Pith. sign in

REVIEW 4 major objections 6 minor 50 references

NeuralPrefix: A Zero-shot Sensory Data Imputation Plugin

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

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

Pith's one-line read NeuralPrefix: a single pretrained model fills missing sensory frames on unseen sensors and modalities without retraining.

desk verdict NeuralPrefix is a clean, well-written idea, but the zero-shot modality claim currently rests on a single experiment without the trivial baselines needed to rule out temporal smoothing. read the letter →

arxiv 2502.05883 v1 pith:RUTB64A3 submitted 2025-02-09 cs.LG cs.AIstat.ML

Abdelwahed Khamis , Sara Khalifa This is my paper · ORCID
classification cs.LGcs.AIstat.ML
keywords zero-shotimputationNeuralODEsensorintermittencyspatio-temporaldatamodalitytransfermotionflow
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 data imputation for intermittently sampled sensors can be made task-agnostic and modality-agnostic: one trained generative component, placed in front of any task model, can fill missing spatio-temporal frames without retraining on the target sensor. The component, NeuralPrefix, models its latent state as a continuous trajectory governed by a Neural ODE, so it can estimate a state at any time point even when observations are irregularly spaced. On radar and pressure-mat datasets with 50 percent of frames dropped, the recovered frames reach SSIM between 0.93 and 0.96, and the same trained prefix transfers to datasets and modalities never seen in training with only modest degradation. If this holds, sensor networks would not need bespoke imputation per device; a single pretrained prefix could restore missing data for new sensors, tasks, and even different sensing modalities.

What carries the argument

The load-bearing mechanism is the continuous latent-state transition of a Neural ODE—a network that parameterizes the derivative of a latent state—defined in Eq. (3) as $h(t_{i+1})=\mathrm{ODESolve}(g_\theta, h(t_i), (t_i,t_{i+1}))$, which lets the latent state be evaluated at arbitrary, unevenly spaced times instead of only at discrete RNN steps. On top of that, modular frame generation (Eqs. 4–5) decomposes each missing frame into a warped version of the previous frame plus appearance residuals, so the generator learns cross-frame motion dynamics rather than the appearance of the source sensor; the shrinkage loss then concentrates the reconstruction budget on foreground pixels in sparse sensory frames. Together these pieces are what allow a single trained prefix to be attached to a new task model and applied to unseen modalities without retraining.

What would settle it

Take a prefix trained on a smooth sensor stream, such as Soli radar, and apply it zero-shot to a modality whose frames change discontinuously or are analytically synthesized, such as the Widar-style body-velocity profiles the paper reports as a failed case, or a high-vibration IMU stream. If the zero-shot SSIM collapses to the level of simple baselines like optical flow or last-observation-carried-forward, then the transferable-dynamics assumption is the limiting factor. A quantitative version: across many source–target sensor pairs, measure the similarity of their apparent motion flows and check whether zero-shot SSIM tracks that similarity; if it does not, the claimed mechanism is not what drives the transfer.

Watch

Extended reading notes

Core claim

NeuralPrefix is an encoder–decoder whose hidden state evolves by an ordinary differential equation: the state at any requested time is obtained by integrating a learned dynamics network from the last observed state, $h(t_i)=h(t_k)+\int_{t_k}^{t_i} g_\theta(h(t),t)\,dt$. During inference the prefix is attached ahead of an existing task model, so the task model sees a continuous stream even when the sensor drops frames. To avoid baking the appearance of the training modality into the weights, the decoder does not generate raw frames; it generates motion flow and a residual, then composes each new frame by warping the previous generated frame (Eqs. 4–5). The paper reports that this recovers 50%-missing frames with SSIM 0.93–0.96 in-domain, and that a prefix trained on one RF dataset imputes the other RF dataset and a tactile carpet dataset zero-shot, with motion-flow visualizations offered as evidence that the learned dynamics, not the training modality's appearance, transfer.

Load-bearing premise

The zero-shot modality transfer rests on the premise that different sensors observing the same kind of physical activity produce data with similar smooth, learnable motion dynamics, so a prefix trained on one modality has learned the right dynamics to impute another; the paper supports this with qualitative examples but does not measure the premise directly.

Editorial extensions

If this is right

  • A task model can be made resilient to intermittent sensor data by attaching NeuralPrefix in front of it, with no retraining or fine-tuning of either the prefix or the task model.
  • The same trained prefix supports interpolation and extrapolation, so it can also be used for early action prediction by reconstructing frames beyond the observed window.
  • At deployment, the ODE solver's tolerance can be changed to trade accuracy for speed, giving about 20% faster processing at tolerance $10^{-3}$ and 50% faster at tolerance 0.5 on MCD without changing the network architecture.
  • The learned motion flow is interpretable and can be inspected to check whether the model is transferring dynamics or memorizing the training modality's appearance.
  • On the hand-tracking downstream task, imputed MCD frames cut median tracking error from 56.85 cm to 14.85 cm, a 3.8x improvement over the Neural ODE video-generation baseline.

Reading between the lines

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

  • If the transferable-dynamics premise is right, the approach should extend beyond human motion to any phenomenon with smooth spatio-temporal evolution, such as weather field interpolation or structural vibration monitoring, but the paper does not test these domains.
  • The paper's report that the analytically synthesized Widar data performed worse suggests a measurable precondition: datasets with abrupt, non-smooth frame-to-frame changes may need a lower missing rate or a different dynamics prior, which could be tested by quantifying frame-to-frame optical-flow smoothness before choosing a prefix.
  • The residual-compensation behavior seen when motion flow is imperfect hints that the modular decomposition is self-correcting; one could deliberately corrupt or ablate the motion-flow branch and measure how much the residual branch compensates, a test the paper does not run.
  • Adapter-based test-time regularization of motion-flow magnitude, which the paper mentions as future work, could turn the qualitative motion-flow exaggeration on unseen modalities into a quantitative robustness improvement.
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 introduces NeuralPrefix, a generative component based on Neural ODEs that is inserted before a task model to impute missing spatio-temporal sensory frames. The training is done on unlabeled data from a source dataset, and the central claim is zero-shot transfer: a single trained prefix can impute missing frames on unseen datasets and even on a different sensory modality without retraining. Experiments on two RF radar datasets (Soli, MCD) and one tactile pressure mat dataset (Intelligent Carpet) report SSIM values in the 0.88–0.98 range at a 50% missing rate, with comparisons against mean, LOCF, EM, optical flow, optimal transport, and a Neural ODE video generation baseline, plus a downstream hand-tracking evaluation. The paper also discusses computational adaptability by varying the ODE solver tolerance.

Significance. If substantiated, the zero-shot modality-transfer claim would be a useful contribution: it would decouple intermittency handling from both the task and the sensor modality, allow unlabeled-data training, and make the imputer parameter-efficient and deployable on edge devices. The paper ships a clear architectural idea (modular frame generation with motion flow and residual composition), a well-motivated continuous-time formulation, and several strengths: training on unlabeled data, comparisons against a reasonable set of baselines, a downstream tracking experiment showing a 3.8x improvement over the Neural ODE baseline, and a practical inference-speed/accuracy trade-off via ODE tolerance. These strengths are real and make the approach credible for the in-domain setting. However, the paper's headline claim is zero-shot cross-modality transfer, and that claim currently rests on a single experimental cell (Soli→Carpet) without the baselines or error bars needed to distinguish genuine transferable dynamics from a static-copy prior. The significance of the paper is therefore real but conditional on additional experimental evidence.

major comments (4)
  1. [IV-F and Table II] The key zero-shot modality evidence, the Soli→Carpet row of Table II, has no baselines. No Mean, LOCF, Optical Flow, or other training-free imputer is evaluated on the Intelligent Carpet dataset. Given that Table I shows LOCF and Optical Flow are nearly as good as NeuralPrefix on Soli (SSIM 0.9646 and 0.9675 vs 0.9817), and given that the modular frame generation in Equations (4)-(5) explicitly composes each new frame as a warped version of the last observed frame plus a residual, a high SSIM on slow, background-dominated tactile data could be produced by a model that simply carries forward or lightly warps the previous frame. Without baseline rows for the Carpet dataset, the observed SSIM 0.8820 does not demonstrate that the model learns or transfers cross-modal dynamics. I request that the authors add at least Mean, LOCF, and Optical Flow results on the same Carpet test split, and also report an in-domain reference for Carpet (a model trained on Carpet) to quantify the transfer gap.
  2. [IV-F] The selection of Soli as the source for the Carpet experiment is test-set-dependent: the text states that 'Soli is picked as the seen dataset, as it is visually more similar to the carpet dataset than the MCD.' This is post hoc selection on the basis of the test set and inflates the apparent transferability. The paper should either report all source-target pairs (MCD→Carpet, Soli→Carpet, and ideally the reverse directions) or justify the choice with an a priori criterion that does not rely on visual similarity to the test data. As it stands, the single favorable transfer direction cannot support the general claim that NeuralPrefix generalises to unseen modalities.
  3. [Tables I and II] All reported numbers are point estimates without error bars, confidence intervals, or repeated runs. Several comparisons in Table I are close (e.g., Soli SSIM 0.9817 vs LOCF 0.9646 and OF 0.9675), and Table II has exactly one number per condition. Given that the paper makes claims about generalization across datasets and modalities, the lack of any variance estimate makes it impossible to assess whether the reported improvements are stable or within noise. The authors should repeat experiments over multiple random seeds/masking patterns and report means and standard deviations, at least for the headline zero-shot results.
  4. [V (Limitations) and IV-A] The paper honestly reports that the Widar 3.0 dataset was dropped because performance was lower, with less smooth cross-frame dynamics. This is a limitation, but it also narrows the scope of the zero-shot claim: the claim is only demonstrated on datasets that are, by the authors' own description, smooth and visually similar in dynamics. The paper should explicitly state that the zero-shot conclusion is limited to data with smooth apparent dynamics, and ideally test on at least one dataset with less smooth or more synthetic transitions to characterize the failure mode. Otherwise, the reported 'generalises well to unseen datasets' overstates the range of the evidence.
minor comments (6)
  1. [Abstract] The abstract says 'SSIM score between 0.93-0.96', but Table I reports extrapolation SSIM values of 0.8569 (MCD) and 0.9202 (Soli), and the zero-shot Soli→Carpet value is 0.8820. The stated range should be reconciled with the full set of results, or the sentence should specify that the range refers to the interpolation mode.
  2. [III-B] The section title reads 'NeuralPrefix Archiecture' (typo for 'Architecture'); there are also typos in the abstract ('NeuralPreifx'), in IV-B ('Structural Similairy'), and in the caption of Figure 2 ('Impuation'). These should be corrected.
  3. [IV-B] The definition of LPIPS is cited, but the paper does not state which backbone network is used for the LPIPS computation, nor how SSIM is averaged across the color/time dimensions for the 4D heatmap data. Please specify the exact metric computation procedure so the numbers are reproducible.
  4. [IV-F] The text says 'A random subject data ( 24 10 TZ) was used for testing' in the Intelligent Carpet description. This is unclear; please specify the subject identifier, the number of trials, and the split procedure, and clarify whether the test subject appeared in the training set of any source model.
  5. [IV-C] The Optimal Transport baseline is cited to reference [20], which is an authors' own survey. This is not itself a problem, but the description would be stronger if it cited the original optimal transport interpolation works used in the implementation, and if the computational cost of OT were reported, since the authors position OT as 'training-free'.
  6. [IV-G] The tolerance experiment reports SSIM 0.9122 for tolerance 1e-3 versus 0.9399 for the default, and the text says this is 'without impacting the quality'. A drop of 0.027 in SSIM is non-negligible; please rephrase to avoid overstating the trade-off, and report whether this difference is within the run-to-run variability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: NeuralPrefix is trained only on source-domain data and is evaluated on genuinely unseen datasets; no prediction reduces to a fitted parameter or to a load-bearing self-citation.

full rationale

The paper's derivation chain is self-contained. NeuralPrefix is trained with a reconstruction objective on masked frames from the seen dataset only, and the zero-shot claims are evaluated on held-out datasets (MCD, Soli, Intelligent Carpet) without retraining or fine-tuning. No parameter is fitted to the target-domain ground truth, and the reported SSIM/MSE/LPIPS/PSNR numbers are external comparisons against standard baselines (Mean, LOCF, EM, Optical Flow, Optimal Transport, and a Neural-ODE baseline). The modular frame generation in Equations 4-5 is an explicit architectural prior (a warped last frame plus residual), but it is not constructed from the target result; it is trained on source data and then applied to unseen data. The only self-citations are to the authors' own optimal-transport review [20] used as a baseline and to a survey [30] used to motivate RF domain shift; neither carries the central zero-shot claim, and both are independently verifiable external references. The paper's limitations section even acknowledges that further evaluation is needed for harder scenarios, which undercuts any suggestion that success is built into the definition. The skeptical concern about missing no-learning baselines on the Carpet dataset is an evaluation-robustness issue, not a circularity issue: it questions whether the reported number demonstrates transferable dynamics, but it does not show that the result is equivalent to its inputs by construction. Overall, the derivation is not circular.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim leans on five assumptions: shared dynamics across modalities, random masking as an intermittency proxy, smooth latent ODE trajectory, numerical solver adequacy, and modular frame composition as a perturbation of the previous frame. The only overtly tunable numbers are loss weights, solver tolerance, and window size; no new physical entities are introduced.

free parameters (4)
  • Loss weights lambda_1, lambda_2, lambda_3 = 0.05, 0.5, 1
    Hand-chosen weights in the combined loss L(G) = lambda1*L_shrinkage + lambda2*L_residual + lambda3*L_content; no sensitivity analysis is reported.
  • ODE solver tolerance = 1e-5 default, 1e-3 and 0.5 variants
    Controls numerical integration accuracy and is tuned to trade speed for quality; affects all reported results.
  • Window size = 10 frames, with 20 in the latency check
    Length of input sequence used for interpolation and extrapolation; not varied in the main accuracy results.
  • Encoder and decoder hidden dimensions and MLP size = Not reported
    Architecture capacity is chosen implicitly; without these details, the model is not fully specified for replication.
assumptions (5)
  • domain assumption Common dynamics prior: natural physical phenomena change smoothly and similar motion dynamics appear across sensor modalities, so dynamics learned on one modality transfer to another.
    Introduced in Section I and Figure 3; load-bearing for modality-OOD zero-shot results, with only qualitative illustration.
  • domain assumption Randomly masking complete frames during training faithfully simulates real data intermittency.
    Training procedure in Section III-B and evaluation in Section IV-D; authors later acknowledge that novel missingness patterns need study in Section V.
  • domain assumption The latent state trajectory is smooth enough that a learned ODE g_theta can integrate across missing intervals.
    Core of Equations 2 and 3; if dynamics are discontinuous or non-smooth, ODE integration will not recover missing frames.
  • standard math Numerical ODE solvers approximate the integral in Equation 2 accurately at the chosen tolerance.
    Black-box solver assumption in Equation 3; standard numerical analysis background, not verified by the paper.
  • domain assumption A new frame can be modeled as a warped version of the previous frame plus a residual, as in Equations 4 and 5.
    Modular frame generation assumes the new frame is mostly a perturbation of the last frame; this may fail when appearance changes abruptly.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NeuralPrefix: A Zero-shot Sensory Data Imputation Plugin." pith.science (2026). https://pith.science/paper/RUTB64A3

@misc{pith2026250205883,
  author       = {Pith},
  title        = {Pith review of: NeuralPrefix: A Zero-shot Sensory Data Imputation Plugin},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RUTB64A3}},
  note         = {Machine review of arXiv:2502.05883}
}
read the original abstract

Real-world sensing challenges such as sensor failures, communication issues, and power constraints lead to data intermittency. An issue that is known to undermine the traditional classification task that assumes a continuous data stream. Previous works addressed this issue by designing bespoke solutions (i.e. task-specific and/or modality-specific imputation). These approaches, while effective for their intended purposes, had limitations in their applicability across different tasks and sensor modalities. This raises an important question: Can we build a task-agnostic imputation pipeline that is transferable to new sensors without requiring additional training? In this work, we formalise the concept of zero-shot imputation and propose a novel approach that enables the adaptation of pre-trained models to handle data intermittency. This framework, named NeuralPrefix, is a generative neural component that precedes a task model during inference, filling in gaps caused by data intermittency. NeuralPrefix is built as a continuous dynamical system, where its internal state can be estimated at any point in time by solving an Ordinary Differential Equation (ODE). This approach allows for a more versatile and adaptable imputation method, overcoming the limitations of task-specific and modality-specific solutions. We conduct a comprehensive evaluation of NeuralPrefix on multiple sensory datasets, demonstrating its effectiveness across various domains. When tested on intermittent data with a high 50% missing data rate, NeuralPreifx accurately recovers all the missing samples, achieving SSIM score between 0.93-0.96. Zero-shot evaluations show that NeuralPrefix generalises well to unseen datasets, even when the measurements come from a different modality.

Figures

Figures reproduced from arXiv: 2502.05883 by the authors.

Figure 1
Figure 1. Zero-shot Imputation (ZSI). We frame the new problem of zero￾shot imputation and highlight potential applications ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Example application of Zero-shot Impuation. (Left) A hazard perception suit encompasses several wearable sensors, such as on-chip radars, to sense the ambience and inform/warn the wearer of important events, including quickly falling objects or approaching vehicles. Intermittency (e.g. transient sensor failure) can result in missing those events and hence dangerous consequences. (Right) NeuralPrefix can preface the … view at source ↗
Figure 3
Figure 3. Common dynamics exist in different datasets in the data space despite the modality disparity. ponent, NeuralPrefix, which can preface the task model and recover the missing spatio-temporal data samples before they are consumed by the task model (see [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: ODE data recovery modes include interpolative, predictive and retrospective. (IVP) that continually integrates the rate of change (of the system’s state) to find the next state. One step of this process (in a discrete-time step) is depicted in [PITH_FULL_IMAGE:figures…
Figure 4
Figure 4. Figure 4: Continuous vs Discrete Latent States in Neural Networks. (Bottom Left) state estimation using a form of ODE. The state at time ti, denoted as h(ti), is updated to next state h(ti+1) through the equation above. Since intermediate steps (∆t) can be arbitrarily small, the…
Figure 6
Figure 6. Figure 6: NeuralPrefix Plugin. NeuralPrefix is a continuous Encoder-Decoder architecture that imputes data in an autoregressive manner. Internal (latent) state transitions are governed by Ordinary Differential Equations. (Left) simplified linear view of the architecture. (Right)…
Figure 7
Figure 7. Figure 7: Comparison of pixel-wise losses. The proposed [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of intermittent data reconstruction using NeuralPrefix vs Neural ODE video generation baseline [18] on MCD dataset. For visual [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Hand tracking through intermittent stream in MCD dataset. CDF curves for the hand tracking error demonstrates the superiority of NeuralPrefix. Median error are denoted by triangular markers on x-axis. Inset plot shows the trivial baselines in which the hand position in…
Figure 11
Figure 11. Figure 11: For example, on Carpet, the vectors magnitudes [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]
Figure 10
Figure 10. Figure 10: Zero-shot Imputation on Intelligent carpet dataset using Neural￾Prefix trained on Soli dataset. (Top) ground truth (bottom) data imputed by the model [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: NeuralPrefix learned Motion Flow Fm (Equation 4) generalises to unseen modalities. background. This suggests that NeuralPrefix learned transfer￾able dynamics from MCD and didn’t memorise to the seen dynamics. The dashed box superimposed on the motion flow frames indic…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 46 canonical work pages

  1. [1]

    B., A ICK - ELIN , U., AND KHORSHIDI , H

    A DHIKARI , D., J IANG , W., Z HAN , J., H E, Z., R AWAT, D. B., A ICK - ELIN , U., AND KHORSHIDI , H. A. A comprehensive survey on imputation of missing data in internet of things. ACM Computing Surveys 55, 7 (2022), 1–38

    work page 2022

  2. [2]

    F., S TELLA , L., T URKMEN , C., Z HANG , X., M ERCADO , P., S HEN , H., S HCHUR , O., R ANGAPURAM , S

    A NSARI , A. F., S TELLA , L., T URKMEN , C., Z HANG , X., M ERCADO , P., S HEN , H., S HCHUR , O., R ANGAPURAM , S. S., A RANGO , S. P., KAPOOR , S., ET AL . Chronos: Learning the language of time series. arXiv preprint arXiv:2403.07815 (2024)

    arXiv 2024

  3. [3]

    T., M ITROVIC , N., D AUWELS , J., AND JAILLET , P

    A SIF, M. T., M ITROVIC , N., D AUWELS , J., AND JAILLET , P. Matrix and tensor based methods for missing data estimation in large traffic networks. IEEE Transactions on Intelligent Transportation Systems 17 , 7 (2016), 1816–1825

    work page 2016

  4. [4]

    Delving deeper into convolutional networks for learning video representations

    B ALLAS , N., Y AO, L., P AL, C., AND COURVILLE , A. Delving deeper into convolutional networks for learning video representations. arXiv preprint arXiv:1511.06432 (2015)

    arXiv 2015

  5. [5]

    T., R UBANOVA , Y., B ETTENCOURT , J., AND DUVENAUD , D

    C HEN , R. T., R UBANOVA , Y., B ETTENCOURT , J., AND DUVENAUD , D. K. Neural ordinary differential equations. Advances in neural information processing systems 31 (2018)

    work page 2018

  6. [6]

    En- semble correlation-based low-rank matrix completion with applications to traffic data imputation

    C HEN , X., W EI, Z., L I, Z., L IANG , J., C AI, Y., AND ZHANG , B. En- semble correlation-based low-rank matrix completion with applications to traffic data imputation. Knowledge-Based Systems 132 (2017), 249– 262

    work page 2017

  7. [7]

    C HENG , H., Z HANG , M., AND SHI, J. Q. A survey on deep neural network pruning: Taxonomy, comparison, analysis, and recommenda- tions. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    work page 2024

  8. [8]

    T., K LJUN , M., S ANDOR , C., AND LEIVA, L

    ˇCOPI ˇC PUCIHAR , K., A TTYGALLE , N. T., K LJUN , M., S ANDOR , C., AND LEIVA, L. A. Solids on soli: Millimetre-wave radar sensing through materials. Proceedings of the ACM on Human-Computer Interaction 6 , EICS (2022), 1–19

    work page 2022

Show all 50 references
  1. [9]

    ˇCOPI ˇC PUCIHAR , K., S ANDOR , C., K LJUN , M., H UERST , W., P LOP - SKI , A., T AKETOMI , T., K ATO, H., AND LEIVA, L. A. The missing interface: micro-gestures on augmented objects. In Extended Abstracts of the 2019 CHI Conference on Human Factors in Computing Systems (201...

    2019

  2. [10]

    G OULART , J., K IBANGOU , A., AND FAVIER , G

    DE M. G OULART , J., K IBANGOU , A., AND FAVIER , G. Traffic data imputation via tensor completion based on soft thresholding of tucker core. Transportation Research Part C: Emerging Technologies 85 (2017), 348–362

    2017

  3. [11]

    P., L AIRD , N

    D EMPSTER , A. P., L AIRD , N. M., AND RUBIN , D. B. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological) 39 , 1 (1977), 1–22

    1977

  4. [12]

    Tsi-bench: Benchmarking time series imputation

    D U, W., WANG , J., Q IAN , L., YANG , Y., IBRAHIM , Z., L IU, F., WANG , Z., L IU, H., Z HAO, Z., Z HOU , Y., ET AL . Tsi-bench: Benchmarking time series imputation. arXiv preprint arXiv:2406.12747 (2024)

    2024 arXiv

  5. [13]

    K., P RABOWO , A., RAHAMAN , M

    G AO, N., X UE, H., S HAO, W., Z HAO, S., Q IN, K. K., P RABOWO , A., RAHAMAN , M. S., AND SALIM , F. D. Generative adversarial networks for spatio-temporal data: A survey. ACM Transactions on Intelligent Systems and Technology (TIST) 13 , 2 (2022), 1–25

    2022

  6. [14]

    G., C HOW, J

    G AO, P., W ANG , M., G HIOCEL , S. G., C HOW, J. H., F ARDANESH , B., AND STEFOPOULOS , G. Missing data recovery by exploiting low- dimensionality in power system synchrophasor measurements. IEEE Transactions on Power Systems 31 , 2 (2016), 1006–1013

    2016

  7. [15]

    Learning dynamics from noisy mea- surements using deep learning with a runge-kutta constraint

    G OYAL, P., AND BENNER , P. Learning dynamics from noisy mea- surements using deep learning with a runge-kutta constraint. In The Symbiosis of Deep Learning and Differential Equations (2021)

    2021

  8. [16]

    Neural ordinary differential equations with irregular and noisy data

    G OYAL, P., AND BENNER , P. Neural ordinary differential equations with irregular and noisy data. R. Soc. Open Sci. 10 (2023), 221475

    2023

  9. [17]

    K., AND SCHUNCK , B

    H ORN , B. K., AND SCHUNCK , B. G. Determining optical flow. Artificial intelligence 17 , 1-3 (1981), 185–203

    1981

  10. [18]

    E., AND PAL, C

    K ANAA , D., V OLETI , V., K AHOU , S. E., AND PAL, C. Simple video generation using neural odes. arXiv preprint arXiv:2109.03292 (2021)

    2021 arXiv

  11. [19]

    Sf-adapter: Computational- efficient source-free domain adaptation for human activity recognition

    K ANG , H., H U, Q., AND ZHANG , Q. Sf-adapter: Computational- efficient source-free domain adaptation for human activity recognition. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiqui- tous Technologies 7, 4 (2024), 1–23

    2024

  12. [20]

    T., R OLLAND , V., AND PE- TERSSON , L

    K HAMIS , A., R USSELL TSUCHIDA , M. T., R OLLAND , V., AND PE- TERSSON , L. Earth movers in the big data era: A review of optimal transport in machine learning. arXiv preprint arXiv:2305.05080 3 (2023)

    2023 arXiv

  13. [21]

    K-nearest temperature trends: A method for weather temperature data imputation

    K IANI , K., AND SALEEM , K. K-nearest temperature trends: A method for weather temperature data imputation. In Proceedings of the 2017 International Conference on Information System and Data Mining (New York, NY , USA, 2017), ICISDM ’17, Association for Computing Machinery, p. 23–27

    2017

  14. [22]

    Continuous-time video generation via learning motion dynamics with neural ode

    K IM, K., P ARK , S., L EE, J., L EE, J., K IM, S., C HOO , J., AND CHOI , E. Continuous-time video generation via learning motion dynamics with neural ode. arXiv preprint arXiv:2112.10960 (2021)

    2021 arXiv

  15. [23]

    Deep sequential context networks for action prediction

    K ONG , Y., TAO, Z., AND FU, Y. Deep sequential context networks for action prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition (2017), pp. 1473–1481

    2017

  16. [24]

    Towards domain-independent and real-time gesture recognition using mmwave signal

    L I, Y., Z HANG , D., C HEN , J., W AN, J., Z HANG , D., H U, Y., S UN, Q., AND CHEN , Y. Towards domain-independent and real-time gesture recognition using mmwave signal. IEEE Transactions on Mobile Computing (2022)

    2022

  17. [25]

    Focal loss for dense object detection

    L IN, T.-Y., G OYAL, P., GIRSHICK , R., H E, K., AND DOLL ´AR, P. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision (2017), pp. 2980–2988

    2017

  18. [26]

    Deep object tracking with shrinkage loss

    L U, X., M A, C., S HEN , J., Y ANG , X., R EID , I., AND YANG , M.-H. Deep object tracking with shrinkage loss. IEEE transactions on pattern analysis and machine intelligence 44 , 5 (2020), 2386–2401

    2020

  19. [27]

    Intelligent carpet: Inferring 3d human pose from tactile signals

    L UO, Y., L I, Y., F OSHEY , M., S HOU , W., S HARMA , P., PALACIOS , T., TORRALBA , A., AND MATUSIK , W. Intelligent carpet: Inferring 3d human pose from tactile signals. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition (2021), pp. 11255– 11265

    2021

  20. [28]

    C., D ING , Y., L IN, C., J IANG , F., W ANG , M., AND ZHAI , C

    M A, J., C HENG , J. C., D ING , Y., L IN, C., J IANG , F., W ANG , M., AND ZHAI , C. Transfer learning for long-interval consecutive missing values imputation without external features in air pollution time series. Advanced Engineering Informatics 44 (2020), 101092

    2020

  21. [29]

    M A, Q., AND GHOSH , S. K. Emflow: Data imputation in latent space via em and deep flow models. arXiv preprint arXiv:2106.04804 (2021)

    2021 arXiv

  22. [30]

    Deep learning for radio-based human sensing: Recent advances and future directions

    N IRMAL , I., K HAMIS , A., H ASSAN , M., H U, W., AND ZHU, X. Deep learning for radio-based human sensing: Recent advances and future directions. IEEE Communications Surveys & Tutorials 23 , 2 (2021), 995–1019

    2021

  23. [31]

    Vid-ode: Continuous-time video generation with neural ordinary differential equation

    P ARK , S., K IM, K., L EE, J., C HOO , J., L EE, J., K IM, S., AND CHOI , E. Vid-ode: Continuous-time video generation with neural ordinary differential equation. In Proceedings of the AAAI Conference on Artificial Intelligence (2021), vol. 35, pp. 2412–2422

    2021

  24. [32]

    P., WANG , X.-Z., AND WU, Q

    P OURPANAH , F., A BDAR , M., L UO, Y., Z HOU , X., W ANG , R., L IM, C. P., WANG , X.-Z., AND WU, Q. J. A review of generalized zero-shot learning methods. IEEE transactions on pattern analysis and machine intelligence 45, 4 (2022), 4051–4070

    2022

  25. [33]

    Q IAN , L., I BRAHIM , Z., D U, W., YANG , Y., AND DOBSON , R. J. Un- veiling the secrets: How masking strategies shape time series imputation. arXiv preprint arXiv:2405.17508 (2024)

    2024 arXiv

  26. [34]

    M., H AJAJ , N., HARDT , M., L IU, P

    R AJKOMAR , A., O REN , E., C HEN , K., D AI, A. M., H AJAJ , N., HARDT , M., L IU, P. J., L IU, X., M ARCUS , J., S UN, M., ET AL . Scalable and accurate deep learning with electronic health records. NPJ digital medicine 1 , 1 (2018), 1–10

    2018

  27. [35]

    W., W U, W., L IN, L., X U, B., AND BERNAL , E

    R ICHARDSON , T. W., W U, W., L IN, L., X U, B., AND BERNAL , E. A. Mcflow: Monte carlo flow models for data imputation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2020), pp. 14205–14214

    2020

  28. [36]

    T., AND DUVENAUD , D

    R UBANOVA , Y., CHEN , R. T., AND DUVENAUD , D. K. Latent ordinary differential equations for irregularly-sampled time series. Advances in neural information processing systems 32 (2019)

    2019

  29. [37]

    Generative adversarial networks assist missing data imputation: A comprehensive survey & evaluation

    S HAHBAZIAN , R., AND GRECO , S. Generative adversarial networks assist missing data imputation: A comprehensive survey & evaluation. IEEE Access (2023)

    2023

  30. [38]

    Ladder: An efficient framework for video frame interpolation

    S HEN , T., L I, D., G AO, Z., T IAN , L., AND BARSOUM , E. Ladder: An efficient framework for video frame interpolation. arXiv preprint arXiv:2404.11108 (2024)

    2024 arXiv

  31. [39]

    Deep learning versus conventional methods for missing data imputation: A review and comparative study

    S UN, Y., L I, J., X U, Y., Z HANG , T., AND WANG , X. Deep learning versus conventional methods for missing data imputation: A review and comparative study. Expert Systems with Applications (2023), 120201

    2023

  32. [40]

    Deep learning for multivariate time series imputation: A survey

    W ANG , J., D U, W., C AO, W., Z HANG , K., W ANG , W., L IANG , Y., AND WEN, Q. Deep learning for multivariate time series imputation: A survey. arXiv preprint arXiv:2402.04059 (2024)

    2024 arXiv

  33. [41]

    Interacting with soli: Exploring fine-grained dynamic gesture recognition in the radio-frequency spectrum

    W ANG , S., S ONG , J., L IEN , J., P OUPYREV , I., AND HILLIGES , O. Interacting with soli: Exploring fine-grained dynamic gesture recognition in the radio-frequency spectrum. In Proceedings of the 29th annual symposium on user interface software and technology (2016), pp. 851– 860

    2016

  34. [42]

    Progressive teacher-student learning for early action prediction

    W ANG , X., H U, J.-F., L AI, J.-H., Z HANG , J., AND ZHENG , W.-S. Progressive teacher-student learning for early action prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019), pp. 3556–3565

    2019

  35. [43]

    Placement matters: Understanding the effects of device placement for wifi sensing

    W ANG , X., N IU, K., X IONG , J., Q IAN , B., Y AO, Z., L OU, T., AND ZHANG , D. Placement matters: Understanding the effects of device placement for wifi sensing. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 6 , 1 (2022), 1–25

    2022

  36. [44]

    C., S HEIKH , H

    W ANG , Z., B OVIK , A. C., S HEIKH , H. R., AND SIMONCELLI , E. P. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13 , 4 (2004), 600–612

    2004

  37. [45]

    St-mvl: filling missing values in geo-sensory time series data

    Y I, X., Z HENG , Y., Z HANG , J., AND LI, T. St-mvl: filling missing values in geo-sensory time series data. In Proceedings of the Twenty- Fifth International Joint Conference on Artificial Intelligence (2016), IJCAI’16, AAAI Press, p. 2704–2710

    2016

  38. [46]

    Y U, Z., Z HANG , Y., X IANG , X., Z OU, D., C HEN , X., AND REN, J. S. Deep bayesian video frame interpolation. In European Conference on Computer Vision (2022), Springer, pp. 144–160

    2022

  39. [47]

    Stgan: Spatio-temporal generative adversarial network for traffic data imputation

    Y UAN, Y., Z HANG , Y., W ANG , B., P ENG , Y., H U, Y., AND YIN, B. Stgan: Spatio-temporal generative adversarial network for traffic data imputation. IEEE Transactions on Big Data 9 , 1 (2022), 200–211

    2022

  40. [48]

    Expecta- tion–maximization approach to fault diagnosis with missing data

    Z HANG , K., G ONZALEZ , R., H UANG , B., AND JI, G. Expecta- tion–maximization approach to fault diagnosis with missing data. IEEE Transactions on Industrial Electronics 62 , 2 (2015), 1231–1240

    2015

  41. [49]

    A., S HECHTMAN , E., AND WANG , O

    Z HANG , R., I SOLA , P., E FROS , A. A., S HECHTMAN , E., AND WANG , O. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition (2018), pp. 586–595

    2018

  42. [50]

    Zero-effort cross-domain gesture recognition with wi-fi

    Z HENG , Y., Z HANG , Y., Q IAN , K., Z HANG , G., L IU, Y., W U, C., AND YANG , Z. Zero-effort cross-domain gesture recognition with wi-fi. In Proceedings of the 17th annual international conference on mobile systems, applications, and services (2019), pp. 313–325

    2019

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.