings learned from unlabeled data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unsupervised contrastive learning pipeline for underwater acoustic target recognition (UATR). A Conformer encoder is trained with the VICReg loss on unlabeled recordings from a single ONC hydrophone near Vancouver, and the learned embeddings are evaluated by training a logistic regression classifier on three labeled benchmarks: Deepship (ship type), ShipsEar (ship type), and The Best of Watkin's (marine mammal vocalizations). The method is compared against supervised contrastive learning baselines (ResNet18 and Conformer) and is additionally analyzed with respect to VICReg loss weights, augmentation functions, reduced labeled data, embedding size, and computational cost. The central claim is that the unsupervised contrastive approach produces robust and generalized embeddings that transfer across geographically and temporally diverse datasets, with accuracy comparable to supervised baselines and greater robustness when labeled data are scarce.","tokens_in":12612,"tokens_out":5170,"duration_ms":49137,"significance":"If the claims hold, the paper offers a useful proof-of-concept that publicly available unlabeled hydrophone data can serve as pretraining data for multiple underwater acoustic classification tasks, reducing reliance on scarce labeled data. The paper is honest about using held-out test sets and a time-wise split for Deepship, and it makes the code available. The study addresses a relevant problem for passive acoustic monitoring and is likely to interest the underwater acoustics and applied self-supervised learning communities. However, the strength of the central claim depends on the missing controls and evaluation issues discussed in the major comments. The manuscript is a solid empirical contribution in need of revision rather than a breakthrough result.","major_comments":[{"comment":"The central claim that VICReg training on unlabeled ONC data produces generalized embeddings is not supported by the current baseline set. Tables 1, 3, and 4 compare only VICReg-trained encoders with supervised-CL-trained encoders; there is no control using the same encoder with random weights or with VICReg trained on a non-representative unlabeled corpus. This control is essential because Table 3 shows that the LowPass augmentation alone matches the full augmentation family (87.00 vs 87.00 on Watkin's, 55.93 vs 55.61 on Deepship), suggesting that a large part of the class signal may be present in the input spectrogram itself. Please add (a) linear-probe accuracy on randomly initialized Conformer and ResNet18 encoders and (b) a VICReg training run on a mismatched unlabeled corpus (e.g., white noise or a different hydrophone) to attribute the observed accuracy to the contrastive objective and the specific ONC data.","section":"Section 3, Table 1"},{"comment":"The VICReg loss weights (Table 2), the augmentation family choice (Table 3), and the decision to expand the augmentation family (Table 4) are selected using accuracy on the Deepship test set, which is also used in the final evaluation in Table 1. This is test-set selection and inflates the reported generalization numbers, undermining the claim that the embeddings generalize without dataset-specific tuning. Use a held-out validation split (separate from the test sets) or nested cross-validation to select hyperparameters, and report the corresponding final accuracies. This is load-bearing because the selected configuration is then presented as the proposed method's performance.","section":"Sections 3.2-3.3, Tables 2-4"},{"comment":"The variance term in the VICReg loss is written as abs(γ − S(z^j, ε)), but the original VICReg formulation (Bardes et al., 2021) uses max(0, γ − S(z^j, ε)), a hinge function that only penalizes standard deviations below the target γ. The absolute value also penalizes standard deviations above γ, which is not part of VICReg and can change the learned representation. If the implementation matches the equation, this is a substantive deviation; if not, the equation misleads. Please correct the formula and, if the implementation used the absolute value, report whether the results change under the correct VICReg variance term.","section":"Equation (2)"},{"comment":"All reported accuracy and F1 scores are point estimates from a single training run. Several key comparisons involve small differences, e.g., Deepship 54.87 vs 55.16 and Watkin's 86.10 vs 87.10 between unsupervised and supervised Conformer in Table 1. Without multiple random seeds and error bars or a significance test, the claims of 'comparable performance' and 'more robust' are not statistically supported. Please report mean ± standard deviation over at least three independent runs, or provide a paired significance test for the main comparisons.","section":"All result tables (Tables 1, 3, 4, Figures 5, 7, 8)"},{"comment":"The load-bearing premise that unlabeled recordings from a single ONC hydrophone near Vancouver are sufficiently representative of the acoustic conditions in all three labeled benchmarks—including ShipsEar off Spain and the decades-spanning Watkins database—is never tested empirically. The paper provides no spectral or distributional analysis comparing the unlabeled data to the evaluation datasets, and no experiment with an alternative unlabeled corpus. This makes it difficult to rule out the possibility that the cross-dataset accuracy reflects generic spectrogram statistics rather than genuinely transferable embeddings learned from the ONC data. Please add an analysis of the acoustic feature distributions or a sensitivity experiment using a different unlabeled corpus.","section":"Section 2.1.1"}],"minor_comments":[{"comment":"The symbol v is used both for the variance component function v(Z) and as the scalar weight for the covariance term in Eq. (8). Rename the scalar weight ν (or similar) to avoid confusing the variance function with the covariance weight.","section":"Equation (8)"},{"comment":"The caption says 'The influence of the augmentation functions on the classification of Ship types,' but the table also reports results on the Watkin's marine mammal dataset. Adjust the caption to reflect that both ship type and marine mammal classification are evaluated.","section":"Table 3 caption"},{"comment":"The Gaussian noise augmentation constrains the output SNR between 0.3 dB and 0.5 dB, which would make the noise overwhelmingly stronger than the signal. Please confirm that this range is intentional or whether the SNR values are defined differently (e.g., SNR in dB of the noise added relative to the signal).","section":"Section 2.2, item 2"},{"comment":"The sentence 'This shows to potential of unsupervised methods' contains a typo; it should be 'This shows the potential of unsupervised methods.'","section":"Abstract and Section 1"},{"comment":"The caption reads 'Visualization fo the proposed augmentation functions'; 'fo' should be 'of.'","section":"Figure 2b caption"},{"comment":"The claim 'The first implementation of unsupervised CL on a separate unlabeled underwater acoustic dataset' is strong given that the reference list includes self-supervised work on underwater acoustics (e.g., Xu et al., 2024). Suggest softening to 'to the best of our knowledge' and clarifying the precise distinction from prior self-supervised underwater acoustic studies.","section":"Section 1, contribution 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript tackles a relevant problem and contains a reasonable empirical setup, but the missing random-initialization/wrong-domain controls and the selection of hyperparameters on the Deepship test set are serious enough that the main claim is not yet established. The VICReg variance term also deviates from the original formulation. These issues are fixable within the scope of a revision. The 'first implementation' claim appears somewhat overstated given the existing self-supervised underwater acoustic literature; the authors should scope it carefully. Overall, the paper is more suitable for a venue that values empirical proof-of-concept studies; with the requested controls and corrected evaluation, it could be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something real: it takes a large unlabeled hydrophone recording, trains a Conformer with VICReg, and shows that logistic regression on those embeddings gets competitive accuracy on three labeled benchmarks (Deepship, ShipsEar, Watkins). It also shows the unsupervised model holds up better than supervised baselines when labeled data is scarce. That is a useful result for the UATR community, and the evaluation is much broader than most work in this area.\n\nWhat is genuinely new is the specific combination: contrastive pretraining on a single unlabeled hydrophone, then transfer to ship and marine mammal classification across different geographic regions. The ablation of loss weights and augmentation choices is a nice touch, even if the tuning is coarse. The paper is clearly written, the method is reproducible in principle, and they include a computational cost table, which is rare.\n\nThe main soft spot is the missing control that would isolate the contribution of the contrastive objective. There is no randomly initialized encoder evaluated with the same linear probe, nor a model trained on irrelevant audio. Without that, you cannot fully attribute the transfer accuracy to VICReg learning meaningful structure; some of the signal might come from the Conformer's inductive bias and generic spectrogram statistics. The LowPass-only augmentation nearly matching the full family in Table 3 reinforces this worry. I would not call it fatal, because the cross-dataset transfer to ShipsEar and Watkins is a meaningful sign of generalization, and the supervised baselines provide some comparison. But the claim in the abstract and contribution list says the embeddings are \"robust and generalized,\" and that claim needs the missing control to be airtight.\n\nTwo other issues, both moderate. Hyperparameters (loss weights, augmentation set) are selected using the same benchmark test sets they report, with no separate validation split. The differences are small except the collapsed λ=25 case, so I would not call this heavy overfitting, but it should be acknowledged. Second, they cite Xu et al. (2024) only for mixup, while that paper also does self-supervised learning for UATR; the positioning should be explicit about what is different. Also, no error bars or repeated runs, and the GitHub link lacks a commit hash, so reproducibility is partially unverified.\n\nOverall, this is a worthwhile proof of concept for practitioners. It is not a breakthrough, but it deserves a serious referee. My advice: send it to review, but ask for a random-initialization baseline, repeated runs, and a clearer differentiation from Xu et al. I would bring it to a reading group if the group cares about self-supervised learning in marine acoustics.","headline":"A solid proof of concept for self-supervised pretraining on unlabeled hydrophone data, but the 'generalized embeddings' claim would be stronger with a random-initialization baseline and repeated runs.","tokens_in":13175,"tokens_out":2977,"would_cite":false,"duration_ms":32460,"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":"A self-supervised contrastive encoder trained on unlabeled single-hydrophone audio produces embeddings that transfer to ship-type and marine-mammal classification as well as supervised baselines.","keywords":["Underwater acoustic target recognition","Contrastive learning","VICReg","Self-supervised learning","Ship radiated noise","Marine mammal vocalizations","Conformer","Transfer learning"],"falsifier":"Train the identical VICReg pipeline on unlabeled recordings from a hydrophone far from Vancouver, ideally near the ShipsEar recording region, then attach the same linear classifiers. If accuracy on ShipsEar and Watkins drops substantially compared with the Vancouver-trained model, the claim that a single representative hydrophone yields generalized embeddings is falsified.","tokens_in":12152,"feed_emoji":"🌊","tokens_out":7178,"duration_ms":65796,"temperature":0.7,"pith_summary":"This paper tries to show that useful underwater acoustic classifiers can be built without labeled training data. It trains a Conformer encoder with the VICReg contrastive loss on unlabeled recordings collected at a single hydrophone near Vancouver, then uses the learned embeddings as features for simple logistic-regression classifiers on three labeled benchmarks: two ship-type datasets and one marine-mammal vocalization dataset. The reported accuracy is close to supervised contrastive baselines, and the unsupervised method degrades less when labeled data are reduced. If correct, passive acoustic monitoring could be automated from the large publicly available stock of unlabeled ocean recordings.","feed_headline":"Unlabeled audio from one hydrophone generalizes to three underwater benchmarks","feed_subtitle":"A contrastive VICReg encoder trained without labels matches supervised baselines on ship and whale recognition.","key_machinery":"The central object is the unsupervised contrastive-learning pipeline built on the VICReg loss, Variance-Invariance-Covariance Regularization. VICReg needs no negative samples: it trains the encoder by pulling augmented views of the same clip together in embedding space, pushing the variance of each embedding dimension toward a target, and minimizing off-diagonal covariances to prevent dimensional collapse. Four Conformer blocks serve as the encoder, and a two-layer MLP projects the embeddings for optimization. The augmentation family is the domain-knowledge injection: low-pass filtering at 1 kHz targets the band where ship-type discriminative information concentrates, and a mixup augmentation combines low- and high-frequency content from nearby time windows. This machinery lets the model learn from raw unlabeled audio and still produce features that separate ship types and whale calls in other recordings.","core_discovery":"The central claim is that an unsupervised contrastive-learning pipeline, trained without labels on raw audio from one hydrophone, yields generalized embeddings for underwater acoustic target recognition. The authors define positive pairs by augmenting two-second audio windows with domain-specific functions, then optimize a four-block Conformer encoder with the VICReg loss, which enforces invariance between augmented views, keeps feature variance alive, and decorrelates embedding dimensions. Linear classifiers built on these embeddings reach 54.87% accuracy on Deepship, 57.42% on ShipsEar, and 86.10% on Watkins, close to the supervised Conformer baseline, and they hold up better than the supervised baselines when the labeled fraction of Deepship falls to 25% and below. The paper offers this as a proof-of-concept that unlabeled ocean audio can replace labeled corpora as the training signal for UATR backbones.","pith_inferences":["A stronger version of the claim would hold if training on a hydrophone far from all three benchmarks still transferred; that test would distinguish general acoustic-structure learning from site-specific similarity to Deepship.","The ShipsEar result, where the unsupervised Conformer exceeds the supervised Conformer baseline, is consistent with supervised training on Deepship overfitting site-specific class boundaries, but the paper does not identify the exact spectral or temporal features driving the transfer.","A direct extension would measure whether linear probes trained on Deepship embeddings transfer to ShipsEar and Watkins labels in a zero-shot or few-shot protocol, and compare against spectrogram-statistics baselines.","A systematic sweep over hydrophone location, depth, season, and recording hardware would define the practical envelope of the single-hydrophone claim."],"forward_implications":["Passive acoustic monitoring pipelines could be initialized from unlabeled ocean recordings, removing the dependence on large labeled corpora for underwater acoustic target recognition.","The same backbone may transfer to other underwater acoustic analysis tasks, such as climate-change monitoring or detection of human-made sounds, because the embeddings are not tied to one dataset's label set.","Time-wise evaluation of Deepship suggests that models trained on a random split overstate real-world performance; time-wise generalization is the harder and more deployment-relevant test.","Unsupervised models are more robust than supervised contrastive baselines to reduced label availability on Deepship, so label collection effort could be shifted to downstream evaluation.","Expanding the augmentation family with speech-based augmentations helps supervised models but can hurt unsupervised transfer, making augmentation choice a key design lever for self-supervised UATR."],"supporting_citations":[{"why":"Supplies the VICReg loss function that is optimized on unlabeled underwater audio.","marker":"Bardes et al., 2021"},{"why":"Provides the SimCLR framework and NTXent loss that the supervised contrastive baseline adapts to labeled data.","marker":"Chen et al., 2020"},{"why":"Defines the Conformer encoder architecture used as the backbone of the proposed pipeline.","marker":"Gulati et al., 2020"},{"why":"Contributes the Deepship benchmark dataset and its ship-type labels for evaluation.","marker":"Irfan et al., 2021"},{"why":"Contributes the ShipsEar benchmark dataset, providing the cross-region ship-type test.","marker":"Santos-Domínguez et al., 2016"},{"why":"Contributes the Watkins marine-mammal vocalization dataset, providing the cross-task test.","marker":"Sayigh et al., 2016"},{"why":"Defines the supervised contrastive loss used to train the baseline models for comparison.","marker":"Khosla et al., 2020"},{"why":"Introduces the LARS optimizer used for stable large-batch training of the unsupervised and supervised backbones.","marker":"You et al., 2017"},{"why":"Survey of ship radiated noise that supplies the 50 Hz to 1 kHz band rationale for the low-pass augmentation.","marker":"Hummel et al., 2024"},{"why":"Inspires the temporal neighborhood coding idea behind the mixup augmentation used to define positive samples.","marker":"Tonekaboni et al., 2021"}],"fun_headline_variants":["Contrastive learning turns unlabeled ocean noise into recognition","No labels: contrastive VICReg learns ship and whale sounds","Unsupervised audio embeddings rival supervised underwater baselines","One hydrophone, zero labels: robust acoustic recognition","VICReg on unlabeled audio matches supervised underwater models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that recordings from one hydrophone near Vancouver are representative enough of the acoustic environments in all three benchmark datasets, including ShipsEar off Spain and the decades-spanning Watkins collection, and this premise is never tested.","fun_headline_variants_meta":{"raw":{"variants":["Contrastive learning turns unlabeled ocean noise into recognition","No labels: contrastive VICReg learns ship and whale sounds","Unsupervised audio embeddings rival supervised underwater baselines","One hydrophone, zero labels: robust acoustic recognition","VICReg on unlabeled audio matches supervised underwater models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000706,"raw_usage":{"total_tokens":3175,"prompt_tokens":930,"completion_tokens":2245,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":2164}},"tokens_in":546,"tokens_out":2245,"duration_ms":17356,"temperature":1.0,"reasoning_tokens":2164,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:23:27.438312+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical VICReg pipeline on unlabeled recordings from a hydrophone far from Vancouver, ideally near the ShipsEar recording region, then attach the same linear classifiers. If accuracy on ShipsEar and Watkins drops substantially compared with the Vancouver-trained model, the claim that a single representative hydrophone yields generalized embeddings is falsified.","supporting_citations":[{"cited_title":", author Jiangbin, Z","cited_arxiv_id":null,"evidence_quote":"Contributes the Deepship benchmark dataset and its ship-type labels for evaluation."},{"cited_title":", author Torres-Guijarro, S","cited_arxiv_id":null,"evidence_quote":"Contributes the ShipsEar benchmark dataset, providing the cross-region ship-type test."},{"cited_title":", author Daher, M.A","cited_arxiv_id":null,"evidence_quote":"Contributes the Watkins marine-mammal vocalization dataset, providing the cross-task test."},{"cited_title":", author van der Mei, R.D","cited_arxiv_id":null,"evidence_quote":"Survey of ship radiated noise that supplies the 50 Hz to 1 kHz band rationale for the low-pass augmentation."}],"review_version":1}