Pith. sign in

REVIEW 4 major objections 6 minor 33 references

MDD-Net: Multimodal Depression Detection through Mutual Transformer

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read MDD-Net claims cross-modal mutual-transform fusion of acoustic and visual features detects depression in vlogs better than prior published methods on D-Vlog.

desk verdict Plausible architecture but the SOTA claim is inflated; the real margin over the strongest baseline is 1.82% and likely within noise given protocol mismatches. read the letter →

arxiv 2508.08093 v1 pith:BULFRVIA submitted 2025-08-11 cs.CV cs.LGcs.MMeess.AS

Md Rezwanul Haque , Md. Milon Islam , S M Taslim Uddin Raju , Hamdi Altaheri , Lobna Nassar , Fakhri Karray This is my paper · ORCID
classification cs.CVcs.LGcs.MMeess.AS
keywords depressiondetectionmultimodalfusionmutualtransformeraudio-visualfeaturesvlogdataD-Vlogdatasetfaciallandmarksacousticdescriptors
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 asks whether exchanging attention between two modalities, acoustic and visual, can improve automatic depression detection from social-media vlogs. It introduces MDD-Net, whose mutual-transformer block lets each modality query the other before features are pooled and classified. On the D-Vlog dataset the reported F1 is 0.7707, recall 0.8065, and precision 0.7392, with F1 gains of 1.82% to 17.37% over published audio-visual methods. The system uses only low-level acoustic descriptors and facial landmarks, so if the result holds it suggests screening for depression can work from privacy-preserving non-verbal cues alone.

What carries the argument

The Mutual Transformer block is the mechanism that carries the argument. It computes audio-to-video and video-to-audio correlations by cross-feeding queries from one modality and keys and values from the other, then applies a joint transformer to the concatenated acoustic-visual sequence. The three resulting representations are pooled and concatenated into a fused representation, which the detection layer scores through learned attention. The same features combined by addition, multiplication, or plain concatenation score lower in the paper's ablation (F1 0.73 to 0.75 versus 0.7707), which is taken as evidence that bidirectional cross-modal correlation is the active ingredient.

What would settle it

Take a random subset of D-Vlog test videos and have clinicians who are blind to the titles assign depression labels; compare with the keyword-derived labels. If agreement is near chance, or if MDD-Net classifies a style-matched control set, slow, low-energy vlogs by people who deny depression, as depressed at high rates, the central claim is unsupported. Separately, re-run MDD-Net and MDA VIF under identical folds, seeds, and early-stopping rules to see whether the 1.82% F1 gap survives.

Watch

Extended reading notes

Core claim

The paper's central claim is that mutual cross-modal attention, rather than simple feature fusion, is what makes depression detectable from vlog audio and video. In MDD-Net, the acoustic branch provides queries while the visual branch provides keys and values, and symmetrically the visual branch queries the acoustic branch; a third transformer fuses the concatenated embeddings. Pooling all three outputs yields a joint representation that scores 72.60% accuracy and 0.7707 F1 on D-Vlog, outperforming the acoustic-only branch (0.4106 F1), the visual-only branch (0.4239 F1), and earlier published audio-visual models, with the smallest reported margin (1.82% F1) against MDA VIF and the largest (1

Load-bearing premise

The result stands or falls on D-Vlog's labels actually marking depression rather than video genre or self-presentation style, and on prior methods' published scores being comparable under the same protocol; the authors themselves note that human, title-based labeling makes mislabeling inevitable.

Editorial extensions

If this is right

  • If MDD-Net is right, fusion strategy matters more than feature choice: mutual-transform fusion beats add, multiply, and concatenate by roughly 2.4 to 3.7 F1 points on the same features.
  • Neither modality is sufficient alone on D-Vlog: acoustic-only and visual-only branches reach only 0.41 to 0.42 F1, so vlog-based screening should use both signals.
  • Depression screening could operate on privacy-light descriptors, 25 acoustic low-level descriptors and 68 facial landmarks per frame, rather than raw audio or video.
  • The claimed improvement over previous methods on this benchmark is at least 1.82% F1 (versus MDA VIF) and as high as 17.37% (versus Bi-LSTM), so the architecture is presented as the new best result on D-Vlog.

Reading between the lines

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

  • A natural test the paper leaves implicit: re-annotate D-Vlog labels with clinicians blind to video titles. If agreement with the keyword-based labels is low, the 0.7707 F1 may partly reflect genre or style rather than depression.
  • Because the mutual-transformer exchange is modality-agnostic, the same audio-to-video attention could transfer to audio-to-text or video-to-physiology pairs, where one modality is often noisy or missing.
  • The tight margin over MDA VIF (1.82% F1) is small enough that fold variation, early stopping, and class imbalance could matter; a multi-seed, matched-protocol rerun would show whether the advantage is stable.
  • A style-matched control experiment, comparing depressed vlogs with equally slow, low-energy but non-depressed vlogs, would test whether the model picks up depression or simply low arousal.
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 / 6 minor

Summary. The paper proposes MDD-Net, a multimodal depression detection architecture that fuses acoustic and visual features on the D-Vlog dataset using a mutual-transformer mechanism. The model extracts acoustic features via a global self-attention module and visual features via patch embeddings with hierarchical multi-head self-attention, then computes audio-to-video and video-to-audio correlations together with a joint transformer fusion, followed by a detection layer with a combined BCE, focal, and L2 loss. The experiments report F1 = 0.7707 and claim state-of-the-art improvement ranging from 1.82% to 17.37% over prior methods. The manuscript also includes ablation studies comparing fusion strategies and unimodal modules, and the source code is publicly available.

Significance. The architectural idea of using mutual transformers for cross-modal audio-visual fusion is plausible and the use of a public dataset plus released code supports reproducibility. If the reported performance were demonstrated under a protocol-matched, subject-disjoint evaluation with proper uncertainty quantification, the paper would be a useful contribution to multimodal depression detection. However, the current empirical validation does not establish the central SOTA claim: the comparison mixes 10-fold cross-validation with published single-split baseline numbers, potential subject-level leakage is not addressed, and no error bars or significance tests are provided. The improvement over the strongest baseline (1.82% in F1) is comparable to the noise level expected from these issues.

major comments (4)
  1. [Section IV-B and Table I] The experimental protocol is not matched to the baselines. Section IV-B states that a 10-fold cross-validation is implemented, while Section IV-A describes D-Vlog as partitioned into a fixed 7:1:2 train/validation/test split. The prior methods listed in Table I were published using that fixed split, not re-run under the authors' 10-fold protocol. Averaging 10 folds and comparing to single-split literature values is not a valid head-to-head comparison, and the reported 1.82% margin over MDA VIF could easily be an artifact of fold selection. The authors should either evaluate all baselines under the same 10-fold protocol or report results on the official fixed split, and provide per-fold metrics.
  2. [Section IV-A and IV-B] Subject-level leakage is a concrete risk. D-Vlog contains 961 vlogs from 816 individuals, so at least 145 videos come from repeated subjects. The manuscript never states that the 10-fold splits are subject-disjoint. If folds are split at the vlog level, the same person's appearance, voice, and recording style can appear in both training and test, inflating F1. Given the class imbalance, a constant majority-class predictor gives F1 ≈ 0.732, and MDD-Net's 0.7707 is only about 3.9 points higher; modest leakage could fully account for this gap. The authors should group videos by subject before splitting or explicitly justify why vlog-level random splits are safe.
  3. [Section IV-B, Table I] No statistical uncertainty is reported. The paper claims 10-fold cross-validation but gives no per-fold results, standard deviations, confidence intervals, or significance tests. The closest competitor, MDA VIF, differs by only 1.82 percentage points in F1, and this margin is within the range that could arise from random fold variation. The authors should report mean ± std across folds and, ideally, a paired test against the strongest baseline under a matched protocol.
  4. [Table II and Section IV-E] The unimodal ablation results are difficult to interpret and may undermine the fusion claim. AFEM accuracy is 0.4770 and VFEM accuracy is 0.4604, both below the majority-class accuracy of approximately 0.577 (555/961) and below random for binary classification. No explanation is given for this. If the unimodal modules are undertrained or evaluated under a different protocol, the reported gains from fusion (24.90% and 26.56% in accuracy) are not meaningful. The authors should clarify the evaluation protocol for the unimodal variants and, if necessary, retrain them to a reasonable baseline.
minor comments (6)
  1. [Abstract, Section IV-D] The phrase 'percentage improvement' is used to describe absolute differences. For example, 0.7707 − 0.5970 = 0.1737 is a percentage-point difference, not a relative percentage improvement. The same applies to the 1.82% figure. Please clarify the wording throughout.
  2. [Section III-A, Equations (2)-(4)] The dimensions and combination of the content attention and positional attention outputs are not fully specified. Please define N, clarify how X^c_A and X^p_A are combined, and state the output dimension d_o in relation to d_a.
  3. [Section III-C] The values d_a = 71 and d_v = 139 are introduced as embedding sizes after feature extraction, but the input dimensions in Eq. (1) are d_a and d_v. Please use distinct notation for input vs. transformed dimensions to avoid confusion.
  4. [Table III] The 'MDD-Net' row duplicates the results already shown in Tables I and II. Consider removing or clearly labeling it as the full model.
  5. [Figure 5] The grayscale reproduction makes it hard to distinguish 'darker' from 'lighter' regions. Please use a more legible colormap or add annotations.
  6. [Section V] The conclusion appropriately acknowledges label noise in D-Vlog, but the discussion would benefit from stating how this limitation affects the interpretation of the absolute F1 score and the comparison with prior work.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MDD-Net's reported performance is a measured empirical outcome on a public benchmark, not a quantity derived from its own fitted inputs.

full rationale

The paper's central claim is an empirical F1-score of 0.7707 on the D-Vlog test partition, achieved by training a neural architecture and evaluating it on held-out data. No load-bearing step reduces by construction to its inputs: the acoustic and visual feature extractors use standard externally published components (global self-attention [26], relative position embeddings [27], hierarchical multi-head self-attention [28]); the mutual transformer is defined by explicit cross-attention equations (9)-(11); and the detection layer applies a learned classifier. There is no fitted parameter that is later renamed as a prediction - hyperparameters such as the label-smoothing factor, focal-loss weights, and network depths are tuned on the validation set, but the paper does not present those validation results as test predictions. The reference list contains no self-citations by the present authors, and no 'uniqueness theorem' is imported from the authors' prior work. The paper does contain validity concerns that the reader's analysis correctly identifies: Section IV-A describes a fixed 7:1:2 train/validation/test split, while Section IV-B states 'a 10-fold cross-validation is implemented,' so the comparison to literature numbers in Table I may not be protocol-matched; and the conclusion admits 'mislabeling is inevitable' in the D-Vlog labels. These issues affect the reliability and interpretation of the reported improvement, but they are not circularity. The derivation chain from input features to output classification is a standard supervised learning pipeline with measured test-set performance, so the appropriate circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper's empirical claim rests on the D-Vlog benchmark and its title-derived labels, the sufficiency of pre-extracted features, and the comparability of prior published results. Numerous hyperparameters (focal loss weights, smoothing, depths, heads) are chosen without reported values, so the configuration is underdetermined from the text.

free parameters (6)
  • Focal loss scaling factor phi
    Eq. (16) includes focal loss with scaling phi; value is not reported and was presumably tuned on validation, affecting precision-recall tradeoff.
  • Focal loss focusing parameter gamma
    Eq. (16) uses gamma to down-weight easy examples; value is not reported.
  • L2 regularization strength lambda
    Eq. (17) adds L2 penalty with strength lambda; value not reported, separate from Adam weight decay (0.1).
  • Label smoothing factor epsilon_s
    Eq. (15) smooths labels with epsilon_s; value not reported.
  • Transformer depth / layer counts
    Section III-B states the VFEM transformer is repeated l(j-1) times for j in {2,3,4,5}, but total depth l is unspecified. This controls model capacity.
  • Attention head count and hidden dimension do
    H-MHSA in Eq. (6) implies multiple heads but head count and output dimensions are not given; chosen by hand without reporting.
assumptions (4)
  • domain assumption D-Vlog dataset labels (inferred from video title keywords and annotated transcripts) are sufficiently valid to measure depression detection.
    Section IV-A describes labels based on title terms such as 'depression episode vlog'; conclusion admits mislabeling is inevitable. The central F1 result is only meaningful if labels are not systematically confounded with video style.
  • domain assumption The pre-extracted acoustic descriptors (25 OpenSmile low-level descriptors) and visual landmarks (68 Dlib facial landmarks) contain enough signal to distinguish depression.
    Section IV-A states these features are used; the model never sees raw audio or video. If these features omit key depression cues, the architecture cannot compensate.
  • domain assumption Prior published baseline results (Bi-LSTM, TFN, TAMFN, CAIINET, STST, MDA VIF) were obtained on the same or equivalent train/test partitions and are directly comparable.
    Section IV-D compares against reported numbers without matched experimental protocols; no error bars or significance tests are given.
  • standard math Standard transformer and attention computations (softmax attention, LayerNorm, residual connections) work as specified in the cited literature.
    The architecture relies on standard transformer blocks; the paper cites Refs. [26]-[28] for these building blocks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MDD-Net: Multimodal Depression Detection through Mutual Transformer." pith.science (2026). https://pith.science/paper/BULFRVIA

@misc{pith2026250808093,
  author       = {Pith},
  title        = {Pith review of: MDD-Net: Multimodal Depression Detection through Mutual Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BULFRVIA}},
  note         = {Machine review of arXiv:2508.08093}
}
read the original abstract

Depression is a major mental health condition that severely impacts the emotional and physical well-being of individuals. The simple nature of data collection from social media platforms has attracted significant interest in properly utilizing this information for mental health research. A Multimodal Depression Detection Network (MDD-Net), utilizing acoustic and visual data obtained from social media networks, is proposed in this work where mutual transformers are exploited to efficiently extract and fuse multimodal features for efficient depression detection. The MDD-Net consists of four core modules: an acoustic feature extraction module for retrieving relevant acoustic attributes, a visual feature extraction module for extracting significant high-level patterns, a mutual transformer for computing the correlations among the generated features and fusing these features from multiple modalities, and a detection layer for detecting depression using the fused feature representations. The extensive experiments are performed using the multimodal D-Vlog dataset, and the findings reveal that the developed multimodal depression detection network surpasses the state-of-the-art by up to 17.37% for F1-Score, demonstrating the greater performance of the proposed system. The source code is accessible at https://github.com/rezwanh001/Multimodal-Depression-Detection.

Figures

Figures reproduced from arXiv: 2508.08093 by the authors.

Figure 1
Figure 1. The overall architecture of the proposed system: a multimodal fusion technique for audio-visual feature fusion using mutual transformers for depression [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Acoustic feature extraction module: the module processes input audio features using [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visual feature extraction module: the module applies [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Transformer encoder architecture: Q, K, and V refer to input queries, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Visualization of features based on weights of the final embedding [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Feature visualization of the proposed MDD-Net through t-SNE for [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 30 canonical work pages

  1. [1]

    Major depressive disorder,

    R. H. Belmaker and G. Agam, “Major depressive disorder,”New England Journal of Medicine, vol. 358, no. 1, pp. 55–68, 2008

    work page 2008

  2. [2]

    World mental health report: Transforming mental health for all,

    World Health Organization, “World mental health report: Transforming mental health for all,” Geneva, 2022, accessed: 2025-07-15. [Online]. Available: https://www.who.int/publications/i/item/9789240049338

    arXiv 2022

  3. [3]

    Review of automated depression detection: Social posts, audio and video, open challenges and future direction,

    U. Yadav, A. K. Sharma, and D. Patil, “Review of automated depression detection: Social posts, audio and video, open challenges and future direction,”Concurrency and Computation: Practice and Experience, vol. 35, no. 1, p. e7407, 2023

    work page 2023

  4. [4]

    Multimodal machine learning for mental disorder detection: A scoping review,

    T. T. Nguyen, V . H.-Q. Pham, D.-T. Le, X.-S. Vu, F. Deligianni, and H. D. Nguyen, “Multimodal machine learning for mental disorder detection: A scoping review,”Procedia Computer Science, vol. 225, pp. 1458–1467, 2023

    work page 2023

  5. [5]

    Depressive semantic awareness from vlog facial and vocal streams via spatio- temporal transformer,

    Y . Tao, M. Yang, Y . Wu, K. Lee, A. Kline, and B. Hu, “Depressive semantic awareness from vlog facial and vocal streams via spatio- temporal transformer,”Digital Communications and Networks, vol. 10, no. 3, pp. 577–585, 2024

    work page 2024

  6. [6]

    Gender- aware estimation of depression severity level in a multimodal setting,

    S. A. Oureshi, G. Dias, S. Saha, and M. Hasanuzzaman, “Gender- aware estimation of depression severity level in a multimodal setting,” inInternational Joint Conference on Neural Networks (IJCNN), 2021, pp. 1–8

    work page 2021

  7. [7]

    A hierarchical depression detection model based on vocal and emotional cues,

    Y . Dong and X. Yang, “A hierarchical depression detection model based on vocal and emotional cues,”Neurocomputing, vol. 441, pp. 279–290, 2021

    work page 2021

  8. [8]

    Multi-modal depression detection and estimation,

    L. Yang, “Multi-modal depression detection and estimation,” in8th International Conference on Affective Computing and Intelligent Inter- action Workshops and Demos (ACIIW). IEEE, 2019, pp. 26–30

    work page 2019

Show all 33 references
  1. [9]

    Multimodal depression classifi- cation using articulatory coordination features and hierarchical attention based text embeddings,

    N. Seneviratne and C. Espy-Wilson, “Multimodal depression classifi- cation using articulatory coordination features and hierarchical attention based text embeddings,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6252–6256

    2022

  2. [10]

    Depression diagnosis and analysis via multimodal multi-order factor fusion,

    C. Yuan, X. Liu, Q. Xu, Y . Li, Y . Luo, and X. Zhou, “Depression diagnosis and analysis via multimodal multi-order factor fusion,” in International Conference on Artificial Neural Networks. Springer, 2024, pp. 56–70

    2024

  3. [11]

    Automatic depression detection: An emotional audio-textual corpus and a gru/bilstm-based model,

    Y . Shen, H. Yang, and L. Lin, “Automatic depression detection: An emotional audio-textual corpus and a gru/bilstm-based model,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6247–6251

    2022

  4. [12]

    Tamfn: time- aware attention multimodal fusion network for depression detection,

    L. Zhou, Z. Liu, Z. Shangguan, X. Yuan, Y . Li, and B. Hu, “Tamfn: time- aware attention multimodal fusion network for depression detection,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 31, pp. 669–679, 2022

    2022

  5. [13]

    D-vlog: Multimodal vlog dataset for depression detection,

    J. Yoon, C. Kang, S. Kim, and J. Han, “D-vlog: Multimodal vlog dataset for depression detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 11, 2022, pp. 12 226–12 234

    2022

  6. [14]

    An overview of tools and technologies for anxiety and depression management using ai,

    A. Pavlopoulos, T. Rachiotis, and I. Maglogiannis, “An overview of tools and technologies for anxiety and depression management using ai,”Applied Sciences, vol. 14, no. 19, p. 9068, 2024

    2024

  7. [15]

    Depression detection using artificial intelligence: A review,

    V . K. Pandey and V . K. Garg, “Depression detection using artificial intelligence: A review,”Computer Science Engineering and Emerging Technologies: Proceedings of ICCS 2022, p. 324, 2024

    2022

  8. [16]

    Comprehensive review of depression detec- tion techniques based on machine learning approach,

    S. J. Pinto and M. Parente, “Comprehensive review of depression detec- tion techniques based on machine learning approach,”Soft Computing, pp. 1–25, 2024

    2024

  9. [17]

    A multimodal approach for detection and assessment of depression using text, audio and video,

    W. Zhang, K. Mao, and J. Chen, “A multimodal approach for detection and assessment of depression using text, audio and video,”Phenomics, pp. 1–16, 2024

    2024

  10. [18]

    Ccfn: Depression detection via multimodal fusion with complex-valued capsule network,

    J. Lei, Q. Yang, B. Li, and W. Zhang, “Ccfn: Depression detection via multimodal fusion with complex-valued capsule network,” inInterna- tional Joint Conference on Neural Networks (IJCNN), 2024, pp. 1–6

    2024

  11. [19]

    Depression detection using emotion artificial intelligence,

    M. Deshpande and V . Rao, “Depression detection using emotion artificial intelligence,” inInternational Conference on Intelligent Sustainable Systems (ICISS). IEEE, 2017, pp. 858–862

    2017

  12. [20]

    Automatic depression detection using smartphone-based text-dependent speech signals: deep convolutional neural network approach,

    A. Y . Kim, E. H. Jang, S.-H. Lee, K.-Y . Choi, J. G. Park, and H.-C. Shin, “Automatic depression detection using smartphone-based text-dependent speech signals: deep convolutional neural network approach,”Journal of Medical Internet Research, vol. 25, p. e34474, 2023

    2023

  13. [21]

    Emotion recognition with eeg-based brain-computer interfaces: a systematic literature review,

    K. Erat, E. B. S ¸ahin, F. Do˘gan, N. Merdano ˘glu, A. Akcakaya, and P. O. Durdu, “Emotion recognition with eeg-based brain-computer interfaces: a systematic literature review,”Multimedia Tools and Applications, pp. 1–48, 2024

    2024

  14. [22]

    Visually interpretable repre- sentation learning for depression recognition from facial images,

    X. Zhou, K. Jin, Y . Shang, and G. Guo, “Visually interpretable repre- sentation learning for depression recognition from facial images,”IEEE Transactions on Affective Computing, vol. 11, no. 3, pp. 542–552, 2018

    2018

  15. [23]

    A com- prehensive review on deep learning-based data fusion,

    M. Hussain, M. O’Nils, J. Lundgren, and S. J. Mousavirad, “A com- prehensive review on deep learning-based data fusion,”IEEE Access, vol. 12, pp. 180 093–180 124, 2024

    2024

  16. [24]

    Detecting depression on video logs using audiovisual features,

    K. Min, J. Yoon, M. Kang, D. Lee, E. Park, and J. Han, “Detecting depression on video logs using audiovisual features,”Humanities and Social Sciences Communications, vol. 10, no. 1, pp. 1–8, 2023

    2023

  17. [25]

    Mdavif: A multi-domain acoustical- visual information fusion model for depression recognition from vlog data,

    T. Ling, D. Chen, and B. Li, “Mdavif: A multi-domain acoustical- visual information fusion model for depression recognition from vlog data,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 8115–8119

    2024

  18. [26]

    Global self- attention networks for image recognition,

    Z. Shen, I. Bello, R. Vemulapalli, X. Jia, and C.-H. Chen, “Global self- attention networks for image recognition,”arXiv:2010.03019, 2020

    2010 arXiv

  19. [27]

    Attention augmented convolutional networks,

    I. Bello, B. Zoph, A. Vaswani, J. Shlens, and Q. V . Le, “Attention augmented convolutional networks,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 3286–3295

    2019

  20. [28]

    Vision transformers with hierarchical attention,

    Y . Liu, Y .-H. Wu, G. Sun, L. Zhang, A. Chhatkuli, and L. Van Gool, “Vision transformers with hierarchical attention,”Machine Intelligence Research, pp. 1–14, 2024

    2024

  21. [29]

    When does label smoothing help?

    R. M ¨uller, S. Kornblith, and G. E. Hinton, “When does label smoothing help?”Advances in Neural Information Processing Systems, vol. 32, 2019

    2019

  22. [30]

    t-sne visualization of large-scale neural recordings,

    G. Dimitriadis, J. P. Neto, and A. R. Kampff, “t-sne visualization of large-scale neural recordings,”Neural computation, vol. 30, no. 7, pp. 1750–1774, 2018

    2018

  23. [31]

    A multi-modal hierarchical recurrent neural network for depression detection,

    S. Yin, C. Liang, H. Ding, and S. Wang, “A multi-modal hierarchical recurrent neural network for depression detection,” inProceedings of the 9th International on Audio/Visual Emotion Challenge and Workshop, 2019, pp. 65–71

    2019

  24. [32]

    Tensor fusion network for multimodal sentiment analysis,

    A. Zadeh, M. Chen, S. Poria, E. Cambria, and L.-P. Morency, “Tensor fusion network for multimodal sentiment analysis,” inProceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, 2017, pp. 1103–1114

    2017

  25. [33]

    Caiinet: Neural network based on contextual attention and information interaction mechanism for depression detection,

    L. Zhou, Z. Liu, X. Yuan, Z. Shangguan, Y . Li, and B. Hu, “Caiinet: Neural network based on contextual attention and information interaction mechanism for depression detection,”Digital Signal Processing, vol. 137, p. 103986, 2023

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.