Pith. sign in

REVIEW 5 major objections 6 minor 24 references

Task and Perception-aware Distributed Source Coding for Correlated Speech under Bandwidth-constrained Channels

T0 review · 5 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a neural distributed PCA-aided autoencoder with task- and perception-aware losses achieves up to 52% higher PSNR than naive distributed speech coding under bandwidth constraints, approaching the joint-encoder upper…

desk verdict Plausible NDPCA-for-speech extension, but the cosine loss inconsistency and missing experimental details undermine the central correlation-exploitation claim. read the letter →

arxiv 2501.17879 v1 pith:5JNQVPKC submitted 2025-01-20 cs.IT cs.AIcs.SDeess.ASeess.SPmath.IT

Sagnik Bhattacharya , Muhammad Ahmed Mohsin , Ahsan Bilal , John M. Cioffi This is my paper · ORCID
classification cs.ITcs.AIcs.SDeess.ASeess.SPmath.IT MSC 94A3494A1568T07
keywords distributedsourcecodingneuralPCAspeechenhancementperception-awarelossrate-distortion-perceptiontrade-offdynamicbandwidthautoencoderwirelessAR/VR
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

This paper tries to establish that distributed speech source coding can be made practically effective for bandwidth-constrained wireless links by combining neural distributed PCA with a task- and perception-aware loss. It claims that when several microphones record correlated speech, the proposed NDPCA-aided encoder-decoder outperforms naive per-source autoencoders by 19% in task-agnostic PSNR and by 52% when a downstream speech-enhancement task and perceptual realism are optimized, approaching the joint-encoder upper bound especially at low bandwidth. The paper also derives the uplink bitrate from channel state information so the model adapts to changing bandwidth without retraining. If true, this gives a concrete path for AR/VR telephony and multi-device speech pipelines where correlated recordings must be compressed in real time. A rate-distortion-perception trade-off curve is provided so operators can choose how much realism to sacrifice for distortion.

What carries the argument

Neural distributed PCA (NDPCA): after each source encoder produces a latent vector, PCA is applied per source and the B components with largest correlation across all sources are selected and transmitted, giving dynamic output dimensions that track channel capacity. The loss function couples this with a cosine-similarity term that penalizes inter-source correlation, a denoising score-matching loss from a score-based generative speech-enhancement model applied after decoding, and an MS-STFT discriminator perceptual loss that keeps reconstructed speech realistic.

What would settle it

Ablation: train the proposed model with and without the cosine-similarity loss, then measure the correlation among the PCA-selected components and the PSNR gap over the naive baseline; if removing the cosine term does not change the gap and the selected components are near-orthogonal, the claimed NDPCA mechanism is not what drives the results.

Watch

Extended reading notes

Core claim

The central claim is that a distributed autoencoder whose per-source latent outputs are projected through neural distributed PCA, with a training loss that combines reconstruction error, a score-based speech-enhancement task loss, and a multi-scale STFT discriminator perceptual loss, achieves significantly higher reconstructed speech PSNR than naive per-source autoencoders under the same total bandwidth. In the task-agnostic setting the proposed E2D1 configuration reaches about 91% of the joint E1D1 upper bound, and in the task-aware setting about 98.1%, with the largest gains at low bandwidth. The paper presents these numbers as evidence that inter-source correlation can be exploited distributively without a shared encoder.

Load-bearing premise

The approach assumes that a training loss which penalizes correlation between source latents can coexist with a distributed-PCA step that needs correlated components to select; if the loss decorrelates the latents, the proposed advantage has no mechanism.

Editorial extensions

If this is right

  • Under bandwidth constraints, the NDPCA-aided distributed codec outperforms naive per-source autoencoders by 19% (task-agnostic) and 52% (task-aware) in average PSNR.
  • The proposed E2D1 variant approaches the joint E1D1 upper bound: 91% of its PSNR task-agnostically and 98.1% task-aware, especially at low bandwidth.
  • Channel-state-information-driven bitrate allocation lets the system vary encoder output dimensions at inference time without retraining for each output size.
  • The rate-distortion-perception trade-off curve shows that higher perceptual weights increase distortion at fixed bandwidth, giving operators a tunable knob for realism versus fidelity.
  • Task-aware training with the speech-enhancement loss substantially raises PSNR over task-agnostic reconstruction at the same bandwidth.

Reading between the lines

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

  • The cosine-similarity loss and the NDPCA component-selection step pull in opposite directions: the loss drives latent vectors toward orthogonality while the PCA step needs correlated components to exploit; ablating the cosine term and measuring the correlation of the selected components would test which side carries the result.
  • The 52% task-aware gain is reported against naive distributed baselines, not against a joint encoder with task loss; the meaningful comparison for practitioners is whether distributed gains persist when all baselines are trained with the same task and perception losses.
  • A direct extension would apply the same NDPCA-plus-task-loss scheme to other correlated modalities, such as multi-camera video or multi-sensor audio plus motion, where the downstream task defines the distortion metric.
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

5 major / 6 minor

Summary. The paper proposes a neural distributed source coding scheme for correlated speech sources, combining distributed PCA (NDPCA) with task-aware and perception-aware losses to adaptively allocate bandwidth based on channel state information. The authors report 19% higher PSNR in task-agnostic settings and 52% higher PSNR in task-aware settings over naive distributed autoencoder baselines, and claim near-optimal performance relative to a joint-encoder 'upper bound.' The method uses variable-rate latent dimensions selected via distributed principal components and a loss comprising MSE, cosine similarity, spectral SNR, PSNR, a speech-enhancement task loss, and a perceptual loss.

Significance. If the reported results were reliable, the work would address a practically relevant gap: enabling dynamic bitrate adaptation for distributed speech coding with correlated sources while balancing task performance and perceptual realism. The paper is one of the first to apply learned distributed source coding to speech with a downstream enhancement task, and the use of CSI-derived bandwidth constraints is timely. However, the significance is substantially undercut by internal inconsistencies in the loss formulation, an unspecified core selection algorithm, a confounded task-aware comparison, and the absence of statistical validation. The claimed gains cannot currently be attributed to the proposed mechanism.

major comments (5)
  1. [Section 4.1, Eq. (11)] The cosine similarity loss is defined as Lcos = 1 - E[sum_{i,j} z_i . z_j / (||z_i|| ||z_j||)], which is minimized by increasing inter-source correlation, not by penalizing it as the text states. Tables 2 and 3 report 'Cosine Loss' values of 0.971 (dim 8) and 0.767 (dim 256) for E4D1; under the stated objective these imply near-orthogonal latents, leaving almost no correlated structure for the distributed PCA to exploit. This directly contradicts the mechanism claimed to produce the PSNR gains. The authors must clarify the actual implemented objective, define the total loss combining all terms, and state what the reported cosine values represent.
  2. [Section 3.2] The distributed PCA selection procedure is underspecified. The paper states that PCA (via SVD) is applied to each source's latent vector and that 'the B maximum correlation components from across all PCA components across all the sources' are selected, but it does not define the optimization problem, the measure of correlation between components, the selection algorithm, or how the decoder handles variable concatenation. Without this, the claimed 18.8% and 16.1% improvements over naive distributed baselines cannot be reproduced or attributed to NDPCA.
  3. [Section 5.3, Fig. 3c] The task-aware comparison is confounded by the downstream speech enhancement model. The pipeline applies a score-based SGMSE model to the decoder output before measuring PSNR, so the 52.2% improvement over the task-agnostic case may result from the enhancement model's denoising rather than from the proposed distributed coding or loss. A controlled ablation is needed, such as running the same enhancement model on the task-agnostic decoder outputs and comparing under identical conditions.
  4. [Section 5.3] The 'theoretical upper bound' (Joint E1D1) is not a theoretical bound; it is the authors' own baseline in which all source data are concatenated at a single encoder. No argument establishes that this architecture achieves the maximum possible PSNR for the given problem. Consequently, the statements that the proposed method reaches 91% (task-agnostic) and 98.1% (task-aware) of this 'upper bound' are self-referential and do not demonstrate near-optimality.
  5. [Section 5.2 and 5.3] Experimental reproducibility is inadequate for the headline claims. All results use a single fixed seed (seed=0), and no standard deviations, confidence intervals, or significance tests are reported for the PSNR gains of 18.8%, 16.1%, 52.2%, and 46.9%. Furthermore, the total loss combining MSE, cosine, spectral SNR, PSNR, task, and perceptual terms is never written down, so it is impossible to determine how the components were weighted or to re-run the experiments.
minor comments (6)
  1. [Section 3.1, Eq. (3)] The worst-case Gaussian assumption for encoder outputs is introduced without justification; the rate-distortion bound in Eq. (1) applies to i.i.d. Gaussian scalar sources, and its application to arbitrary learned latents of dimension l_s,t is not defended.
  2. [Section 4.2, Eq. (17)] The speech enhancement loss is presented as the denoising score matching loss for the score network, but the role of this loss in training the encoders/decoders is not explained, especially whether gradients flow through the pre-trained score model and how the enhancement output is differentiated.
  3. [Section 4.3] The perceptual loss for the generator is not explicitly defined; Eq. (22) is the discriminator loss, and the text only mentions minimizing 'the difference of their logits' without giving the generator objective or the relative weight of the perceptual term.
  4. [General] Table 1 appears empty in the submitted manuscript; it is referenced as 'Comparison of task-agnostic and task-aware loss' but contains no data.
  5. [General] There are several typos and inconsistencies: 'Spectogram' in Figure 1, 'to simply things' in Section 3.1, 'asign' in the Conclusion, and the abstract's '19%' versus Section 5.3's reported 18.8% and 16.1% gains (the abstract should clarify how the average is computed).
  6. [References] Reference 'han Li et al. 2023' should be formatted as 'P. Han Li et al.' or 'Li et al.' to match standard citation style; the duplicated Welker, Richter, and Gerkmann 2022a/2022b entries should be consolidated or distinguished.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the PSNR claims are empirical comparisons against in-paper baselines, and the distributed-PCA mechanism is imported from non-overlapping prior work and benchmarked internally.

full rationale

The paper's central claims are experimental measurements: the 19% and 52% PSNR gains and the 91% and 98.1% approach-to-upper-bound figures are read off Figures 3b and 3c, not derived algebraically from the method's own definitions. The bandwidth-allocation formula in Eqs. (1)-(3) is standard Gaussian rate-distortion theory; it contains no fitted parameter that is later reported as a prediction. The NDPCA mechanism is adopted from han Li et al. (2023), a prior work whose author list does not overlap with the present paper, and the paper internally benchmarks the mechanism by comparing DPCA E4D1/E2D1 against naive E4D1/E2D1, so the import is not an unverified self-citation. The phrase 'theoretical upper bound' is a misnomer in that the Joint E1D1 comparator is the authors' own empirical baseline rather than a proven information-theoretic bound, but the closeness to that comparator is an experimental result, not a construction. The cosine-similarity loss in Eq. (11) is textually described as penalizing high correlation while the displayed expression is minimized by increasing correlation, and the full combined loss is never written; these are internal-consistency and reproducibility concerns, not circularity. No equation in the paper is defined in terms of the result it is used to predict, and no fitted quantity is renamed as a prediction. The derivation chain is therefore self-contained with respect to circularity.

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

The central claim rests on the rate-distortion budget formula, which uses the Gaussian worst-case assumption and an unspecified distortion target D, plus a training procedure whose full loss function and weights are not reported. No new physical entities are introduced; the method reuses NDPCA (han Li et al. 2023) and MS-STFT discriminator (Defossez et al. 2022).

free parameters (3)
  • Perceptual loss weight (w) = 0, 10, 30, 50, 70
    Section 5.3 and Fig. 4 sweep w to generate the rate-distortion-perception trade-off; the chosen weights are hand-designated operating points.
  • Loss component weights (alpha, beta, gamma) = not reported
    Section 4 describes three loss components (task-agnostic, downstream task, perceptual) but never gives the total loss equation or the relative weights; these are implicit free parameters of the trained model.
  • Distortion target D in rate-distortion budget = not reported
    Eqs. (1)-(3) use D to convert channel capacity into encoder output dimensions, but D is never selected or reported; the experiments bypass the formula by setting bandwidth directly as the number of transmitted PCA components.
assumptions (5)
  • standard math Gaussian rate-distortion function R(D) = 0.5 log2(sigma^2 / D)
    Used in Eqs. (1)-(3) to derive the per-source bit budget from channel capacity (Zamir, Kochman, and Erez 2008).
  • domain assumption Perfect channel state information (CSI) at the receiver
    Section 3.1 states: 'We assume perfect wireless CSI availability at the receiver, which is crucial for efficient source coding.' The dynamic bitrate allocation depends on this assumption.
  • ad hoc to paper Encoder outputs follow a worst-case Gaussian distribution
    Section 3.1: 'We design the encoders for the worst case scenario, that is, where the encoder output follows a Gaussian distribution', used to set l_Gaussian,s,t and the dimension budget.
  • domain assumption The score-based generative model provides a valid differentiable task loss
    Section 4.2 defines the task loss through denoising score matching (Eq. 17) but does not specify whether the score network is frozen or trained jointly, or how the gradient flows to the autoencoder.
  • domain assumption CHiME-6 microphone recordings are aligned and correlated enough for distributed coding
    Supplementary A describes alignment and clock-skew compensation; the method assumes the selected microphones (c=3,4,5,6) capture the same conversational speech with exploitable correlation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Task and Perception-aware Distributed Source Coding for Correlated Speech under Bandwidth-constrained Channels." pith.science (2026). https://pith.science/paper/5JNQVPKC

@misc{pith2026250117879,
  author       = {Pith},
  title        = {Pith review of: Task and Perception-aware Distributed Source Coding for Correlated Speech under Bandwidth-constrained Channels},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5JNQVPKC}},
  note         = {Machine review of arXiv:2501.17879}
}
read the original abstract

Emerging wireless AR/VR applications require real-time transmission of correlated high-fidelity speech from multiple resource-constrained devices over unreliable, bandwidth-limited channels. Existing autoencoder-based speech source coding methods fail to address the combination of the following - (1) dynamic bitrate adaptation without retraining the model, (2) leveraging correlations among multiple speech sources, and (3) balancing downstream task loss with realism of reconstructed speech. We propose a neural distributed principal component analysis (NDPCA)-aided distributed source coding algorithm for correlated speech sources transmitting to a central receiver. Our method includes a perception-aware downstream task loss function that balances perceptual realism with task-specific performance. Experiments show significant PSNR improvements under bandwidth constraints over naive autoencoder methods in task-agnostic (19%) and task-aware settings (52%). It also approaches the theoretical upper bound, where all correlated sources are sent to a single encoder, especially in low-bandwidth scenarios. Additionally, we present a rate-distortion-perception trade-off curve, enabling adaptive decisions based on application-specific realism needs.

Figures

Figures reproduced from arXiv: 2501.17879 by the authors.

Figure 1
Figure 1. End to end proposed pipeline for distributed downstream speech enhancement using perceptual loss [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Proposed task aware speech enhancement using [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. (a) Efficient source coding using CSI-aware feedback vs. channel capacit, (b) Dimensions of latent space vs. task agnostic PSNR [db] for baseline E1D1, E2D1 and E4D1 autoencoders, (c) Dimensions of latent space vs. task aware PSNR [db] for baseline E1D1, E2D1 and E4D1 autoencoders. 20 40 60 80 100 Total bandwidth 0 0.2 0.4 0.6 0.8 1 N o r m aliz e d dis t o r tio n lo s s Perception w = 0 Perception w = 10 Perceptio… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Experimental distortion versus total bandwidth un [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Floor plan for correlated audio data retrieval for distributed task-aware source coding (Barker et al. 2018). [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Magnitude and phase spectrogram distributions for clean and noisy audio signals with [FFT= 2048, Hop length= 512, [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: 3D power spectrogram distribution of clean and audio signals against time [s] and frequency [Hz] [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Task agnostic loss for different baseline models. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 21 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Barker, J.; Watanabe, S.; Vincent, E.; and Trmal, J. 2018. The Fifth 'CHiME' Speech Separation and Recognition Challenge: Dataset, Task and Baselines . In Proc. Interspeech 2018, 1561--1565

    work page 2018

  4. [4]

    B.; and Gündüz, D

    Bourtsoulatze, E.; Kurka, D. B.; and Gündüz, D. 2019. Deep Joint Source-channel Coding for Wireless Image Transmission. In ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 4774--4778

    work page 2019

  5. [5]

    Casebeer, J.; Vale, V.; Isik, U.; Valin, J.-M.; Giri, R.; and Krishnaswamy, A. 2021. Enhancing into the Codec: Noise Robust Speech Coding with Vector-Quantized Autoencoders. In ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 711--715

    work page 2021

  6. [6]

    Défossez, A.; Copet, J.; Synnaeve, G.; and Adi, Y. 2022. High Fidelity Neural Audio Compression. arXiv:2210.13438

    arXiv 2022

  7. [7]

    A.; Tung, T.-Y.; Zhang, P.; and Xiao, Y

    Gündüz, D.; Wigger, M. A.; Tung, T.-Y.; Zhang, P.; and Xiao, Y. 2024. Joint Source–Channel Coding: Fundamentals and Recent Progress in Practical Designs. Proceedings of the IEEE, 1--32

    work page 2024

  8. [8]

    Task-aware Distributed Source Coding under Dynamic Bandwidth

    han Li, P.; Ankireddy, S. K.; Zhao, R.; Mahjoub, H. N.; Moradi-Pari, E.; Topcu, U.; Chinchali, S.; and Kim, H. 2023. Task-aware Distributed Source Coding under Dynamic Bandwidth. arXiv:2305.15523

    work page Pith review arXiv 2023

Show all 24 references
  1. [9]

    He, Y.; Yu, G.; and Cai, Y. 2024. Rate-Adaptive Coding Mechanism for Semantic Communications With Multi-Modal Data. IEEE Transactions on Communications, 72(3): 1385--1400

    2024

  2. [10]

    Hu, F.; Deng, Y.; Saad, W.; Bennis, M.; and Aghvami, A. H. 2020. Cellular-Connected Wireless Virtual Reality: Requirements, Challenges, and Solutions. IEEE Communications Magazine, 58(5): 105--111

    2020

  3. [11]

    Kavukcuoglu, K.; Sermanet, P.; Boureau, Y.-l.; Gregor, K.; Mathieu, M.; and Cun, Y. 2010. Learning Convolutional Feature Hierarchies for Visual Recognition. In Lafferty, J.; Williams, C.; Shawe-Taylor, J.; Zemel, R.; and Culotta, A., eds., Advances in Neural Information Proces...

    2010

  4. [12]

    Li, W.; Huang, J.; Su, Q.; Jiang, W.; and Wang, J. 2024. VASE: Enhancing Adaptive Bitrate Selection for VBR-Encoded Audio and Video Content With Deep Reinforcement Learning. IEEE Transactions on Mobile Computing, 23(12): 14889--14902

    2024

  5. [13]

    Petrangeli, S.; Simon, G.; Wang, H.; and Swaminathan, V. 2019. Dynamic Adaptive Streaming for Augmented Reality Applications. In 2019 IEEE International Symposium on Multimedia (ISM), 56--567

    2019

  6. [14]

    B.; Chen, J.; Yu, W.; and Khisti, A

    Salehkalaibar, S.; Phan, T. B.; Chen, J.; Yu, W.; and Khisti, A. J. 2023. On the choice of Perception Loss Function for Learned Video Compression. In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  7. [15]

    H.; and Zhang, L

    Wang, Z.; Wen, M.; Xu, Y.; Zhou, Y.; Wang, J. H.; and Zhang, L. 2023. Communication compression techniques in distributed deep learning: A survey. Journal of Systems Architecture, 142: 102927

    2023

  8. [16]

    Welker, S.; Richter, J.; and Gerkmann, T. 2022 a . Speech Enhancement with Score-Based Generative Models in the Complex STFT Domain. In Interspeech 2022, 2928--2932

    2022

  9. [17]

    Welker, S.; Richter, J.; and Gerkmann, T. 2022 b . Speech Enhancement with Score-Based Generative Models in the Complex STFT Domain. In Proc. Interspeech 2022, 2928--2932

    2022

  10. [18]

    Weng, Z.; Qin, Z.; Tao, X.; Pan, C.; Liu, G.; and Li, G. Y. 2023. Deep Learning Enabled Semantic Communications With Speech Recognition and Synthesis. IEEE Transactions on Wireless Communications, 22(9): 6227--6240

    2023

  11. [19]

    Whang, J.; Nagle, A.; Acharya, A.; Kim, H.; and Dimakis, A. G. 2024. Neural Distributed Source Coding. IEEE Journal on Selected Areas in Information Theory, 5: 493--508

    2024

  12. [20]

    D.; Marković, D.; and Richard, A

    Wu, Y.-C.; Gebru, I. D.; Marković, D.; and Richard, A. 2023. Audiodec: An Open-Source Streaming High-Fidelity Neural Audio Codec. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1--5

    2023

  13. [21]

    Xiao, Z.; Yao, S.; Dai, J.; Wang, S.; Niu, K.; and Zhang, P. 2023. Wireless Deep Speech Semantic Transmission. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1--5

    2023

  14. [22]

    Yue, W.; Dai, J.; Wang, S.; Si, Z.; and Niu, K. 2023. Learned Source and Channel Coding for Talking-Head Semantic Transmission. In 2023 IEEE Wireless Communications and Networking Conference (WCNC), 1--6

    2023

  15. [23]

    Zamir, R.; Kochman, Y.; and Erez, U. 2008. Achieving the Gaussian rate--distortion function by prediction. IEEE Transactions on Information Theory, 54(7): 3354--3364

    2008

  16. [24]

    Zhang, G.; Qian, J.; Chen, J.; and Khisti, A. J. 2021. Universal Rate-Distortion-Perception Representations for Lossy Compression. In Beygelzimer, A.; Dauphin, Y.; Liang, P.; and Vaughan, J. W., eds., Advances in Neural Information Processing Systems

    2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.