Pith. sign in

REVIEW 4 major objections 5 minor 32 references

CS-SHRED: Enhancing SHRED for Robust Recovery of Spatiotemporal Dynamics

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read CS-SHRED claims that inserting compressed-sensing recovery into SHRED's forward pass reconstructs heavily subsampled spatiotemporal fields with markedly higher fidelity than SHRED alone.

desk verdict Plausible SHRED extension, but training-target setup and availability-mask bug undermine the reported numbers; worth a corrected re-run. read the letter →

arxiv 2507.22303 v2 pith:SSRE7PIZ submitted 2025-07-30 cs.LG

classification cs.LG MSC 68T0794A12
keywords spatiotemporalreconstructioncompressedsensingSHREDbasispursuitdenoisingLSTMsensorsubsamplingmissingdatarecoverySNR-adaptiveloss
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 tries to show that SHRED's weakest point—reconstructing dynamics from sensor data with gaps, corrupt readings, and sparse coverage—can be fixed by adding a compressed-sensing recovery stage before the LSTM decoder. The proposed CS-SHRED solves a basis-pursuit denoising problem per sensor column, using a Fourier sparsifying basis and an $\ell_1$ penalty, then feeds the recovered time series into SHRED's recurrent decoder. On viscoelastic flow, humidity, sea-surface temperature, and rotating turbulence benchmarks, the authors report that CS-SHRED consistently beats SHRED on structural similarity, PSNR, normalized error, and perceptual metrics, especially on the hardest final snapshot. A sympathetic reader would take the claim as: high-fidelity spatiotemporal recovery from very few sensors is achievable by making the neural decoder consume denoised, sparsity-recovered inputs rather than raw subsampled ones.

What carries the argument

The load-bearing object is the per-column basis-pursuit denoising (BPDN) recovery step inserted into SHRED's forward pass. For each spatial column's time series $y_{\mathrm{sub}}(t)$, the algorithm sets available indices $\texttt{iava} = \{i : y_{\mathrm{sub}}(t)_i > 0\}$, builds a row-selection operator $R_{\mathrm{op}}$ from them, sets $\Theta = R_{\mathrm{op}} F_{\mathrm{op}}^H$ where $F_{\mathrm{op}}$ is the discrete Fourier transform, and solves $\min_\xi \|\Theta \xi - \bar{y}_{\mathrm{sub}}(t)\|_2^2 + \lambda \|\xi\|_1$. This supplies the LSTM with a denoised, gap-filled input series, and the shallow decoder turns the LSTM's final hidden state into the full high-dimensional field. A second mechanism is the SNR-adaptive loss, which adds a piecewise $\pm\lambda_{\mathrm{SNR}}\,\mathrm{SNR}$ term to the MSE and MAE losses with an $\ell_2$ weight-decay regularizer.

What would settle it

Take a dataset whose valid field contains negative values and zero crossings (e.g., a velocity component), skip the min-max scaling that would make zeros ambiguous, subsample with an explicit mask, and compare CS-SHRED against SHRED on ground truth; if the reported SSIM and PSNR gains vanish or reverse, the central claim of robust recovery from incomplete data is falsified.

Watch

Extended reading notes

Core claim

The central claim is that coupling a Fourier-domain compressed-sensing denoiser with SHRED's recurrent decoder produces reconstructions that preserve small-scale structures under heavy subsampling and noise, where plain SHRED blurs or distorts them. Concretely, for every sensor time series the method builds a restriction operator $R_{\mathrm{op}}$ selecting the entries flagged as observed, forms $\Theta = R_{\mathrm{op}}F_{\mathrm{op}}^H$ with the inverse Fourier operator, and solves $\arg\min_\xi \|\Theta\xi - \bar{y}_{\mathrm{sub}}\|_2^2 + \lambda\|\xi\|_1$ via a spectral projected-gradient $\ell_1$ solver; the recovered series then enters the LSTM and shallow decoder. The paper reports, for example, SSIM rising from 0.730 to 0.952 and PSNR from 20.12 dB to 27.47 dB on the Oldroyd-B test, with similar directional gains on qmax, SST, and TURB-Rot. The authors also introduce an SNR-weighted loss that switches from penalizing inverse SNR to penalizing negative SNR, so that low-quality reconstructions are down-weighted while fine features in high-SNR regions are retained.

Load-bearing premise

The pipeline assumes that a sensor reading of exactly zero marks a missing sample, so after normalization any genuine zero or negative field value is treated as a gap and replaced by compressed-sensing imputation.

Editorial extensions

If this is right

  • Under the reported settings, CS-SHRED recovers fine-scale structures in the final snapshot that SHRED misses, with SSIM and PSNR improving substantially on every benchmark the paper tests.
  • The method inherits SHRED's sensor flexibility: with one randomly placed sensor on three benchmarks and five on the turbulent flow, the recovered time series suffice for full-state decoding.
  • Because the CS stage runs per training batch, the model trains on complete signals and evaluates on incomplete ones, making the denoising jointly learned with the recurrent decoder rather than a separate preprocessing step.
  • The gains appear across viscoelastic, humidity, SST, and turbulent regimes, indicating the benefit comes from the sparsity-recovery mechanism rather than from any single dataset's structure.

Reading between the lines

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

  • Editorial inference: the reported advantage is largest on the last snapshot, which is always the subsampled one; a natural check is whether the gap closes when the evaluation snapshot is not degraded, since the CS stage may be doing disproportionate work exactly there.
  • Editorial inference: replacing the zero-mask with an explicit missing-data mask should extend the same architecture to fields with genuine zero or negative values, such as velocity components with zero crossings, without silently imputing valid measurements.
  • Editorial inference: the Fourier basis is a fixed, physics-agnostic prior; swapping it for a data-driven dictionary learned from training snapshots could improve recovery on strongly anisotropic fields, at the cost of losing the basis's universality and the simple per-column solve.
  • Editorial inference: the method roughly tripled training time on the viscoelastic benchmark, so deployment choices will likely hinge on whether the fidelity gain justifies the extra compute in applications with tight retraining budgets.
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 manuscript proposes CS-SHRED, an extension of the SHRED architecture in which a basis-pursuit denoising (BPDN) recovery stage, implemented with a Fourier dictionary and an SPGL1 solver, is applied to subsampled sensor time series before the LSTM and shallow decoder reconstruct the full spatiotemporal field. Training is guided by a composite loss that combines MSE, MAE, an SNR-based term, and L2 weight decay. The method is evaluated on four datasets: viscoelastic Oldroyd-B flow, maximum specific humidity, sea surface temperature, and rotating turbulent flow. The central claim is that CS-SHRED reconstructs heavily subsampled fields with substantially higher fidelity than SHRED, measured by SSIM, PSNR, normalized error, and LPIPS.

Significance. If the reported gains are real, CS-SHRED would be a useful, readily applicable extension of SHRED for incomplete sensor data, and the public code release is a concrete asset for the community. The choice of four diverse spatiotemporal datasets is a strength, and the detailed computational cost analysis in Section 3.2.3 is informative. However, the paper currently does not establish the central quantitative claim: the training-target construction in Section 3.1 appears to train the model to predict zeroed missing entries, the availability mask in Algorithm 1 misclassifies genuine zero and negative measurements, the SNR term in Equation (10) is essentially a monotone function of the same MSE already in the loss, and the comparisons use per-model hyperparameter optimization with a single seed and no error bars. These are load-bearing issues rather than presentation concerns, so the significance of the empirical results, in the form presented, is not yet established.

major comments (4)
  1. [3.1] The dataset construction in Section 3.1 defines D_train and D_val with targets x^(t+l)_sub,p, i.e., subsampled frames, while D_test uses original targets x^(t+l)_ori,p. For any training target whose time index lies in T_sub, the target entries in the missing columns are zero, so the MSE loss in Equation (9) and Algorithm 3 actively penalizes nonzero predictions at exactly the locations the method is supposed to recover. The reported test metrics in Tables 3, 9, 11, and 13 are computed against original frames, so they do not measure what the training procedure optimizes unless the code silently replaces the stated targets. This affects all four datasets and must be resolved by re-training with original data targets (or equivalently by masking the loss to observed entries) and reporting the resulting metrics.
  2. [2.1.1, Algorithm 1] Algorithm 1, line 13, defines the available indices iava as the set where y_sub(t) > 0. After Min-Max normalization to [0,1] in Section 3.1, genuine zero-valued measurements are treated as missing, and for TURB-Rot, which is normalized to zero mean and unit variance in Section 3.6.1, the velocity field contains many negative values that would also be classified as missing. Since the entire CS recovery stage is driven by this mask, the imputation replaces valid measurements with CS estimates, corrupting both the recovered time series and the reported metrics. The mask must be defined from the known subsampling pattern rather than from the sign of the data, and the TURB-Rot experiment must be rerun with such a mask.
  3. [2.1, Eqs. (9)-(10)] The SNR term in Equation (10) defines noise power as the squared reconstruction error ||x(t) - H({y*})||_2^2 averaged over the batch, which is exactly the MSE term L_MSE up to normalization. Therefore the term lambda_SNR / SNR (or -lambda_SNR * SNR) is a monotone function of the same objective that already appears as lambda_L2 * L_MSE. Adding it double-counts the MSE objective and does not introduce an independent SNR constraint. The claim that the loss 'suppresses noise and outliers in low-SNR regions while preserving fine-scale features in high-SNR regions' is not supported by the equations as written. The authors should either reformulate the SNR term so that it depends on an estimate of noise distinct from the reconstruction error, or remove the claim and provide an ablation showing what the SNR term contributes beyond L_MSE.
  4. [3.2.2 and Tables 2, 8, 10, 12] The experimental comparison is confounded. Each model is tuned separately with Optuna, and the resulting configurations differ not only in the presence of the CS stage and SNR loss but also in hidden size, number of layers, batch size, learning rate, number of lags, and number of epochs. The SHRED baseline is therefore not an ablated version of CS-SHRED, and the large metric differences in Tables 3, 9, 11, and 13 could be due to hyperparameter search or architectural capacity rather than the proposed components. Additionally, all experiments use a single seed (915) with no error bars, so the claimed 'significantly higher fidelity' is not statistically supported. I ask for a controlled comparison: identical architecture and lag settings where possible, multiple seeds with mean and standard deviation, and ablations that remove the CS recovery stage, the SNR term, and the L1 term one at a time.
minor comments (5)
  1. [Title and Abstract] The title 'EnhancingSHREDfor Robust Recovery' has missing spaces and should read 'Enhancing SHRED for Robust Recovery'; the same typo appears in the running header.
  2. [3.2.2, Table 2] The paragraph after Table 2 states that SHRED has regularization parameters lambda_L1 = 0.00665, lambda_L2 = 0.15933, and lambda_SNR = 0.04275, but Table 2 shows dashes for these SHRED entries; the text and table are inconsistent and should be reconciled.
  3. [3.6.1] The description says the extracted data are organized into three-dimensional arrays 'in (x,y,z)', but the experiments appear to use two-dimensional slices; please clarify the exact dimensionality used for training and evaluation.
  4. [2.1, Eq. (9)] Equation (9) is described as a piecewise loss, but the pieces differ only in the sign of the SNR term; the notation and the surrounding text should make clear whether the switch is evaluated per batch, per sample, or per spatial location, since 'piecewise SNR regularization' is claimed but never precisely defined.
  5. [Section 4] The final paragraph states that frequency-domain constraints such as bandpass filtering were 'not implemented in this study'; this is a useful limitation statement, but it should appear in the experimental section where the Fourier dictionary is introduced, so that readers know the dictionary is not adapted to the data spectra.

Circularity Check

2 steps flagged · score 5.0 of 10

The SNR-regularization term is defined from the same MSE it is added to, and the training targets are subsampled frames, so two headline mechanisms do not establish the claimed reconstruction gains.

  1. self definitional [Section 2.1, Eqs. (9)-(10)]
    "SNR= 10 log 10( signal power noise power+ϵ ) (10) ... noise power = 1/T Σ ||x(t)−H({y⋆(t)}... )||²₂ ... L={λsnr·snr−1 +λ L2·L MSE +λ L1·L MAE +Rℓ2, snr>0 ...}"

    Eq. (10) defines the SNR regularizer's denominator, 'noise power', as exactly the per-sample squared reconstruction error that Eq. (9)'s L_MSE term minimizes. Adding λsnr·SNR^{-1} (or −λsnr·SNR) to λL2·L_MSE therefore adds no independent constraint: it is a monotone re-weighting of the same error. The abstract's claim of a 'piecewise SNR regularization' that 'suppresses noise and outliers in low-SNR regions' is not supported by Eq. (9), which applies a global scalar SNR with only a sign branch. The innovation is, by construction, the MSE objective it is added to.

  2. fitted input called prediction [Section 3.1, dataset definitions]
    "Dtrain ={(x (t)_sub,p, x (t+l)_sub,p)} ... Dtest ={(x (t)_sub,p, x (t+l)_ori,p )} ... The training and validation use subsampled data for both input and target. Testing properly evaluates reconstruction by comparing subsampled inputs against original data targets"

    By Eq. (12), x_sub is zero on Y_sub×T_sub. Algorithm 3 computes L_MSE = ||xhat(tcurrent) − x(tcurrent)||² with x(tcurrent) drawn from D_train, i.e. a subsampled target. Thus the training loss explicitly penalizes nonzero predictions in the missing columns, the exact locations the paper says CS-SHRED recovers. The test metrics in Tables 3, 9, 11, and 13 are computed against original targets, so the reported reconstruction fidelity is not a learned recovery of missing structure; the model was constructed to reproduce the corruption instead. The claimed 'prediction' is therefore an artifact of the train/test target mismatch, not a consequence of the architecture.

full rationale

The central empirical comparison between CS-SHRED and SHRED is not, by itself, circular: both models are evaluated on the same four datasets, and the CS pre-processing stage could plausibly improve reconstruction. However, two construction-level problems undermine the paper's headline claims. First, the SNR regularizer in Eq. (10) is defined directly from the same per-sample MSE that appears in the L_MSE term of Eq. (9), so adding it to the loss does not introduce an independent, spatially localized constraint; the claimed 'piecewise' SNR behavior is not present in the equations. Second, Section 3.1 defines D_train and D_val targets as subsampled frames, which are zeroed at the missing columns by Eq. (12). During training, the model is therefore penalized for outputting nonzero values at exactly the locations where the paper claims it recovers missing data, while the test metrics are computed against original frames. This means the reported absolute fidelity numbers do not measure the learned recovery the text describes; the relative ranking may survive because SHRED is trained under the same target definition, but the absolute claim of high-fidelity reconstruction is unsupported as written. I also note two non-circular concerns: Algorithm 1's y_sub(t)>0 availability mask conflicts with the zero-mean/unit-variance TURB-Rot data, which contain valid negative velocities, and the CS recovery stage is a fixed SPGL1 solver with no trainable parameters, so the 'jointly trained CS and SHRED' phrasing is not supported by the architecture. The self-citations to prior SHRED work are appropriate baseline references and are not load-bearing in a circular way.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

This ledger shows the main uncharged assumptions. The BPDN penalty lambda is never reported. All loss weights and lag lengths are tuned per dataset with Optuna, and the two models are not compared under matched budgets. The biggest unstated assumption is that zero-valued entries represent missing data; after normalization, this is false for many valid samples. The CS recovery assumes per-column Fourier sparsity and ignores spatial coherence. No new physical entities are introduced.

free parameters (6)
  • BPDN sparsity penalty lambda = not reported
    Eq. 7 and Algorithm 1 rely on lambda to balance fidelity and sparsity in the CS recovery stage; no value or tuning schedule is given for any dataset.
  • lambda_L1 (L1 regularization) = Oldroyd 0.00314, qmax 0.013111, SST 0.0041, TURB-Rot 0.0091
    Chosen per dataset by Optuna (Tables 2, 8, 10, 12); contributes to the loss and differs across models, so comparisons are not matched.
  • lambda_L2 (L2 regularization) = Oldroyd 0.74010, qmax 0.376978, SST 0.3220, TURB-Rot 0.2513
    Chosen per dataset by Optuna; the loss weight is not held constant between CS-SHRED and SHRED.
  • lambda_SNR (SNR regularization) = Oldroyd 0.01183, qmax 0.003763, SST 0.00130, TURB-Rot 0.8552
    Chosen per dataset by Optuna; the SNR term is a function of MSE, so this weight partly rescales the MSE contribution.
  • number of lags l = Oldroyd 10, qmax 24, SST 36, TURB-Rot 15
    Temporal embedding length tuned per model; SHRED uses different lags (20, 54, 24, 15), confounding comparisons.
  • architecture hyperparameters (hidden size, layers, batch size) = Oldroyd 256x2/512 vs 128x1/128; qmax 256x1/512 vs 512x2/256; SST 512x2/64 vs 512x2/512; TURB-Rot 256x3/32 vs 128x3/128
    Tuned separately with Optuna; the CS-SHRED model is larger on the Oldroyd dataset, so the authors' claim that gains come from methodology rather than complexity is not demonstrated.
assumptions (5)
  • domain assumption Each sensor time series is approximately sparse in the Fourier basis
    Algorithm 1 and Eq. 7 solve basis pursuit denoising in the FFT domain; recovery quality depends on this sparsity, which is not verified per dataset.
  • ad hoc to paper Missing entries are exactly the zero-valued entries of the subsampled tensor
    Algorithm 1 line 13 uses y_sub(t)>0 as the availability mask; after normalization this is false for genuine zeros and negative values.
  • domain assumption Takens-style time-lag embedding suffices for LSTM state estimation
    CS-SHRED inherits this premise from SHRED [1] when feeding lagged recovered sequences to the LSTM.
  • standard math Standard FFT and BPDN theory (convex relaxation recovers sparse signals)
    Background for the CS stage; the paper does not numerically verify recovery guarantees on these datasets.
  • ad hoc to paper SNR loss is an independent objective from MSE
    By Eq. 10, SNR for a fixed target is a deterministic function of the MSE; the piecewise branch adds no spatial information, so parts of the loss design are redundant with L_MSE.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CS-SHRED: Enhancing SHRED for Robust Recovery of Spatiotemporal Dynamics." pith.science (2026). https://pith.science/paper/SSRE7PIZ

@misc{pith2026250722303,
  author       = {Pith},
  title        = {Pith review of: CS-SHRED: Enhancing SHRED for Robust Recovery of Spatiotemporal Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SSRE7PIZ}},
  note         = {Machine review of arXiv:2507.22303}
}
abstract

We present CS-SHRED, a novel deep learning architecture that integrates Compressed Sensing (CS) into a Shallow Recurrent Decoder (SHRED) to reconstruct spatiotemporal dynamics from incomplete, compressed, or corrupted data. Our approach introduces two key innovations. First, by incorporating CS techniques into the SHRED architecture, our method leverages a batch-based forward framework with $\ell_1$ regularization to robustly recover signals even in scenarios with sparse sensor placements, noisy measurements, and incomplete sensor acquisitions. Second, an adaptive loss function dynamically combines Mean Squared Error (MSE) and Mean Absolute Error (MAE) terms with a piecewise Signal-to-Noise Ratio (SNR) regularization, which suppresses noise and outliers in low-SNR regions while preserving fine-scale features in high-SNR regions. We validate CS-SHRED on challenging problems including viscoelastic fluid flows, maximum specific humidity fields, sea surface temperature distributions, and rotating turbulent flows. Compared to the traditional SHRED approach, CS-SHRED achieves significantly higher reconstruction fidelity -- as demonstrated by improved SSIM and PSNR values, lower normalized errors, and enhanced LPIPS scores-thereby providing superior preservation of small-scale structures and increased robustness against noise and outliers. Our results underscore the advantages of the jointly trained CS and SHRED design architecture which includes an LSTM sequence model for characterizing the temporal evolution with a shallow decoder network (SDN) for modeling the high-dimensional state space. The SNR-guided adaptive loss function for the spatiotemporal data recovery establishes CS-SHRED as a promising tool for a wide range of applications in environmental, climatic, and scientific data analyses.

Figures

Figures reproduced from arXiv: 2507.22303 by the authors.

Figure 1
Figure 1. Overview of the CS-SHRED pipeline: (1) The original dynamics x are uniformly sub-sampled in one spatial dimension for each snapshot. (2) A few sub-sampled sensors (s1, s2, s3) = y sub i are randomly positioned to capture the multivariate time series {y sub i } tcurrent i=tcurrent−l with a predefined percentage of missing information. (3) The missing data in the multivariate time series collected by the sensors are r… view at source ↗
Figure 2
Figure 2. Temporal dynamics captured by a sensor fixed at a randomly chosen spatial position (a), and (b) [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Temporal dynamics captured by a sensor fixed at a randomly chosen spatial position (a), and [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison between SHRED and CS-SHRED for the specific humidity (qmax) data: (a) shows the original data of the last snapshot, (b) shows the respective subsampled data, (c) shows the temporal dynamics captured by a sensor fixed at a randomly chosen spatial position. Th…
Figure 5
Figure 5. Figure 5: Comparison between SHRED and CS-SHRED for the Sea Surface Temperature data: (a) shows the original data of the last snapshot, (b) shows the respective subsampled data, (c) shows the temporal dynamics captured by a sensor fixed at a randomly chosen spatial position. The…
Figure 6
Figure 6. Figure 6: Visualization of the rotating turbulent flow focusing on the [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Comparison between SHRED and CS-SHRED for the Rotating Turbulent Flow data: (a) shows the original data of the last snapshot, (b) shows the respective subsampled data, (c) shows the temporal dynamics captured by a sensor fixed at a randomly chosen spatial position. The…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 25 canonical work pages

  1. [1]

    Williams, Olivia Zahn, and J

    Jan P. Williams, Olivia Zahn, and J. Nathan Kutz. Sensing with shallow recurrent decoder networks. arXiv preprint arXiv:2301.12011, 2023

  2. [2]

    Leveraging arbitrary mobile sensor trajectories with shallow recurrent decoder networks for full-state reconstruction.IEEE Access, 2024

    Megan R Ebers, Jan P Williams, Katherine M Steele, and J Nathan Kutz. Leveraging arbitrary mobile sensor trajectories with shallow recurrent decoder networks for full-state reconstruction.IEEE Access, 2024

  3. [3]

    Reduced order modeling with shallow recurrent decoder networks.arXiv preprint arXiv:2502.10930, 2025

    Matteo Tomasetto, Jan P Williams, Francesco Braghin, Andrea Manzoni, and J Nathan Kutz. Reduced order modeling with shallow recurrent decoder networks.arXiv preprint arXiv:2502.10930, 2025

  4. [4]

    Sparse identification of nonlinear dynamics and koopman operators with shallow recurrent decoder networks.arXiv preprint arXiv:2501.13329, 2025

    Mars Liyao Gao, Jan P Williams, and J Nathan Kutz. Sparse identification of nonlinear dynamics and koopman operators with shallow recurrent decoder networks.arXiv preprint arXiv:2501.13329, 2025

  5. [5]

    Robust State Estimation from Partial Out-Core Measurements with Shallow Recurrent Decoder for Nuclear Reactors

    Stefano Riva, Carolina Introini, Antonio Cammi, and J Nathan Kutz. Robust state estimation from partial out-core measurements with shallow recurrent decoder for nuclear reactors.arXiv preprint arXiv:2409.12550, 2024

  6. [6]

    Simply denoise: Wavefield reconstruction via jittered under- sampling.Geophysics, 73(3):V19–V28, 2008

    Gilles Hennenfent and Felix J Herrmann. Simply denoise: Wavefield reconstruction via jittered under- sampling.Geophysics, 73(3):V19–V28, 2008

  7. [7]

    P. C. de Assis, R. B. da Silva, I. A. L. Neto, A. C. Galante, A. L. Campi, A. R. de Oliveira, and M. A. Ceia. Compressive sensing framework using the linear radon transform for 3d ultrasonic data reconstruction in a pinch-out reservoir model. InSecond International Meeting for Applied Geoscience & Energy, pages 2621–2625. Society of Exploration Geophysici...

  8. [8]

    Recovering from missing data in population imaging—cardiac mr image imputation via conditional generative adversarial nets.Medical Image Analysis, 67:101812, 2021

    Yan Xia, Wenjia Bai, Pengxiang Hu, Li Wang, Kuangyu Shi, Qingsong Chen, and Pheng-Ann Heng. Recovering from missing data in population imaging—cardiac mr image imputation via conditional generative adversarial nets.Medical Image Analysis, 67:101812, 2021

Show all 32 references
  1. [9]

    Optimization-inspired compact deep compressive sensing.IEEE Journal of Selected Topics in Signal Processing, 14(4):765–774, 2020

    Jian Zhang, Chen Zhao, and Wen Gao. Optimization-inspired compact deep compressive sensing.IEEE Journal of Selected Topics in Signal Processing, 14(4):765–774, 2020

  2. [10]

    Comparison of common algorithms for single-pixel imaging via compressed sensing.Sensors, 23(10):4678, 2023

    Wenjing Zhao, , andet al. Comparison of common algorithms for single-pixel imaging via compressed sensing.Sensors, 23(10):4678, 2023

  3. [11]

    Determining the number of measurements for compressive sensing of traffic-induced vibration data.Measurement, 152:107259, 2020

    Fereidoun Amini, Yousef Hedayati, and Hadi Zanddizari. Determining the number of measurements for compressive sensing of traffic-induced vibration data.Measurement, 152:107259, 2020

  4. [12]

    Convo- lutional lstm network: A machine learning approach for precipitation nowcasting.Advances in Neural Information Processing Systems, pages 802–810, 2015

    Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Convo- lutional lstm network: A machine learning approach for precipitation nowcasting.Advances in Neural Information Processing Systems, pages 802–810, 2015

  5. [13]

    Robust spatial–temporal imputation based on spatio-temporal generative adversarial nets.Knowledge-Based Systems, 279:110919, 2023

    Longji Huang, Jianbin Huang, He Li, and Jiangtao Cui. Robust spatial–temporal imputation based on spatio-temporal generative adversarial nets.Knowledge-Based Systems, 279:110919, 2023

  6. [14]

    A survey on deep transfer learning

    Chuanqi Tan, Fangxiang Sun, Tao Kong, Wenchao Zhang, Chao Yang, and Chunfeng Liu. A survey on deep transfer learning. InInternational Conference on Artificial Neural Networks, pages 270–279. Springer, 2018

  7. [15]

    Zeng Chen, Huan Xu, Peng Jiang, Shanen Yu, Guang Lin, Igor Bychkov, Alexey Hmelnov, Gennady Ruzhnikov, Ning Zhu, and Zhen Liu. A transfer-learning-based lstm strategy for imputing large-scale consecutive missing data and its application in a water quality prediction system.Jou...

  8. [16]

    A transfer-learning approach for accelerated mri using deep neural networks

    Salman Ul Hassan Dar et al. A transfer-learning approach for accelerated mri using deep neural networks. arXiv e-prints, page arXiv:1710.02615, 2017

  9. [17]

    Reviewing autoencoders for missing data imputation: Technical trends, applications and outcomes

    Ricardo Cardoso Pereira, Miriam Seoane Santos, Pedro Pereira Rodrigues, and Pedro Henriques Abreu. Reviewing autoencoders for missing data imputation: Technical trends, applications and outcomes. Journal of Artificial Intelligence Research, 69:1255–1285, 2020

  10. [18]

    Saits: Self-attention-based imputation for time series.Expert Systems with Applications, 219:119619, 2023

    Wenjie Du, David Côté, and Yan Liu. Saits: Self-attention-based imputation for time series.Expert Systems with Applications, 219:119619, 2023

  11. [19]

    Attention is all you need.Advances in Neural Information Processing Systems, 30:5998–6008, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in Neural Information Processing Systems, 30:5998–6008, 2017. 29 EnhancingSHREDfor Robust Recovery of Spatiotemporal Dynamics

  12. [20]

    Basis pursuit

    Shaobing Chen and David Donoho. Basis pursuit. InProceedings of the 1994 28th Asilomar Conference on Signals, Systems and Computers, pages 41–44. IEEE, 1994

  13. [21]

    Friedlander

    Ewout van den Berg and Michael P. Friedlander. Probing the pareto frontier for basis pursuit solutions. SIAM Journal on Scientific Computing, 31(2):890–912, 2009

  14. [22]

    Pylops—a linear-operator python library for scalable algebra and optimization.SoftwareX, 11:100361, 2020

    Matteo Ravasi and Ivan Vasconcelos. Pylops—a linear-operator python library for scalable algebra and optimization.SoftwareX, 11:100361, 2020

  15. [23]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

  16. [24]

    Turb-rot

    Luca Biferale, Fabio Bonaccorso, Michele Buzzicotti, and Patricio Clark di Leoni. Turb-rot. a large database of 3d and 2d snapshots from turbulent rotating flows.ArXiv, abs/2006.07469, 2020

  17. [25]

    Oishi andet al

    Cassio M. Oishi andet al. Nonlinear parametric models of viscoelastic fluid flows.Royal Society Open Science, 11(10):240995, 2024

  18. [26]

    Hart and Kevin Martinez

    Jane K. Hart and Kevin Martinez. Environmental sensor networks: A revolution in the earth system science?Earth-Science Reviews, 78(3-4):177–191, 2006

  19. [27]

    Trenberth

    Kevin E. Trenberth. Challenges of a sustained climate observing system. InClimate Science for Serving Society, pages 13–50. Springer, 2013

  20. [28]

    Quality control of ocean temperature and salinity profiles—historical and real-time data.Journal of Marine Systems, 69(1-2):1–4, 2008

    Eleanor Anderson, Ruth Pettifer, and Mark Bell. Quality control of ocean temperature and salinity profiles—historical and real-time data.Journal of Marine Systems, 69(1-2):1–4, 2008

  21. [29]

    David L. Donoho. Compressed sensing.IEEE Transactions on Information Theory, 52(4):1289–1306, 2006

  22. [30]

    Candès and Michael B

    Emmanuel J. Candès and Michael B. Wakin. An introduction to compressive sampling.IEEE Signal Processing Magazine, 25(2):21–30, 2008

  23. [31]

    James E. Fowler. Compressive-projection principal component analysis.IEEE Transactions on Image Processing, 21(4):1863–1871, 2012

  24. [32]

    Alexakis and L

    A. Alexakis and L. Biferale. Cascades and transitions in turbulent flows.Physics Reports, 767–769:1–101, November 2018. 30

Pith tools

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