Pith. sign in

REVIEW 4 major objections 7 minor 29 references

Language-Guided Contrastive Audio-Visual Masked Autoencoder with Automatically Generated Audio-Visual-Text Triplets from Videos

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that training a contrastive audio-visual masked autoencoder with additional audio-text and visual-text contrastive losses, using captions automatically generated from videos and filtered by CLAP, improves audio-visual…

desk verdict Useful incremental extension of CAV-MAE with automatic language supervision; results probably real but test-set selection and reporting errors keep it from being publishable as is. read the letter →

arxiv 2507.11967 v1 pith:PEKCP2SB submitted 2025-07-16 cs.CV eess.ASeess.IV

Yuchi Ishikawa , Shota Nakada , Hokuto Munakata , Kazuhiro Saito , Tatsuya Komatsu , Yoshimitsu Aoki This is my paper
classification cs.CVeess.ASeess.IV
keywords audio-visualrepresentationlearningcontrastivemaskedautoencoderaudio-visual-texttripletsCLAPfilteringimagecaptioningretrievalclassification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a contrastive audio-visual masked autoencoder can learn sharper audio-visual correspondences if it is also trained to align both modalities to natural-language captions. Since existing audio-visual datasets rarely provide text or guarantee that the soundtrack matches the visuals, the paper generates the captions itself: an image captioning model describes sampled frames, and CLAP is used to keep only the captions whose text matches the video's audio. The resulting model, LG-CAV-MAE, is reported to outperform CAV-MAE and DETECLAP on audio-to-visual and visual-to-audio retrieval and on AudioSet20K classification, with gains up to 5.6% recall@10 and 3.2% mAP. The significance would be that noisy, unlabeled video could be converted at scale into fine-grained tri-modal supervision without manual annotation.

What carries the argument

The central machinery is the tri-modal loss extension of CAV-MAE: two InfoNCE losses, $\mathcal{L}_{a2t}$ (audio-text) and $\mathcal{L}_{v2t}$ (visual-text), are added to the reconstruction and audio-visual contrastive losses, using a frozen pretrained text encoder (CLAP) whose embeddings are projected into the shared space. A second mechanism carries the data: an image captioning model generates several frame captions per video, and the CLAP audio-text score selects the best caption and filters the resulting triplets to the top $k\%$. The conceptual hinge is the paper's stated assumption that a caption matching the audio is evidence that audio and visual content correspond.

What would settle it

A reader could test this by taking held-out videos, generating captions, and having human annotators rate whether each caption describes the soundtrack and whether the audio matches the visuals; if CLAP scores do not rank-order those human ratings, the filtering step is not selecting aligned triplets and the reported gains need a different explanation.

Watch

Extended reading notes

Core claim

The central claim, stated on the authors' terms, is that text acts as a semantic bridge: the InfoNCE losses between audio and text and between visual and text pull the two perceptual embeddings toward the same linguistic description, letting the model capture fine details such as the difference between intense and gentle guitar playing that object-label supervision misses. The paper also claims that a caption generated from a video frame is a reliable text label when CLAP judges it similar to the video's audio, and that filtering triplets by CLAP score removes noisy pairs better than random subsampling. Trained with objective $L = L_{\mathrm{rec}} + \lambda_1 L_c + \lambda_2(L_{a2t} + L_{v2t})$ on VGGSound plus the top 30% of Kinetics700 triplets, LG-CAV-MAE reaches recall@10 of 54.5 for audio-to-visual and 56.5 for visual-to-audio retrieval on VGGSound, and 42.8 mAP on AudioSet20K, surpassing both baselines.

Load-bearing premise

The load-bearing premise is that a caption generated from a video frame, when judged similar to the video's audio by the pretrained audio-text model CLAP, also implies that the video's audio and visual content are well aligned.

Editorial extensions

If this is right

  • Automatically generated and CLAP-filtered triplets should let audio-visual pretraining scale to large unlabeled video collections without annotation costs.
  • The top-30% CLAP-filtered subset of Kinetics700 added to VGGSound consistently improves both retrieval directions and classification, so data-quality filtering matters more than raw data volume.
  • Text-guided contrastive losses improve AudioSet20K classification (42.8 vs 39.6 mAP for the DETECLAP baseline), suggesting the learned representations transfer to downstream tasks.
  • Because the method does not need predefined object labels, it can capture caption-level nuances that object-label supervision like DETECLAP misses.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: because CLAP is both the filtering oracle and the frozen text encoder, part of the reported gain may be CLAP's own audio-text prior being distilled into the audio-visual encoder rather than newly discovered audio-visual structure; a cleaner isolation would filter with one model and train with another.
  • Inference: the caption-selection rule, one caption per video chosen by the highest CLAP score, likely biases the triplet dataset toward visible sound sources and away from off-screen or ambient audio; a dedicated evaluation on classes with off-screen sound would test whether the gains concentrate on visually grounded events.
  • Inference: the same generate-score-filter recipe could transfer to other unlabeled multimodal corpora, such as egocentric video or surveillance footage, since nothing in the pipeline depends on category labels or curated metadata.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes LG-CAV-MAE, an extension of CAV-MAE that adds audio-text and visual-text InfoNCE losses (Eqs. 9-11) using a frozen CLAP text encoder, and trains this model on audio-visual-text triplets automatically generated from unlabeled videos. The triplets are produced by captions from image captioning models (BLIP2/LLaVa) and filtered by CLAP audio-text similarity scores. The method is evaluated on VGGSound and AudioSet20K for audio-to-visual and visual-to-audio retrieval and for classification, reporting up to +5.6% R@10 and +3.2% mAP over existing methods. The paper also includes ablations over the triplet filtering ratio k, the caption model, and the contrastive weight lambda_2.

Significance. If the reported gains are robust, the paper makes a useful contribution by showing a simple, annotation-free way to leverage generated text for audio-visual representation learning, and the automatic triplet-generation pipeline could be reused by others. The inclusion of a random-sampling ablation (Table 4) is a positive step, since it provides some evidence that CLAP-based filtering, rather than merely adding more text data, drives the improvement. However, the absence of a held-out validation protocol for hyperparameter selection, the suspicious duplicate rows in Table 3, and the unresolved role of CLAP as both filter and text encoder currently make the headline claims uncertain. The central idea is defensible, but the evidence as presented is not yet convincing.

major comments (4)
  1. [Section 3.2, Tables 1-3] The values of lambda_2, the caption model, and the filtering ratio k are all selected by their R@K performance on VGGSound and AudioSet20K, which are the same evaluation benchmarks used for the headline comparisons in Tables 5 and 6. Since no held-out validation split is described, the reported improvements may reflect test-set selection rather than a genuine advantage of the method. Please provide a validation protocol (e.g., a split of the training data) or demonstrate that the chosen hyperparameters are stable across multiple evaluation subsets or training runs.
  2. [Table 3] In Table 3, the rows for K(10%) are numerically identical on VGGSound and AudioSet20K for both retrieval directions (audio-to-visual: 4.4/11.8/17.4; visual-to-audio: 3.1/10.3/15.2). This is implausible as a genuine result and strongly suggests a copy-paste error. Please correct these numbers or re-run the experiments, and provide multiple-seed results or confidence intervals for the main tables to establish reliability.
  3. [Section 2.3 and Eqs. (9)-(10)] The same CLAP model is used both to filter triplets (Section 2.3) and as the frozen text encoder in the audio-text and visual-text InfoNCE losses (Eqs. 9-10). This creates a feedback loop in which the training signal is biased toward data that CLAP already considers aligned, so the observed gain could be due to distilling CLAP's audio-text prior rather than learning genuinely better audio-visual correspondences. The random-sampling ablation (Table 4) shows that filtering matters, but it does not separate the filtering effect from the choice of text encoder. Please add an experiment using a different text encoder (e.g., a general sentence encoder not pre-trained on audio) or otherwise analyze how much of the improvement is attributable to the CLAP prior.
  4. [Section 3.1 and Table 5] The comparison with CAV-MAE and DETECLAP in Table 5 is central to the claim of significant improvement, but the manuscript does not report training details for these baselines (epochs, batch size, learning rate, data augmentation, number of frames per video). Please provide these details to establish that the comparison is fair, and state whether the baselines were tuned on the same validation protocol as the proposed method.
minor comments (7)
  1. [Throughout] There are inconsistent spellings of the method name: 'CA V-MAE' appears in many places (e.g., the abstract and Section 2.1) while the standard form 'CAV-MAE' is used elsewhere; please standardize.
  2. [Table 3 header] The header contains a typo: 'Perfromance' should be 'Performance'.
  3. [Figure 2 caption] The phrase 'an image caption models' is grammatically incorrect; it should be 'an image captioning model'.
  4. [Section 2.3] The sentence 'In our experiments, we generate captions for frames per second' is incomplete; please specify the frame rate (e.g., 1 frame per second).
  5. [Section 3.1] The description 'the 2023 version of CLAP' is vague; please cite the exact model checkpoint or version used.
  6. [Table 2] The choice of BLIP2 over LLaVa is not clearly justified by the numbers, since LLaVa is better on several AudioSet20K metrics (e.g., visual-to-audio R@1 and R@10); please state the criterion used for selecting the caption model.
  7. [Table 4] The random-sampling baseline '+K (rand. 30%)' appears to be a single random subset; please report results with multiple random seeds or provide the variance, as random sampling can affect the outcome.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the empirical pipeline is self-contained, with only a minor self-referential use of CLAP as both filter and text encoder.

full rationale

The paper is an empirical systems paper, not a derivation, so there is no definitional circularity. The closest to self-reference is that CLAP is used both to filter audio-visual-text triplets (Section 2.3) and as the frozen text encoder in the audio-text and visual-text InfoNCE losses (Equation 11). This is not circular: the reported predictions are audio-visual retrieval and classification scores on VGGSound and AudioSet20K, not CLAP similarity, and the ablation in Table 4 shows that CLAP-filtered triplets outperform randomly selected triplets, so the filtering contributes content-based signal beyond simply using the same embedding space. The only self-citation is [8] (DETECLAP), which is used as a baseline and for experimental setting; it is not load-bearing for the central claim. Hyperparameters such as lambda_2, the captioning model, and the CLAP filtering ratio are selected on the same evaluation sets (Tables 1-3), which is a test-set selection risk rather than a circularity, since the final numbers are still computed from the selected models. No step in the paper reduces by construction to its own input.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. Its central claim rests on empirical assumptions about CLAP as a scoring function, image captioning quality, and dataset transfer, plus a few tuned hyperparameters.

free parameters (3)
  • lambda_2 (text contrastive weight) = 0.01
    Chosen via ablation in Table 1; higher values (0.05, 0.1) degrade retrieval performance.
  • k (CLAP filtering threshold) = 30%
    Selected from {10%, 30%, 50%} in Table 3 as the best balance when combining VGGSound and Kinetics700; selection is based on evaluation dataset results.
  • frames-per-second caption rate = not reported
    The paper's text leaves the sampling rate blank ('we generate captions for frames per second'), but it affects triplet quality.
assumptions (5)
  • domain assumption CLAP audio-text similarity is a reliable proxy for audio-visual alignment
    Used in Section 2.3 to filter triplets; if false, high-scoring clips may still have mismatched audio and video.
  • domain assumption Image captioning models (BLIP2/LLaVa) produce captions that accurately describe visual content
    Captions are the source of text; errors propagate to triplets.
  • domain assumption The pretrained text encoder CLAP 2023 provides a stable and meaningful text embedding space for contrastive training
    The audio-text and visual-text losses depend on the quality and stability of CLAP text embeddings.
  • domain assumption Training on VGGSound and Kinetics700 transfers to AudioSet20K evaluation
    The model is evaluated on AudioSet20K without pretraining on it, assuming the learned representations generalize across datasets.
  • standard math Gradient-based optimization converges and does not introduce systematic bias
    Standard assumption in deep learning experiments; the paper reports a single run per setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Language-Guided Contrastive Audio-Visual Masked Autoencoder with Automatically Generated Audio-Visual-Text Triplets from Videos." pith.science (2026). https://pith.science/paper/PEKCP2SB

@misc{pith2026250711967,
  author       = {Pith},
  title        = {Pith review of: Language-Guided Contrastive Audio-Visual Masked Autoencoder with Automatically Generated Audio-Visual-Text Triplets from Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PEKCP2SB}},
  note         = {Machine review of arXiv:2507.11967}
}
read the original abstract

In this paper, we propose Language-Guided Contrastive Audio-Visual Masked Autoencoders (LG-CAV-MAE) to improve audio-visual representation learning. LG-CAV-MAE integrates a pretrained text encoder into contrastive audio-visual masked autoencoders, enabling the model to learn across audio, visual and text modalities. To train LG-CAV-MAE, we introduce an automatic method to generate audio-visual-text triplets from unlabeled videos. We first generate frame-level captions using an image captioning model and then apply CLAP-based filtering to ensure strong alignment between audio and captions. This approach yields high-quality audio-visual-text triplets without requiring manual annotations. We evaluate LG-CAV-MAE on audio-visual retrieval tasks, as well as an audio-visual classification task. Our method significantly outperforms existing approaches, achieving up to a 5.6% improvement in recall@10 for retrieval tasks and a 3.2% improvement for the classification task.

Figures

Figures reproduced from arXiv: 2507.11967 by the authors.

Figure 1
Figure 1. Existing audio-visual-text data vs. our generated triplets. (left) Existing datasets often lack strong alignment be￾tween audio and vision or have missing text. (right) Our pro￾posed triplet generation method is based on the idea that if a caption generated from a video describes the audio accurately, the audio and visual components are also likely to correspond. Therefore, this method can generate well-aligned trip… view at source ↗
Figure 2
Figure 2. (a) The upper part illustrates the training process of the contrastive audio-visual masked autoencoder (CAV-MAE) [1]. The lower part shows how we introduce additional audio-text and visual-text contrastive losses using a pretrained text encoder, extending CAV-MAE to a tri-modal framework (LG-CAV-MAE). (b) For LG-CAV-MAE training, we propose to automatically generate audio￾visual-text triplets from unlabeled videos u… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 18 canonical work pages

  1. [1]

    Introduction Understanding the relationship between audio and visual modalities is crucial for various applications, including video content analysis and autonomous systems. Existing meth- ods [1, 2, 3] have sought to learn correspondences between au- ditory and visual signals through self-supervised audio-visual learning, achieving high performance on cl...

    work page Pith review arXiv 2025

  2. [2]

    We also propose to automatically gener- ate audio-visual-text triplets from unlabeled videos, which are subsequently used for training LG-CA V-MAE

    Proposed Method In this paper, we propose Language-Guided CA V-MAE (LG- CA V-MAE) which is built upon CA V-MAE to enhance audio- visual representations. We also propose to automatically gener- ate audio-visual-text triplets from unlabeled videos, which are subsequently used for training LG-CA V-MAE. In the follow- ing sections, we first provide an overvie...

    work page

  3. [3]

    Experiments 3.1. Experimental settings To evaluate the efficacy of our LG-CA V-MAE and the automatic generation of audio-visual-text triplets, we pretrain LG-CA V- MAE with the generated audio-visual-text triplets and assess its performance on audio-to-visual and visual-to-audio retrieval tasks, as well as on audio-visual classification. Dataset: For data...

    work page

  4. [4]

    Conclusion In this paper, we proposed LG-CA V-MAE, a tri-modal ex- tension of CA V-MAE that incorporates audio-text and visual- text contrastive losses. To address the limited availability of high-quality audio-visual-text data, we introduced an auto- matic method for generating audio-visual-text triplets from un- labeled videos via image captioning and C...

    work page

  5. [5]

    Contrastive audio-visual masked au- toencoder,

    Y . Gong, A. Rouditchenko, A. H. Liu, D. Harwath, L. Karlinsky, H. Kuehne, and J. Glass, “Contrastive audio-visual masked au- toencoder,” arXiv preprint arXiv:2210.07839, 2022

    arXiv 2022

  6. [6]

    Mavil: Masked audio-video learners,

    P.-Y . Huang, V . Sharma, H. Xu, C. Ryali, Y . Li, S.-W. Li, G. Ghosh, J. Malik, C. Feichtenhofer et al. , “Mavil: Masked audio-video learners,” Advances in Neural Information Process- ing Systems, vol. 36, 2024

    work page 2024

  7. [7]

    Audiovisual masked autoencoders,

    M.-I. Georgescu, E. Fonseca, R. T. Ionescu, M. Lucic, C. Schmid, and A. Arnab, “Audiovisual masked autoencoders,” in Proceed- ings of the IEEE/CVF International Conference on Computer Vi- sion, 2023, pp. 16 144–16 154

    work page 2023

  8. [8]

    Audio set: An ontology and human-labeled dataset for audio events,

    J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” in 2017 IEEE inter- national conference on acoustics, speech and signal processing (ICASSP). IEEE, 2017, pp. 776–780

    2017

Show all 29 references
  1. [9]

    The kinetics human action video dataset,

    W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev et al., “The kinetics human action video dataset,” arXiv preprint arXiv:1705.06950, 2017

    2017 arXiv

  2. [10]

    Vggsound: A large-scale audio-visual dataset,

    H. Chen, W. Xie, A. Vedaldi, and A. Zisserman, “Vggsound: A large-scale audio-visual dataset,” in ICASSP 2020-2020 IEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2020, pp. 721–725

    2020

  3. [11]

    Acav100m: Automatic curation of large-scale datasets for audio-visual video representation learning,

    S. Lee, J. Chung, Y . Yu, G. Kim, T. Breuel, G. Chechik, and Y . Song, “Acav100m: Automatic curation of large-scale datasets for audio-visual video representation learning,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 10 274–10 284

    2021

  4. [12]

    Deteclap: Enhancing audio-visual representation learning with object information,

    S. Nakada, T. Nishimura, H. Munakata, M. Kondo, and T. Ko- matsu, “Deteclap: Enhancing audio-visual representation learning with object information,”arXiv preprint arXiv:2409.11729, 2024

    2024 arXiv

  5. [13]

    Clap learning audio concepts from natural language supervision,

    B. Elizalde, S. Deshmukh, M. Al Ismail, and H. Wang, “Clap learning audio concepts from natural language supervision,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  6. [14]

    Dense-captioning events in videos,

    R. Krishna, K. Hata, F. Ren, L. Fei-Fei, and J. Carlos Niebles, “Dense-captioning events in videos,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 706–715

    2017

  7. [15]

    Howto100m: Learning a text-video embedding by watching hundred million narrated video clips,

    A. Miech, D. Zhukov, J.-B. Alayrac, M. Tapaswi, I. Laptev, and J. Sivic, “Howto100m: Learning a text-video embedding by watching hundred million narrated video clips,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2019, pp. 2630–2640

    2019

  8. [16]

    Vatex: A large-scale, high-quality multilingual dataset for video- and-language research,

    X. Wang, J. Wu, J. Chen, L. Li, Y .-F. Wang, and W. Y . Wang, “Vatex: A large-scale, high-quality multilingual dataset for video- and-language research,” inProceedings of the IEEE/CVF interna- tional conference on computer vision, 2019, pp. 4581–4591

    2019

  9. [17]

    Avset-10m: An open large-scale audio-visual dataset with high correspondence,

    X. Cheng, Z. Zhang, Z. Wang, M. Fang, R. Huang, S. Zheng, R. Hu, B. Jionghao, T. Jin, and Z. Zhao, “Avset-10m: An open large-scale audio-visual dataset with high correspondence,” 2024

    2024

  10. [18]

    Sound event envelope estimation in polyphonic mixtures,

    I. Mart ´ın-Morat´o, A. Mesaros, T. Heittola, T. Virtanen, M. Cobos, and F. J. Ferri, “Sound event envelope estimation in polyphonic mixtures,” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2019, pp. 935–939

    2019

  11. [19]

    Wavcaps: A chatgpt-assisted weakly- labelled audio captioning dataset for audio-language multimodal research,

    X. Mei, C. Meng, H. Liu, Q. Kong, T. Ko, C. Zhao, M. D. Plumb- ley, Y . Zou, and W. Wang, “Wavcaps: A chatgpt-assisted weakly- labelled audio captioning dataset for audio-language multimodal research,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, 2024

    2024

  12. [20]

    Lp-musiccaps: Llm-based pseudo music captioning,

    S. Doh, K. Choi, J. Lee, and J. Nam, “Lp-musiccaps: Llm-based pseudo music captioning,” arXiv preprint arXiv:2307.16372 , 2023

    2023 arXiv

  13. [21]

    Sound-vecaps: Improving audio generation with visual enhanced captions,

    Y . Yuan, D. Jia, X. Zhuang, Y . Chen, Z. Liu, Z. Chen, Y . Wang, Y . Wang, X. Liu, X. Kang et al. , “Sound-vecaps: Improving audio generation with visual enhanced captions,” arXiv preprint arXiv:2407.04416, 2024

    2024 arXiv

  14. [22]

    A short note on the kinetics-700 human action dataset,

    J. Carreira, E. Noland, C. Hillier, and A. Zisserman, “A short note on the kinetics-700 human action dataset,” arXiv preprint arXiv:1907.06987, 2019

    1907 arXiv

  15. [23]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in International conference on machine learning . PMLR, 2020, pp. 1597–1607

    2020

  16. [24]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  17. [25]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,” in International conference on machine learning. PMLR, 2023, pp. 19 730–19 742

    2023

  18. [26]

    Improved baselines with visual instruction tuning,

    H. Liu, C. Li, Y . Li, and Y . J. Lee, “Improved baselines with visual instruction tuning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 26 296– 26 306

    2024

  19. [27]

    Audiovisual slowfast networks for video recognition,

    F. Xiao, Y . J. Lee, K. Grauman, J. Malik, and C. Feichtenhofer, “Audiovisual slowfast networks for video recognition,” arXiv preprint arXiv:2001.08740, 2020

    2001 arXiv

  20. [28]

    Masked autoencoders that lis- ten,

    P.-Y . Huang, H. Xu, J. Li, A. Baevski, M. Auli, W. Galuba, F. Metze, and C. Feichtenhofer, “Masked autoencoders that lis- ten,” Advances in Neural Information Processing Systems, vol. 35, pp. 28 708–28 720, 2022

    2022

  21. [29]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , Y . Bengio and Y . LeCun, Eds., 2015. [Online]. Available: http://arx...

    2015 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.