Pith. sign in

REVIEW 4 major objections 5 minor 30 references

D-CAT: Decoupled Cross-Attention Transfer between Sensor Modalities for Unimodal Inference

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A cross-attention loss transfers knowledge between sensor modalities during training, allowing a model to run on a single sensor at inference and improving F1 scores by up to 10 percentage points.

desk verdict The paper's central loss derivation doesn't hold up: Eq. 6 does not reduce to Eq. 7, and the proof fails, so the 'cross-attention transfer' mechanism is not established — but the underlying problem and code are real enough to warrant peer review. read the letter →

arxiv 2509.09747 v1 pith:IC3V6EYG submitted 2025-09-11 cs.LG cs.AIcs.RO

Leen Daher , Zhaobo Wang , Malcolm Mielle This is my paper · ORCID
classification cs.LGcs.AIcs.RO
keywords cross-modaltransferlearningself-attentioncross-attentionlosshumanactivityrecognitionsensormodalitiesunimodalinferenceknowledgedistillationFrobeniusnorm
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 proposes D-CAT, a training framework that aligns the feature spaces of two sensor modalities using a cross-attention loss, while keeping each modality's classifier independent. The goal is to let a model trained on paired multi-sensor data run at deployment on a single sensor. On three human-activity datasets (IMU, video, audio), the method improves F1 scores for weaker modalities in in-distribution settings, and in out-of-distribution settings even weaker source modalities can boost stronger targets. If it holds, perception systems could get multi-sensor training benefits without the hardware cost of carrying all sensors at inference.

What carries the argument

The cross-attention loss L_CA = ||K_B^T V_B - K_A^T V_A||_F, computed from the key and value projections of the target (B) and frozen source (A) self-attention modules, combined with an indicator mask that keeps only correctly classified source samples. The paper also relies on Theorem 4.1, a matrix-factorization result, to argue that minimizing this soft loss drives the target's key and value matrices toward linear mappings of the source's, preserving alignment without softmax or scaling.

What would settle it

Measure the actual distance ||Q_B K_B^T V_B - Q_B K_A^T V_A||_F between the true cross-attention outputs before and after training. If the loss L_CA shrinks while this true distance does not, or if removing Q_B from the loss changes results drastically, the claimed alignment mechanism is not what is doing the work.

Watch

Extended reading notes

Core claim

D-CAT's central claim is that a well-chosen loss between attention key-value products can transfer knowledge between modalities without coupling their classifiers. Specifically, the paper argues that minimizing ||K_B^T V_B - K_A^T V_A||_F on correctly classified source samples, with the source frozen, aligns the target's self-attention features to the source's, so the target network learns from the source's embeddings. The authors show empirically that this improves the weaker modality's F1 by up to 10 percentage points in in-distribution tests and can help even strong targets under distribution shift. They justify dropping softmax and scaling via a rank-factorization theorem, and mask out m

Load-bearing premise

The load-bearing assumption is that the target query matrix can be factored out of the alignment loss, which holds only if the query is orthogonal or the norm ignores it, and the supporting theorem applies only to exact matrix equality, not to a soft penalty.

Editorial extensions

If this is right

  • If D-CAT works, robots and wearables can train on rich multi-sensor data but deploy with a single cheap sensor, cutting cost and power.
  • The decoupled design means each modality keeps its own classifier, so adding or removing sensors does not require retraining the whole system.
  • Masking by source correctness suggests that the quality of knowledge transfer depends on the source model's reliability, guiding when to trust cross-modal alignment.
  • In out-of-distribution settings, weak-to-strong transfer can improve generalization, implying D-CAT could help with domain shift without adding sensors.

Reading between the lines

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

  • The formal gap in Eq. (7) suggests the loss may actually be aligning unnormalized key-value products rather than true cross-attention outputs; if the query matrix is near-orthogonal in practice, the practical effect could still hold, but this should be tested.
  • The masking rule could be extended to confidence-weighted alignment instead of a hard 0/1 mask, potentially improving transfer on uncertain samples.
  • The framework's claim that it generalizes to arbitrary sensor types is plausible but untested beyond IMU, video, and audio; a stress test on, say, radar or depth sensors would clarify the scope.
  • The authors note overfitting limits out-of-distribution gains; a natural next step, which they mention, is using multiple sources or regularizing the target to reduce overfitting.
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 / 5 minor

Summary. The paper proposes D-CAT, a training-time cross-modal transfer framework for human activity recognition that aligns the feature spaces of a frozen source modality and a learned target modality through a cross-attention loss. The target model is trained with a weighted sum of cross-entropy and the proposed loss, optionally masked to consider only source-correctly-classified samples. Experiments on three two-modality datasets (video-IMU, audio-IMU, audio-video) compare against uni-modal baselines in both in-distribution (ID) and out-of-distribution (OOD) settings. The paper reports F1-score gains when transferring from a higher-performing source to a lower-performing target in ID settings, and reports some OOD gains in the reverse direction. The claimed contribution is a decoupled loss that allows single-modality inference while capturing multi-modal training benefits.

Significance. If the claimed mechanism were correct, D-CAT would be a practically useful contribution: it addresses a real deployment constraint (single-sensor inference with multi-modal training) and the code is publicly available. The paper also includes ablations on the masking strategy and the transfer weight. However, the central theoretical derivation is invalid: the Frobenius-norm reduction from Eq. (6) to Eq. (7) is not justified, and Theorem 4.1 does not repair the flaw. The actual loss being optimized is an unexamined key-value Gram-matrix alignment term, not the cross-attention distance the paper claims. This undermines the core contribution. Empirical results also lack error bars, and the headline improvement is stated inconsistently. The contribution, as presented, is therefore not established.

major comments (4)
  1. [Section IV-B, Eq. (6) to Eq. (7)] The derivation claiming L_CA = ||Q_B K_B^T V_B - Q_B K_A^T V_A||_F factors out Q_B to obtain ||K_B^T V_B - K_A^T V_A||_F. This is valid only if left multiplication by Q_B preserves the Frobenius norm, i.e., if Q_B is orthogonal. No such constraint is imposed: Q_B = E_B W_Q with W_Q an unconstrained learned matrix (Eq. 1). In general ||Q_B X||_F != ||X||_F and the gradients w.r.t. K_B and V_B differ by the metric Q_B^T Q_B. The optimized objective is therefore not the cross-attention distance claimed in Eq. (6). This is a load-bearing flaw because the paper's central mechanism is introduced as a cross-attention alignment loss.
  2. [Theorem 4.1 and surrounding text] Theorem 4.1 does not justify dropping the softmax and 1/sqrt(d_out) scaling from Eq. (2) when moving to Eq. (7). The theorem addresses exact matrix factorization M = AB = A'B' and asserts existence of linear maps R,S with A'=AR and B'=SB. It says nothing about approximate equality under a soft Frobenius penalty, and it does not imply that K_B^T V_B converges to K_A^T V_A. Even in the exact case, the theorem does not provide R,S with RS = I. Thus the claim that 'as proven in Theorem 4.1, K_B converges to a linear mapping of K_A ... ensuring that the alignment between modalities is preserved even without softmax and scaling' is not supported.
  3. [Tables I-III and abstract/conclusion] All reported results are single runs (one seed per dataset in Table I) with no error bars or repeated-seed statistics. Given the modest gains (e.g., Cough-IMU F1 from 0.205 to 0.256) and occasional slight losses, the empirical claims are not established at the reported precision. Additionally, the abstract states 'up to 10% F1-score gains' in ID settings, while the conclusion says 'up to 7%'. This discrepancy needs to be resolved, and the numbers should be reported with variance.
  4. [Section IV-C and Tables II-III] The Masked Cross-Modal Alignment (MCMA) is presented as a component of the method, but Tables II and III show that in the OOD setting it is sometimes worse than the unmasked variant (e.g., Cough-Audio and VGGSound-Audio). The authors conclude that MCMA 'should be used on a per-dataset basis.' This is a post-hoc selection effect: the reported 'Ours' results are chosen after noting which masking variant performs better. The paper should pre-specify the masking strategy or report both variants without selecting per dataset, otherwise the headline improvements are not well-defined.
minor comments (5)
  1. [Section IV-B, Eq. (6)] The phrase 'where ||·|| denotes the normalized matrices' is confusing. Eq. (5) defines the Frobenius norm of a matrix, not a normalization of matrices. Please clarify.
  2. [Table II and Section VI-B] The acronym is inconsistent: the method is called 'MCMA' in Section IV-C, but Section VI-B and Tables II/III refer to 'MMCA' and 'no MMCA.' Please standardize.
  3. [Section VI-A.2] 'OD scenarios' should be 'OOD scenarios'; the abbreviation is introduced as OOD but used as OD in several places.
  4. [Overall] The paper does not compare against any prior cross-modal transfer or feature-alignment baseline (e.g., contrastive losses such as CoCoA or simple feature-space alignment). Comparisons only against uni-modal baselines leave open the question of whether the specific proposed loss is necessary for the observed gains.
  5. [Theorem 4.1 proof] The proof contains loose statements, e.g., 'If that is not the case, then A' cannot produce the columns of M' is not a formal argument, and the notation in Eq. (10) has an indexing inconsistency (r_rj vs. r_kj). This should be rewritten.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: D-CAT's loss and evaluation are not equivalent to their inputs by construction.

full rationale

The paper's claimed derivation chain is not circular. The cross-attention loss L_CA (Eqs. 6-7, 17) is an explicit, fixed objective defined on the source and target key/value matrices; it is not fitted to the reported F1 outcomes, and it demonstrably fails to help in several reported rows (e.g., Table II Image and Cough-Audio, Table III UESTC-IMU), so the central empirical claim is not statistically forced by construction. The only self-citation is [6], used as a pointer for ID/OOD split conventions, but the paper defines both splits in the same paragraph (Section VI), so this citation is not load-bearing. The masked-alignment ablation is reported for both variants, and the paper explicitly states that in OOD it 'should be used on a per-dataset basis', which is a transparent model-selection caveat rather than a fitted input renamed as a prediction. The mathematical objections to Eq. 7 (that factoring out Q_B requires orthogonality, and that Theorem 4.1 concerns exact factorization rather than a soft Frobenius penalty) concern the validity and rigor of the derivation, not circularity: an invalid or unsupported step is not the same as the conclusion being equivalent to the premise. The paper is self-contained against external uni-modal baselines and does not rely on a self-citation chain to establish its central claim, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the mathematically false factoring of Q_B out of the norm, and on the unproven assertion that softmax/scaling can be dropped. No entirely new physical or mathematical entities are introduced; the new 'entity' is the loss function itself, which is a method, not an object.

free parameters (3)
  • lambda (transfer weight) = 1.0 for main tables; ablation shows best varies (0.01-10) per dataset
    Hyperparameter balancing classification and alignment losses, chosen via validation; not fitted to test, but varies across datasets.
  • MCMA mask on/off = On for main ID results; off or on per dataset in OOD
    The paper switches masking on/off per dataset depending on which achieves better metrics, a post-hoc modeling choice.
  • Per-dataset hyperparameters (learning rate, dropout, window/stride, epochs, seed) = Table I lists values per dataset
    Tuned via grid search on validation; standard hyperparameter selection, not a fitted constant of the method itself.
assumptions (4)
  • ad hoc to paper The query matrix Q_B can be factored out of the Frobenius norm without changing the loss
    Used to derive Eq. (7) from Eq. (6). Only valid if Q_B is orthogonal or the norm is unitarily invariant, which is not guaranteed for a learned projection.
  • ad hoc to paper Dropping softmax and scaling from attention preserves alignment quality
    Asserted after Theorem 4.1, but the theorem only states that if two matrix products are equal, factors are linearly related; it does not show that softmax(Q K_A^T) V_A and softmax(Q K_B^T) V_B are close under the loss.
  • domain assumption Frozen source representations are informative for the target modality
    The entire transfer relies on the source K and V providing useful alignment targets; no joint optimization or adaptation of the source is performed.
  • standard math Self-attention and cross-attention definitions from Vaswani et al. are standard
    Used in Preliminaries, Eqs. (1)-(3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of D-CAT: Decoupled Cross-Attention Transfer between Sensor Modalities for Unimodal Inference." pith.science (2026). https://pith.science/paper/IC3V6EYG

@misc{pith2026250909747,
  author       = {Pith},
  title        = {Pith review of: D-CAT: Decoupled Cross-Attention Transfer between Sensor Modalities for Unimodal Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IC3V6EYG}},
  note         = {Machine review of arXiv:2509.09747}
}
read the original abstract

Cross-modal transfer learning is used to improve multi-modal classification models (e.g., for human activity recognition in human-robot collaboration). However, existing methods require paired sensor data at both training and inference, limiting deployment in resource-constrained environments where full sensor suites are not economically and technically usable. To address this, we propose Decoupled Cross-Attention Transfer (D-CAT), a framework that aligns modality-specific representations without requiring joint sensor modality during inference. Our approach combines a self-attention module for feature extraction with a novel cross-attention alignment loss, which enforces the alignment of sensors' feature spaces without requiring the coupling of the classification pipelines of both modalities. We evaluate D-CAT on three multi-modal human activity datasets (IMU, video, and audio) under both in-distribution and out-of-distribution scenarios, comparing against uni-modal models. Results show that in in-distribution scenarios, transferring from high-performing modalities (e.g., video to IMU) yields up to 10% F1-score gains over uni-modal training. In out-of-distribution scenarios, even weaker source modalities (e.g., IMU to video) improve target performance, as long as the target model isn't overfitted on the training data. By enabling single-sensor inference with cross-modal knowledge, D-CAT reduces hardware redundancy for perception systems while maintaining accuracy, which is critical for cost-sensitive or adaptive deployments (e.g., assistive robots in homes with variable sensor availability). Code is available at https://github.com/Schindler-EPFL-Lab/D-CAT.

Figures

Figures reproduced from arXiv: 2509.09747 by the authors.

Figure 1
Figure 1. Cross-modal transfer architecture. Sensor modality A (source) and B (target) classification networks each consist of an encoder and a self-attention module. Modality A network is pretrained and frozen during the training of modality B’s classification network. The classification network of modality B is then trained against a standard classification loss as well as a novel cross-attention loss that aims to align the… view at source ↗
Figure 4
Figure 4. Audio encoder architecture. The model processes augmented spectrograms through two consecutive convolu￾tional blocks, followed by a sequence of pooling and dropout layers repeated six times, and finally a fully connected layer with ReLU and dropout. C. Implementation Details Unless otherwise stated, models are trained using the Adam optimizer with an attention size of 512 and ReLU activation functions. Dropout and w… view at source ↗
Figure 3
Figure 3. Image encoder architecture. The model begins with an initial convolution block and pooling layer, followed by two stacked convolution blocks. A residual block, repeated 33 times, combines pairs of 2D convolutions with batch normalization and a ReLU activation. follows the Somasa architecture, the specific parameters of the convolutional and subsequent layers (kernel sizes, strides, and number of filters) are adjuste… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 4 linked inside Pith

  1. [1]

    Human activity recognition from 3d data: A review,

    J. K. Aggarwal and L. Xia, “Human activity recognition from 3d data: A review,”Pattern Recognition Letters, vol. 48, pp. 70–80, 2014

    2014

  2. [2]

    A comprehensive survey of rgb-based and skeleton-based human action recognition,

    C. Wang and J. Yan, “A comprehensive survey of rgb-based and skeleton-based human action recognition,”IEEE Access, vol. 11, pp. 53 880–53 898, 2023

    2023

  3. [3]

    A review of human activity recognition methods,

    M. Vrigkas, C. Nikou, and I. A. Kakadiaris, “A review of human activity recognition methods,”Frontiers in Robotics and AI, vol. 2, p. 28, 2015

    2015

  4. [4]

    Human-robot interactions in manufacturing: A survey of hu- man behavior modeling,

    R. Jahanmahin, S. Masoud, J. Rickli, and A. Djuric, “Human-robot interactions in manufacturing: A survey of hu- man behavior modeling,”Robotics and Computer-Integrated Manufacturing, vol. 78, p. 102 404, 2022,ISSN: 0736-5845

    2022

  5. [5]

    Semantic- based surveillance video retrieval,

    W. Hu, D. Xie, Z. Fu, W. Zeng, and S. Maybank, “Semantic- based surveillance video retrieval,”IEEE Transactions on Image Processing, vol. 16, no. 4, pp. 1168–1181, 2007

    2007

  6. [6]

    UAC: uncertainty- aware calibration of neural networks for gesture detection,

    F. Al Haddad, Y . Wang, and M. Mielle, “UAC: uncertainty- aware calibration of neural networks for gesture detection,” arXiv preprint arXiv:2504.02895, 2025

    arXiv 2025

  7. [7]

    Combing rgb and depth map features for human activity recognition,

    Y . Zhao, Z. Liu, L. Yang, and H. Cheng, “Combing rgb and depth map features for human activity recognition,” in Proceedings of the 2012 Asia Pacific signal and information processing association annual summit and conference, IEEE, 2012, pp. 1–4

    2012

  8. [8]

    Audio-based activities of daily living (adl) recognition with large-scale acoustic embeddings from online videos,

    D. Liang and E. Thomaz, “Audio-based activities of daily living (adl) recognition with large-scale acoustic embeddings from online videos,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 3, no. 1, pp. 1–18, Mar. 2019,ISSN: 2474-9567

    2019

Show all 30 references
  1. [9]

    Skeleton-based human action recog- nition in low-resolution infrared images,

    L. Min and B. Yang, “Skeleton-based human action recog- nition in low-resolution infrared images,” in2023 IEEE 18th Conference on Industrial Electronics and Applications (ICIEA), 2023, pp. 1363–1368

    2023

  2. [10]

    A deeply coupled con- vnet for human activity recognition using dynamic and rgb images,

    T. Singh and D. K. Vishwakarma, “A deeply coupled con- vnet for human activity recognition using dynamic and rgb images,”Neural Computing and Applications, vol. 33, no. 1, pp. 469–485, 2021

    2021

  3. [11]

    Human activity recognition in rgb-d videos by dynamic images,

    S. Mukherjee, L. Anvitha, and T. M. Lahari, “Human activity recognition in rgb-d videos by dynamic images,”Multimedia Tools and Applications, vol. 79, no. 27, pp. 19 787–19 801, 2020

    2020

  4. [12]

    Audio-and video-based human activity recognition systems in healthcare,

    S. Cristina, V . Despotovic, R. P ´erez-Rodr´ıguez, and S. Aleksic, “Audio-and video-based human activity recognition systems in healthcare,”IEEE Access, vol. 12, pp. 8230–8245, 2024

    2024

  5. [13]

    Recognition of human activities based on ambient audio and vibration data,

    M. Koch, T. Pfitzinger, F. Schlenke, F. Kohlmorgen, R. Groll, and H. W ¨ohrle, “Recognition of human activities based on ambient audio and vibration data,”IEEE Access, vol. 12, pp. 174 399–174 412, 2024

    2024

  6. [14]

    Design of human activity recognition algorithms based on a single wearable imu sensor,

    W. Zhuang, Y . Chen, J. Su, B. Wang, and C. Gao, “Design of human activity recognition algorithms based on a single wearable imu sensor,”International Journal of Sensor Net- works, vol. 30, no. 3, pp. 193–206, 2019

    2019

  7. [15]

    A study on imu-based human activity recognition using deep learning and traditional machine learning,

    C. Hou, “A study on imu-based human activity recognition using deep learning and traditional machine learning,” in 2020 5th International Conference on Computer and Com- munication Systems (ICCCS), IEEE, 2020, pp. 225–234

    2020

  8. [16]

    An efficient human activity recognition framework based on wearable imu wrist sensors,

    A. Ayman, O. Attalah, and H. Shaban, “An efficient human activity recognition framework based on wearable imu wrist sensors,” in2019 IEEE International Conference on Imaging Systems and Techniques (IST), IEEE, 2019, pp. 1–5

    2019

  9. [17]

    Multisense: Cross-labelling and learn- ing human activities using multimodal sensing data,

    L. Zhang et al., “Multisense: Cross-labelling and learn- ing human activities using multimodal sensing data,”ACM Transactions on Sensor Networks, vol. 19, no. 3, pp. 1–26, Apr. 2023,ISSN: 1550-4867

    2023

  10. [18]

    W. Ma, S. Li, L. Cai, and J. Kang,Learning modality knowl- edge alignment for cross-modality transfer, 2024. arXiv: 2406.18864 [cs.CV]

    2024 arXiv

  11. [19]

    Kwon et al.,Imutube: Automatic extraction of virtual on- body accelerometry from video for human activity recogni- tion, 2020

    H. Kwon et al.,Imutube: Automatic extraction of virtual on- body accelerometry from video for human activity recogni- tion, 2020. arXiv:2006.05675 [cs.CV]

    2020 arXiv

  12. [20]

    Class-aware sample weight learning for cross-modal unsupervised domain adap- tation in cross-user wearable human activity recognition,

    T. Wu, Y . Liu, and S. Yongchareon, “Class-aware sample weight learning for cross-modal unsupervised domain adap- tation in cross-user wearable human activity recognition,” in ECAI 2024. IOS Press, Oct. 2024,ISBN: 9781643685489

    2024

  13. [21]

    Hu- man: Complex activity recognition with multi-modal multi- positional body sensing,

    P. Bharti, D. De, S. Chellappan, and S. K. Das, “Hu- man: Complex activity recognition with multi-modal multi- positional body sensing,”IEEE Transactions on Mobile Computing, vol. 18, no. 4, pp. 857–870, Apr. 2019,ISSN: 2161-9875

    2019

  14. [22]

    Cocoa: Cross modality contrastive learning for sensor data,

    S. Deldari, H. Xue, A. Saeed, D. V . Smith, and F. D. Salim, “Cocoa: Cross modality contrastive learning for sensor data,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 6, no. 3, pp. 1–28, Sep. 2022,ISSN: 2474-9567

    2022

  15. [23]

    Multi- modality cross attention network for image and sentence matching,

    X. Wei, T. Zhang, Y . Li, Y . Zhang, and F. Wu, “Multi- modality cross attention network for image and sentence matching,” in2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 10 938– 10 947

    2020

  16. [24]

    Cocoa: Cross modality contrastive learning for sensor data,

    S. Deldari, H. Xue, A. Saeed, D. V . Smith, and F. D. Salim, “Cocoa: Cross modality contrastive learning for sensor data,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 6, no. 3, pp. 1–28, 2022

    2022

  17. [25]

    Attention is all you need,

    A. Vaswani et al., “Attention is all you need,” inAdvances in Neural Information Processing Systems, I. Guyon et al., Eds., vol. 30, Curran Associates, Inc., 2017

    2017

  18. [26]

    Towards continual egocentric activity recog- nition: A multi-modal egocentric activity dataset for con- tinual learning,

    L. Xu et al., “Towards continual egocentric activity recog- nition: A multi-modal egocentric activity dataset for con- tinual learning,”IEEE Transactions on Multimedia, vol. 26, pp. 2430–2443, 2024,ISSN: 1941-0077

    2024

  19. [27]

    Robust multimodal cough and speech detec- tion using wearables: A preliminary analysis,

    Y . Chen et al., “Robust multimodal cough and speech detec- tion using wearables: A preliminary analysis,” in2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2024, pp. 1–6

    2024

  20. [28]

    H. Chen, W. Xie, A. Vedaldi, and A. Zisserman,Vggsound: A large-scale audio-visual dataset, 2020. arXiv:2004.14368 [cs.CV]

    2020 arXiv

  21. [29]

    Samosa: Sensing activities with motion and subsampled audio,

    V . Mollyn, K. Ahuja, D. Verma, C. Harrison, and M. Goel, “Samosa: Sensing activities with motion and subsampled audio,”Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 6, no. 3, Sep. 2022

    2022

  22. [30]

    Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumbley,Panns: Large-scale pretrained audio neural net- works for audio pattern recognition, 2020. arXiv:1912 . 10211 [cs.SD]

    2020

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.