Pith. sign in

REVIEW 4 major objections 6 minor 65 references

Met$^2$Net: A Decoupled Two-Stage Spatio-Temporal Forecasting Model for Complex Meteorological Systems

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A single two-stage model can forecast multiple weather variables more accurately than dedicated single-variable models by aligning a shared-latent training phase with an inter-variable prediction phase.

desk verdict Per-variable encoders are a good idea, but the paper's two-stage training story is not actually tested: the ablation conflates the extra loss with the freezing/momentum scheme. read the letter →

arxiv 2507.17189 v1 pith:B2MLBU2K submitted 2025-07-23 cs.LG

Shaohan Li , Hao Yang , Min Chen , Xiaolin Qin This is my paper · ORCID
classification cs.LG
keywords spatiotemporalforecastingmultivariateweatherpredictiontwo-stagetraininglatentspacemomentumupdatesrepresentationconsistencyvariableattentionBench
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

Met2Net asks why feeding several meteorological variables into one spatiotemporal model does not help, and sometimes hurts, relative to forecasting each variable separately. The paper's answer is that naive multivariable training suffers from representation inconsistency—variables like temperature, humidity, cloud cover, and wind have such different distributions that a shared network either loses variable-specific features or lets one variable dominate—and from task inconformity, where borrowing the two-stage recipe from generative models shifts from reconstruction in stage one to prediction in stage two. The proposed fix is an implicit two-stage loop: stage one freezes the Translator while per-variable encoders and decoders learn a shared latent space; stage two freezes the encoders and decoders while the Translator predicts in that space, guided by a latent-space MSE loss, with momentum updates smoothing the frozen weights. A single resulting model lowers T2M (2-meter temperature) MSE by 28.82% and relative-humidity MSE by 23.39% against the strongest single-variable baseline on WeatherBench, and improves MSL, U10, V10, and T2M on a cropped ERA5 region plus typhoon-track forecasting. If the attribution is right, heterogeneous physical variables are best integrated as separate modalities fused in latent space rather than as extra channels in one network.

What carries the argument

The load-bearing mechanism is the implicit two-stage training loop. In stage one the Translator's gradients are frozen while per-variable encoders and decoders learn a shared latent space; in stage two the encoders and decoders are frozen while the Translator learns temporal and inter-variable dynamics, and a latent-space prediction loss $L_2(Z'_y, Z_y)$ keeps both stages on the forecasting task. Momentum updates, $\theta_m \leftarrow \alpha\theta_m + (1-\alpha)\theta$ with $\alpha = 0.999$, transfer information into the frozen modules so the two halves stay aligned inside one training run, and the total loss is $L = L_1(Y', Y) + L_2(Z'_y, Z_y)$. Around this core sit two supporting choices: per-variable encoders and decoders that avoid forcing heterogeneous physical variables into one representation, and a variable-attention operation $A = \mathrm{softmax}(QK^\top/\sqrt{d})$ over the variable axis in the latent space that captures inter-variable couplings such as the U/V wind pair or the T2M–TCC pair.

What would settle it

Train the same per-variable encoder–decoder architecture on WeatherBench end-to-end with per-variable normalization but without the freezing schedule or latent-space loss; if its T2M MSE reaches Met2Net's 0.8271, the two-stage recipe is not what produces the gain. A complementary check is to run the full implicit two-stage scheme with the momentum EMA replaced by a plain stop-gradient operation and compare accuracy on T2M and R.

Watch

Extended reading notes

Core claim

The central discovery the paper argues for is that representation inconsistency and task inconformity are separable, fixable causes of poor multivariable forecasting, and that both dissolve under a single continuous training cycle that never changes its objective. Concretely, Met2Net assigns each variable its own encoder–decoder pair, freezes the spatiotemporal Translator while those pairs learn to represent and reconstruct each variable in a shared latent space (stage one), then freezes the pairs and trains the Translator to predict the encoded future, adding a latent-space loss $L_2(Z'_y, Z_y)$ so the second stage is still forecasting rather than reconstruction (stage two); momentum updates $\theta_m \leftarrow \alpha\theta_m + (1-\alpha)\theta$ keep the frozen modules moving smoothly. A self-attention mechanism over the variable axis inside the Translator fuses the variables. The paper reports that this one model achieves state-of-the-art MSE, MAE, and RMSE across low-resolution, high-resolution, and high-altitude settings, with the headline numbers being a 28.82% T2M MSE reduction and a 23.39% relative-humidity MSE reduction against TAU on WeatherBench, and higher R²/ACC on a cropped ERA5 region.

Load-bearing premise

The argument credits the gains to two-stage training, but its multivariable ablation baseline stacks the four variables as plain image channels without per-variable normalization, so the claim assumes that a single-stage multivariable model given the same inputs, properly normalized and tuned, would not close most of the distance to Met2Net's numbers.

Editorial extensions

If this is right

  • One Met2Net model replaces four separate single-variable models on WeatherBench, using 8.65M inference parameters and lowering MSE by 5.46–28.82% across the four variables.
  • The gains persist at high spatial resolution and for high-altitude variables (150–850 hPa), so the result is not an artifact of one dataset's scale.
  • On the cropped ERA5 region, MSL, U10, V10, and T2M all show lower MSE, lower MAE, and higher R² and ACC than TAU, and the model tracks Typhoon Mawar more closely at 3-hour lead time.
  • The training recipe transfers to a synthetic multivariate video dataset (MvMmfnist) and keeps single-variable performance on TaxiBJ, indicating it generalizes beyond weather.
  • Inference cost is identical to a standard end-to-end model; the extra parameters and computation are confined to training, because the frozen modules are dropped at inference.

Reading between the lines

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

  • A fair share of the reported gain may come from per-variable encoders acting as per-variable normalization, rather than from the freezing schedule itself; the paper's ablation tree does not include a single-stage model with per-variable encoders but no two-stage objective.
  • The latent-space loss $L_2$ doubles as deep supervision on the encoder features; a testable conjecture is that part of the improvement would survive even without gradient freezing, which would weaken the task-inconformity explanation.
  • Because the learned attention maps recover physically expected couplings (U10–V10, T2M–TCC), this latent-space design could double as an interpretability probe for which variables actually drive forecast skill in other multivariate physical systems, such as air quality or ocean state.
  • If the two-stage alignment is the real driver, the recipe should transfer to any heterogeneous multi-channel forecasting problem—traffic flow plus occupancy, energy demand plus weather covariates—and the cheapest validation is to run the identical pipeline on such data with the same encoder–decoder counts.
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. This paper proposes Met2Net, a multivariate spatiotemporal forecasting model in which each meteorological variable has a separate encoder and decoder, and a shared Translator (implemented with TAU blocks) performs variable attention and temporal modeling in the latent space. The central methodological contribution is an 'implicit two-stage' training procedure: in each iteration, one module group is updated by backpropagation while the other is updated via momentum, and the loss is the sum of an output-space MSE and a latent-space prediction MSE between momentum-encoded targets and predicted latents. The authors evaluate on WeatherBench (low-resolution, high-resolution, and high-altitude), a cropped ERA5 dataset, a synthetic multi-channel MNIST dataset, and TaxiBJ, and report state-of-the-art numbers, including 28.82% and 23.39% MSE reductions for T2M and relative humidity relative to TAU.

Significance. If the causal claims were fully supported, the contribution would be significant for multivariate weather forecasting: a clean recipe (per-variable encoders/decoders + latent prediction loss + variable attention) that outperforms single-variable spatiotemporal baselines without changing the inference-time architecture. The paper ships code, reports a wide range of metrics, and includes a new multi-channel benchmark (MvMmfnist), which is commendable. However, the significance is currently limited by the ablations: the table that is supposed to validate the training paradigm changes two factors at once, and the headline comparison against TAU is confounded by the number of models and input variables. The reported gains are promising but their attribution is not yet established.

major comments (4)
  1. [Sec. 3.3, Eq. (2), Table 8] The ablation in Table 8 does not isolate the proposed implicit two-stage mechanism. The row '+ITS' adds both the latent prediction loss L2 and the momentum/stop-gradient decoupling relative to '+VA', so the observed improvement could come from the auxiliary deep-supervision loss alone. The paper never reports a control trained with L1+L2 in a fully end-to-end manner (with gradients flowing through Encoder, Translator, and Decoder and no freezing or momentum). Without that control, the central claim that the decoupled two-stage training resolves 'task inconformity' (Sec. 3.3) is not causally supported. This is load-bearing because the two-stage training is the paper's main contribution.
  2. [Sec. 4.2, Table 1] The headline comparison against TAU is not apples-to-apples. In Table 1, TAU is listed with N=4, meaning four separate single-variable models are trained and their metrics are reported individually, while Met2Net has N=1 and receives all four variables as input. The reported MSE reductions of 28.82% and 23.39% therefore conflate the benefit of the proposed architecture with the benefit of additional input information. The 'Baseline' row in Table 8 is a naive channel-stack variant that degrades performance, not a properly tuned multivariable TAU with per-variable normalization. A fair multivariable baseline must be included before the superiority of the method can be claimed.
  3. [Sec. 4.2 'Tracking Tropical Cyclones', Table 5] The tropical cyclone experiment is not quantitatively evaluated as claimed. Figure 6 shows one typhoon (Mawar) and the text states that Met2Net 'achieves higher accuracy with a 3-hour lead time', but Table 5 reports MSE/MAE/R2 on the cropped ERA5 grid for four variables, not track forecast error (e.g., distance error in km). A single qualitative trajectory plot is insufficient evidence for improved cyclone track prediction. The authors should either provide track-error metrics for multiple typhoons or temper the claim to 'field-level improvement on a cyclone-affected region'.
  4. [Sec. 3.3, Eq. (1), Algorithm 2] The description of the method as 'freezing' is inconsistent with the implementation. Eq. (1) and Algorithm 2 update the so-called frozen modules with momentum (theta_m <- alpha*theta_m + (1-alpha)*theta), so the frozen parameters are not static; they are an exponential moving average of the trained parameters. This is a legitimate design, but it means the method is not literally a two-stage freeze-train procedure. The paper should clarify whether the momentum update is essential, and should report sensitivity to the momentum coefficient alpha (currently fixed to 0.999 in Sec. 4.1) and a comparison against a plain stop-gradient baseline with truly static frozen parameters.
minor comments (6)
  1. [Abstract] The abstract contains the typo 'resectively' (should be 'respectively') and the phrase 'In detailed' (should be 'In detail').
  2. [Figure 2 caption] The sentence 'The Translator module is responsible for learning spatiotemporal features.' is duplicated in the caption.
  3. [Sec. 3.3] The sentence 'However, the Translator and Encoder & Decoder cannot be optimized alternatively with naive stop-gradient operation' is unclear; please explain the failure mode or provide a reference for this claim.
  4. [Table 8] The 'Baseline' row is not defined in the main text; state explicitly whether it is TAU with channel-stacked variables and how the variables are normalized.
  5. [Sec. 4.1] The momentum coefficient alpha is a free hyperparameter, but no sensitivity analysis for it is provided; consider adding an ablation to Table 8 or the appendix.
  6. [Appendix, Table 11] Met2Net uses 119.0 GFLOPs and 23,078 MiB versus TAU's 18.3 GFLOPs and 11,942 MiB; the main text should acknowledge this increased training cost rather than implying competitive efficiency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical, benchmarked on held-out external data against independent baselines, and no result reduces by construction to its own inputs.

full rationale

Met2Net's central claims are empirical: on WeatherBench, cropped ERA5, TaxiBJ, and its MvMmfnist dataset, it reports lower MSE/MAE/RMSE than independently published baselines such as ConvLSTM, PredRNN, SimVP, and TAU. These numbers are computed on held-out test splits, so the headline improvements are not algebraic consequences of the model definition. The two-stage training objective L = L1(Y',Y) + L2(Z'_y,Zy) does use the model's own momentum encoder to define the latent target Zy = E(Y), but this is a standard self-supervised / teacher-student training mechanism, not a derived prediction that reduces to its own input; the paper does not claim L2 alone constitutes the forecast or that the benchmark results follow from the loss definition. The only self-citation (Ref. [8], the authors' STAA paper) appears in passing related-work context and is not load-bearing for the proposed method; no uniqueness theorem or ansatz is imported from prior work by the same authors. The ablation in Table 8 changes both the auxiliary latent loss and the stop-gradient/momentum schedule simultaneously, which weakens causal attribution of the gains to 'implicit two-stage training' specifically, but this is an experimental confound (a correctness risk), not circularity, because the reported gains are still measured against external baselines and are not fitted in-sample. Accordingly, no circular step can be exhibited with the required specificity, and the paper's derivation chain is self-contained with respect to circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method has one hand-set momentum coefficient and a learning-rate selection, plus two domain assumptions about meteorological variables as modalities and about the two-stage objective alignment. No new physical entities or fitted scientific constants are introduced.

free parameters (2)
  • Momentum coefficient alpha = 0.999
    Set by hand in Section 4.1; controls the exponential moving average in Eq. 1 and is not fitted to the data or derived from theory.
  • Learning rate for MvMmfnist = chosen from {1e-4, 1e-3, 5e-3}
    The appendix reports selecting the best learning rate from this set for the synthetic dataset, a hyperparameter choice rather than a fitted physical constant.
assumptions (3)
  • standard math Softmax attention over the variable axis (Eq. 7) provides a valid fusion mechanism.
    The paper applies the standard scaled dot-product attention formula with 2D-CNN-extracted queries, keys, and values.
  • domain assumption Meteorological variables are heterogeneous enough to be treated as distinct modalities requiring separate encoders and decoders.
    Section 3.4 asserts this based on Figure 1, which shows different distributions; no formal criterion is given for when separate encoders help.
  • ad hoc to paper Freezing one module group and momentum-updating it resolves 'task inconformity' between reconstruction and prediction.
    Section 3.3 and Table 8 provide empirical support, but the mechanism is not proven; Algorithm 2 actually updates both groups in one backward pass.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Met$^2$Net: A Decoupled Two-Stage Spatio-Temporal Forecasting Model for Complex Meteorological Systems." pith.science (2026). https://pith.science/paper/B2MLBU2K

@misc{pith2026250717189,
  author       = {Pith},
  title        = {Pith review of: Met$^2$Net: A Decoupled Two-Stage Spatio-Temporal Forecasting Model for Complex Meteorological Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2MLBU2K}},
  note         = {Machine review of arXiv:2507.17189}
}
read the original abstract

The increasing frequency of extreme weather events due to global climate change urges accurate weather prediction. Recently, great advances have been made by the \textbf{end-to-end methods}, thanks to deep learning techniques, but they face limitations of \textit{representation inconsistency} in multivariable integration and struggle to effectively capture the dependency between variables, which is required in complex weather systems. Treating different variables as distinct modalities and applying a \textbf{two-stage training approach} from multimodal models can partially alleviate this issue, but due to the inconformity in training tasks between the two stages, the results are often suboptimal. To address these challenges, we propose an implicit two-stage training method, configuring separate encoders and decoders for each variable. In detailed, in the first stage, the Translator is frozen while the Encoders and Decoders learn a shared latent space, in the second stage, the Encoders and Decoders are frozen, and the Translator captures inter-variable interactions for prediction. Besides, by introducing a self-attention mechanism for multivariable fusion in the latent space, the performance achieves further improvements. Empirically, extensive experiments show the state-of-the-art performance of our method. Specifically, it reduces the MSE for near-surface air temperature and relative humidity predictions by 28.82\% and 23.39\%, respectively. The source code is available at https://github.com/ShremG/Met2Net.

Figures

Figures reproduced from arXiv: 2507.17189 by the authors.

Figure 1
Figure 1. (a) Spatiotemporal distributions of the meteorological [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) End-to-End (E2E) training strategy. (b) Generative [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Training pipeline and Inference pipeline. The training pipeline consists of two stages: In Stage I, the Translator (blue snowflake [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Visualization of prediction results for different lead times. (a) Results at a forecast time of 1 hour. The background in white [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Met [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Comparison of predicted and ground truth tracks of Ty [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Performance comparison of T2M and TCC prediction [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 9
Figure 9. Figure 9: Tracking tropical cyclones. 6.11. Additional visualization results [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 8
Figure 8. Figure 8: Cross-variable attention map extracted from the transla [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 10
Figure 10. Figure 10: Visualization of prediction results for different lead times. (a) Results at a forecast time of 1 hour. The background in white [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Visualization of prediction results for different lead times. (a) Results at a forecast time of 1 hour. The background in white [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Visualization of prediction results for different lead times. (a) Results at a forecast time of 1 hour. The background in white [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Visualization of prediction results for different lead times. (a) Results at a forecast time of 1 hour. The background in white [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]
Figure 14
Figure 14. Figure 14: Visualization of prediction results for different lead times on the Mv [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: Visualization of prediction results for different lead times on the Mv [PITH_FULL_IMAGE:figures/full_fig_p016_15.png]
Figure 16
Figure 16. Figure 16: Visualization of prediction results for different lead times on the Mv [PITH_FULL_IMAGE:figures/full_fig_p016_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

65 extracted references · 52 canonical work pages

  1. [1]

    Accurate medium-range global weather forecasting with 3d neural networks

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiao- tao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619(7970): 533–538, 2023. 1, 3

    work page 2023

  2. [2]

    Align your latents: High-resolution video synthesis with la- tent diffusion models

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dock- horn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with la- tent diffusion models. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 22563–22575, 2023. 4

    work page 2023

  3. [3]

    Unsupervised learning of visual features by contrasting cluster assignments

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Pi- otr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Ad- vances in neural information processing systems , 33:9912– 9924, 2020. 3

    work page 2020

  4. [4]

    Mau: A motion- aware unit for video prediction and beyond

    Zheng Chang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, Yan Ye, Xiang Xinguang, and Wen Gao. Mau: A motion- aware unit for video prediction and beyond. Advances in Neural Information Processing Systems , 34:26950–26962,

    work page

  5. [5]

    Fengwu: Pushing the skillful global medium- range weather forecast beyond 10 days lead

    Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium- range weather forecast beyond 10 days lead. arXiv preprint arXiv:2304.02948, 2023. 1, 3

    arXiv 2023

  6. [6]

    Mixed autoencoder for self- supervised visual representation learning

    Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Zhen- guo Li, and Dit-Yan Yeung. Mixed autoencoder for self- supervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22742–22751, 2023. 2, 3

    work page 2023

  7. [7]

    Fuxi: A cascade ma- chine learning forecasting system for 15-day global weather forecast

    Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: A cascade ma- chine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science , 6(1):190,

    work page

  8. [8]

    Staa: Spatiotemporal alignment attention for short-term precipita- tion forecasting

    Min Chen, Hao Yang, Shaohan Li, and Xiaolin Qin. Staa: Spatiotemporal alignment attention for short-term precipita- tion forecasting. IEEE Geoscience and Remote Sensing Let- ters, 21:1–5, 2024. 1, 2, 3

    work page 2024

Show all 65 references
  1. [9]

    Prompt federated learning for weather forecasting: Toward foundation models on meteorological data

    Shengchao Chen, Guodong Long, Tao Shen, and Jing Jiang. Prompt federated learning for weather forecasting: Toward foundation models on meteorological data. arXiv preprint arXiv:2301.09152, 2023. 2

    2023 arXiv

  2. [10]

    The atlas of mortality and economic losses from weather, climate and water extremes (1970-2019)

    James Douris and Geunhye Kim. The atlas of mortality and economic losses from weather, climate and water extremes (1970-2019). 2021. 1

    1970

  3. [11]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873–12883, 2021. 2

    2021

  4. [12]

    Realistic tropical cyclone wind and pressure fields can be reconstructed from sparse data using deep learn- ing

    Ryan Eusebi, Gabriel A Vecchi, Ching-Yao Lai, and Mingjing Tong. Realistic tropical cyclone wind and pressure fields can be reconstructed from sparse data using deep learn- ing. Communications Earth & Environment, 5(1):8, 2024. 1, 2, 3

    2024

  5. [13]

    Spatiotemporal attention for multivariate time series prediction and interpretation

    Tryambak Gangopadhyay, Sin Yong Tan, Zhanhong Jiang, Rui Meng, and Soumik Sarkar. Spatiotemporal attention for multivariate time series prediction and interpretation. In ICASSP 2021-2021 IEEE international conference on acous- tics, speech and signal processing (ICASSP) , pages 3560–

    2021

  6. [14]

    Simvp: Simpler yet better video prediction

    Zhangyang Gao, Cheng Tan, Lirong Wu, and Stan Z Li. Simvp: Simpler yet better video prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3170–3180, 2022. 3, 5, 6, 7, 8

    2022

  7. [15]

    Bootstrap your own latent-a new approach to self-supervised learning

    Jean-Bastien Grill, Florian Strub, Florent Altch ´e, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Ghesh- laghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neur...

    2020

  8. [16]

    Disentangling physi- cal dynamics from unknown factors for unsupervised video prediction

    Vincent Le Guen and Nicolas Thome. Disentangling physi- cal dynamics from unknown factors for unsupervised video prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 11474– 11484, 2020. 1, 3

    2020

  9. [17]

    Precipitation nowcasting using ground radar data and simpler yet better video prediction deep learn- ing

    Daehyeon Han, Minki Choo, Jungho Im, Yeji Shin, Juhyun Lee, and Sihun Jung. Precipitation nowcasting using ground radar data and simpler yet better video prediction deep learn- ing. GIScience & Remote Sensing, 60(1):2203363, 2023. 1, 3

    2023

  10. [18]

    Momentum contrast for unsupervised visual rep- resentation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual rep- resentation learning. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 9729–9738, 2020. 2, 3

    2020

  11. [19]

    Imagen video: High definition video generation with diffusion mod- els

    Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion mod- els. arXiv preprint arXiv:2210.02303, 2022. 4

    arXiv

  12. [20]

    A dynamic multi-scale voxel flow network for video prediction

    Xiaotao Hu, Zhewei Huang, Ailin Huang, Jun Xu, and Shuchang Zhou. A dynamic multi-scale voxel flow network for video prediction. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 6121–6131, 2023. 3

    2023

  13. [21]

    Relative importance of meteorological vari- ables on air quality and role of boundary layer height

    Yaxuan Huang, Bin Guo, Haoxuan Sun, Huijie Liu, and Song Xi Chen. Relative importance of meteorological vari- ables on air quality and role of boundary layer height. Atmo- spheric Environment, 267:118737, 2021. 2

    2021

  14. [22]

    Applicability analysis of transformer to wind speed forecasting by a novel deep learn- ing framework with multiple atmospheric variables

    Wenjun Jiang, Bo Liu, Yang Liang, Huanxiang Gao, Pengfei Lin, Dongqin Zhang, and Gang Hu. Applicability analysis of transformer to wind speed forecasting by a novel deep learn- ing framework with multiple atmospheric variables. Applied Energy, 353:122155, 2024. 3, 4

    2024

  15. [23]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 6

    2015

  16. [24]

    Videopoet: A large language model for zero-shot video generation

    Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jos ´e Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2, 3, 4

    2023 arXiv

  17. [25]

    Similarity of neural network represen- tations revisited

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network represen- tations revisited. In International conference on machine learning, pages 3519–3529. PMLR, 2019. 7

    2019

  18. [26]

    Learning skillful medium-range global weather forecasting

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416–1421, 2023. 4

    2023

  19. [27]

    A comprehensive sur- vey on design and application of autoencoder in deep learn- ing

    Pengzhi Li, Yan Pei, and Jianqiang Li. A comprehensive sur- vey on design and application of autoencoder in deep learn- ing. Applied Soft Computing, 138:110176, 2023. 2, 3

    2023

  20. [28]

    Moganet: Multi-order gated aggregation network

    Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z Li. Moganet: Multi-order gated aggregation network. In Inter- national Conference on Learning Representations, 2024. 8

    2024

  21. [29]

    Moganet: Multi-order gated aggregation network

    Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z Li. Moganet: Multi-order gated aggregation network. In ICLR,

  22. [30]

    Mergevq: A unified framework for visual generation and representation with dis- entangled token merging and quantization

    Siyuan Li, Luyuan Zhang, Zedong Wang, Juanxi Tian, Cheng Tan, Zicheng Liu, Chang Yu, Qingsong Xie, Hao- nan Lu, Haoqian Wang, and Zhen Lei. Mergevq: A unified framework for visual generation and representation with dis- entangled token merging and quantization. InProceedings o...

    2025

  23. [31]

    Conditional local convolution for spatio- temporal meteorological forecasting

    Haitao Lin, Zhangyang Gao, Yongjie Xu, Lirong Wu, Ling Li, and Stan Z Li. Conditional local convolution for spatio- temporal meteorological forecasting. In Proceedings of the AAAI conference on artificial intelligence, pages 7470–7478,

  24. [32]

    Spacetime separable la- tent diffusion model with intensity structure information for precipitation nowcasting

    XuDong Ling, ChaoRong Li, LiHong Zhu, FengQing Qin, Ping Zhu, and Yuanyuan Huang. Spacetime separable la- tent diffusion model with intensity structure information for precipitation nowcasting. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1

    2024

  25. [33]

    A unified multi-step wind speed forecasting framework based on numerical weather prediction grids and wind farm monitoring data

    Xingdou Liu, Li Zhang, Jiangong Wang, Yue Zhou, and Wei Gan. A unified multi-step wind speed forecasting framework based on numerical weather prediction grids and wind farm monitoring data. Renewable Energy, 211:948–963, 2023. 4

    2023

  26. [34]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 11976–11986,

  27. [35]

    Histgnn: Hierarchical spatio- temporal graph neural network for weather forecasting

    Minbo Ma, Peng Xie, Fei Teng, Bin Wang, Shenggong Ji, Junbo Zhang, and Tianrui Li. Histgnn: Hierarchical spatio- temporal graph neural network for weather forecasting. In- formation Sciences, 648:119580, 2023. 4

    2023

  28. [36]

    Latte: Latent diffusion transformer for video generation

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zi- wei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2, 3

    2024 arXiv

  29. [37]

    Sensitivity of meteorological-forcing resolution on hy- drologic variables

    Fadji Z Maina, Erica R Siirila-Woodburn, and Pouya Vah- mani. Sensitivity of meteorological-forcing resolution on hy- drologic variables. Hydrology and Earth System Sciences, 24 (7):3451–3474, 2020. 2

    2020

  30. [38]

    Conditional image-to-video gener- ation with latent flow diffusion models

    Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video gener- ation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444–18455, 2023. 2, 3

    2023

  31. [39]

    Triplet attention transformer for spatiotemporal predictive learning

    Xuesong Nie, Xi Chen, Haoyuan Jin, Zhihang Zhu, Yun- feng Yan, and Donglian Qi. Triplet attention transformer for spatiotemporal predictive learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7036–7045, 2024. 3

    2024

  32. [40]

    Wavelet-driven spatiotemporal predictive learning: Bridg- ing frequency and time variations

    Xuesong Nie, Yunfeng Yan, Siyuan Li, Cheng Tan, Xi Chen, Haoyuan Jin, Zhihang Zhu, Stan Z Li, and Donglian Qi. Wavelet-driven spatiotemporal predictive learning: Bridg- ing frequency and time variations. In Proceedings of the AAAI Conference on Artificial Intelligence , pages ...

    2024

  33. [41]

    Recurrent neural network modeling of multivariate time series and its application in temperature forecasting

    Edward Appau Nketiah, Li Chenlong, Jing Yingchuan, and Simon Appah Aram. Recurrent neural network modeling of multivariate time series and its application in temperature forecasting. Plos one, 18(5):e0285713, 2023. 2

    2023

  34. [42]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3

    2023 arXiv

  35. [43]

    An interpretable multi-stage forecasting framework for energy consumption and co2 emissions for the transportation sector

    Qingyao Qiao, Hamidreza Eskandari, Hassan Saadatmand, and Mohammad Ali Sahraei. An interpretable multi-stage forecasting framework for energy consumption and co2 emissions for the transportation sector. Energy, 286:129499,

  36. [44]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  37. [45]

    Hornet: Efficient high- order spatial interactions with recursive gated convolutions

    Yongming Rao, Wenliang Zhao, Yansong Tang, Jie Zhou, Ser Nam Lim, and Jiwen Lu. Hornet: Efficient high- order spatial interactions with recursive gated convolutions. Advances in Neural Information Processing Systems , 35: 10353–10366, 2022. 5, 6, 7, 8

    2022

  38. [46]

    Weather- bench: a benchmark data set for data-driven weather fore- casting

    Stephan Rasp, Peter D Dueben, Sebastian Scher, Jonathan A Weyn, Soukayna Mouatadid, and Nils Thuerey. Weather- bench: a benchmark data set for data-driven weather fore- casting. Journal of Advances in Modeling Earth Systems, 12 (11):e2020MS002203, 2020. 6

    2020

  39. [47]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 3, 4

    2022

  40. [48]

    Convolutional lstm network: A machine learning approach for precipitation nowcasting

    Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional lstm network: A machine learning approach for precipitation nowcasting. Advances in neural information processing sys- tems, 28, 2015. 2, 5, 6, 7, 8

    2015

  41. [49]

    Unsupervised learning of video representations using lstms

    Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudi- nov. Unsupervised learning of video representations using lstms. In International conference on machine learning , pages 843–852. PMLR, 2015. 8

    2015

  42. [50]

    Temporal attention unit: To- wards efficient spatiotemporal predictive learning

    Cheng Tan, Zhangyang Gao, Lirong Wu, Yongjie Xu, Jun Xia, Siyuan Li, and Stan Z Li. Temporal attention unit: To- wards efficient spatiotemporal predictive learning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770–18782, 2023....

    2023

  43. [51]

    Open- stl: A comprehensive benchmark of spatio-temporal predic- tive learning

    Cheng Tan, Siyuan Li, Zhangyang Gao, Wenfei Guan, Ze- dong Wang, Zicheng Liu, Lirong Wu, and Stan Z Li. Open- stl: A comprehensive benchmark of spatio-temporal predic- tive learning. Advances in Neural Information Processing Systems, 36:69819–69831, 2023. 1, 6

    2023

  44. [52]

    Hydrologic extremes in a changing climate: a review of extremes in east africa

    Meron Teferi Taye and Ellen Dyer. Hydrologic extremes in a changing climate: a review of extremes in east africa. Cur- rent Climate Change Reports, 10(1):1–11, 2024. 2

    2024

  45. [53]

    Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms

    Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and Philip S Yu. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in neural information processing systems, 30, 2017. 2, 5, 6, 7, 8

    2017

  46. [54]

    Predrnn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learn- ing

    Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and S Yu Philip. Predrnn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learn- ing. In International conference on machine learning, pages 5123–5132. PMLR, 2018. 2, 5, 6, 8

    2018

  47. [55]

    Eidetic 3d lstm: A model for video prediction and beyond

    Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Ming- sheng Long, and Li Fei-Fei. Eidetic 3d lstm: A model for video prediction and beyond. In International conference on learning representations, 2018. 3

    2018

  48. [56]

    Memory in memory: A predictive neural network for learning higher-order non- stationarity from spatiotemporal dynamics

    Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order non- stationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF conference on computer vision and patter...

    2019

  49. [57]

    A review of wind speed and wind power fore- casting with deep neural networks

    Yun Wang, Runmin Zou, Fang Liu, Lingjun Zhang, and Qianyi Liu. A review of wind speed and wind power fore- casting with deep neural networks. Applied Energy, 304: 117766, 2021. 4

    2021

  50. [58]

    Predrnn: A recurrent neural network for spatiotemporal predictive learn- ing

    Yunbo Wang, Haixu Wu, Jianjin Zhang, Zhifeng Gao, Jian- min Wang, S Yu Philip, and Mingsheng Long. Predrnn: A recurrent neural network for spatiotemporal predictive learn- ing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2208–2225, 2022. 3

    2022

  51. [59]

    ivideogpt: Interac- tive videogpts are scalable world models

    Jialong Wu, Shaofeng Yin, Ningya Feng, Xu He, Dong Li, Jianye Hao, and Mingsheng Long. ivideogpt: Interac- tive videogpts are scalable world models. arXiv preprint arXiv:2405.15223, 2024. 2

    2024 arXiv

  52. [60]

    Fashion- mnist: a novel image dataset for benchmarking machine learning algorithms

    Han Xiao, Kashif Rasul, and Roland V ollgraf. Fashion- mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 8

    2017 arXiv

  53. [61]

    Videogpt: Video generation using vq-vae and trans- formers

    Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and trans- formers. arXiv preprint arXiv:2104.10157, 2021. 2

    2021 arXiv

  54. [62]

    Magvit: Masked generative video transformer

    Lijun Yu, Yong Cheng, Kihyuk Sohn, Jos ´e Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming- Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p...

    2023

  55. [63]

    Video probabilistic diffusion models in projected latent space

    Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 18456–18466,

  56. [65]

    Experimental Setup for Variable Distributions Data Distributions

    Appendix 6.1. Experimental Setup for Variable Distributions Data Distributions. The experiment utilized the 2018 T2M (2-meter air temperature) and TCC (Total Cloud Cover) data from the WeatherBench dataset to analyze spa- tial and temporal distribution patterns. The data were ...

    2018

  57. [2023]

    # E1, E2: encoders

    2, 3 Met2Net: A Decoupled Two-Stage Spatio-Temporal Forecasting Model for Complex Meteorological Systems Supplementary Material Algorithm 1 Pseudocode of Implicit Two-Stage Process in a PyTorch-like Style Integrated Within a Inference Pipeline. # E1, E2: encoders. # D1, D2: de...

Pith tools

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