Pith. sign in

REVIEW 3 major objections 6 minor 28 references

Latent State Inference in a Spatiotemporal Generative Model

T0 review · 3 major / 6 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read A distributed spatiotemporal network can infer hidden static causal factors from observed dynamics, including a never-seen continent's land-sea mask.

desk verdict The wave experiment genuinely supports the latent-inference mechanism, but the WeatherBench claim that AT improves predictions is unsupported by the reported experiments. read the letter →

arxiv 2009.09823 v2 pith:AFY6LTW3 submitted 2020-09-21 cs.LG stat.ML

Matthias Karlbauer , Tobias Menge , Sebastian Otte , Hendrik P.A. Lensch , Thomas Scholten , Volker Wulfmeyer , Martin V. Butz This is my paper · ORCID
classification cs.LGstat.ML
keywords spatiotemporalpredictionlatentstateinferencegraphneuralnetworkactivetuninghiddencausalfactorsweatherforecastingland-seamask
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 seeks to show that hidden, location-specific causes of spatiotemporal dynamics can be recovered from the observed dynamics alone. The authors combine a distributed graph neural network, DISTANA, with active tuning, a retrospective gradient-based procedure that adjusts a static input per grid cell to reduce local prediction errors. In a 2D wave benchmark, the combined model infers the local wave-speed map, preserves the ordering of wave speeds never seen in training, and predicts more accurately than ConvLSTM or TCN with fewer parameters. In a global weather benchmark, it infers the missing third of Earth's land-sea mask, including a continent never shown during training, from temperature dynamics alone and uses that inference to improve its own forecasts. If correct, the paper establishes a general recipe for latent causal-factor inference in spatiotemporal processes.

What carries the argument

The machinery is active tuning applied to DISTANA's per-cell static input. DISTANA arranges shared-weight prediction kernels in a grid; each kernel is a recurrent LSTM-like unit that takes local dynamic input, lateral messages from neighboring cells, and a static input, modeling universal local dynamics. Active tuning decouples that static input from the direct signal, starts it at zero, and iteratively optimizes it by backpropagating the prediction error over a recent history, with the static input clamped to a fixed range in the weather experiments. A normalization step over the inferred static field, with a compensating update to the preprocessing layer's weights and bias, prevents drift and stabilizes inference. The static input therefore acts as a learned parametric bias that absorbs location-specific causal information.

What would settle it

Run DISTANA+AT on a synthetic dataset with two confounded static factors that both affect the observed dynamics, such as local wave speed and local damping, and check whether the inferred latent factors separate them. If the inferred static input is the same for factor pairs that produce identical observed fields, or if it only reproduces one monotone combination, the claim that the inferred factors are the hidden causes is falsified; recovering the two factors individually would support it.

Watch

Extended reading notes

Core claim

Using only observed dynamics, the DISTANA+AT model derives local static factors that shape those dynamics. The static input is initialized at zero and then optimized by active tuning over prediction-error gradients, while the shared prediction kernels learn the universal dynamics. In the wave task the inferred context map matches the true speed-modulating field up to a monotone transform, including values never seen during training. In the weather task, after the model was trained on two thirds of the land-sea mask, retrospective tuning recovered the remaining third, including the Americas, from T850 temperature observations, and feeding that inferred mask back improved the model's future temperature predictions. The authors conclude that projecting gradient information onto static neural activities identifies local parametric biases that characterize hidden causal factors, and that these factors tend to be causal because they form to improve prediction accuracy.

Load-bearing premise

The central assumption is that optimizing the local static input to reduce prediction error recovers the true hidden causal factor, rather than any other latent variable that also improves prediction, and that the learned effect of a factor generalizes to locations where it was never observed.

Editorial extensions

If this is right

  • DISTANA+AT produces more accurate closed-loop spatiotemporal predictions than ConvLSTM and TCN with fewer parameters, because the self-inferred static context keeps dynamics stable over long horizons.
  • Because the static context is inferred by optimizing local prediction errors, the model can recover hidden factors it was never given, such as the local wave-speed map or the missing third of Earth's land-sea mask.
  • On the weather benchmark, DISTANA achieves an RMSE of 816 for Z500 at a 72-hour lead time, better than the best comparable iterative CNN baseline, while remaining below physics-based numerical weather prediction.
  • Freely tuned latent neurons develop encodings that resemble physically meaningful fields, such as land-sea mask, latitude, albedo, and orography, with the encoding depending on the predicted variable and the static inputs supplied.
  • The approach assumes hidden causes are constant in time and universal in effect; under that assumption, inference works during both training and testing.

Reading between the lines

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

  • The wave experiment's inferred range of -0.6 to 2.6 for a true context in [0,1] suggests the recovered factors are determined only up to a monotone transform, so converting them to physical units would require calibration or additional observations, not just prediction error.
  • The same inference recipe could be applied to other spatiotemporal systems with quasi-constant local properties, such as soil properties, traffic demand, or energy-grid bottlenecks, where the hidden factor is static but varies across space.
  • A natural testable extension is to vary the number of independently tunable latent neurons per cell; the paper's third weather experiment suggests more neurons can develop disentangled physical codes, which could be checked by correlating each latent neuron with held-out geographic variables.
  • If the identifiability limitation can be overcome, the approach offers a path toward causal discovery in settings where interventions are impossible, because the inferred per-location factors could be used as hypotheses for controlled experiments.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper combines the DISTANA spatiotemporal graph neural network with active tuning (AT), a retrospective gradient-based optimization of a static latent input, in order to infer location-specific hidden causal factors from observed dynamics. Two application domains are considered: a synthetic two-dimensional wave propagation task with locally varying propagation speeds, and the WeatherBench weather prediction benchmark. For the wave task, the authors report that DISTANA+AT outperforms ConvLSTM and TCN baselines and that the inferred static context preserves the monotone ordering of the true wave-speed factor. For WeatherBench, they report that a model trained with two thirds of the land-sea mask can infer the missing third (the Americas) from T850 temperature dynamics, and they present latent codes that appear to encode land-sea and latitude information. The paper claims that DISTANA+AT 'reliably derive[s] location-respective hidden causal factors' and that the inferred land-sea mask improves the model's own temperature predictions.

Significance. If the claims are fully supported, the combination of DISTANA with active tuning would be a noteworthy contribution to unsupervised latent-state inference in spatiotemporal modeling, and the WeatherBench results would demonstrate a practical use of self-inferred static context. The paper has strengths: it provides a concrete algorithmic modification (AT on a spatially distributed static input with the normalization update of Eq. (1)), it compares against reasonable baselines on the wave task, and the WeatherBench Z500 result (RMSE 816 vs 1114 for the iterative CNN baseline) is a useful data point. However, the central claim about reliable derivation of hidden causal factors is currently supported mainly by qualitative inspection and by a confounded performance comparison, so the significance is conditional on the empirical gaps identified below.

major comments (3)
  1. [§3.1–§3.2] The wave-experiment comparison between DISTANA and DISTANA+AT is confounded by test-time AT inference. As described in §3.1, DISTANA+AT is given 50 separate test sequences to infer the static context before evaluation, using H=30, c=1, and a dedicated inference learning rate, while plain DISTANA receives no analogous test-time adaptation. The improved closed-loop MSE in §3.2 (3.87e-4 versus 8.69e-4) could therefore be due to test-time optimization of an extra static input rather than to the model's ability to identify the true causal factor. The authors should report an ablation in which DISTANA+AT is evaluated without the test-time tuning loop, or in which an equally flexible test-time adaptation mechanism is added to the baselines.
  2. [§3.4 and Abstract] The abstract's claim that the inferred land-sea mask 'uses the self inferred information to improve its own future temperature predictions' is not supported by any reported quantitative comparison in §3.4. The second WeatherBench experiment infers the missing third of the LSM, but no results are shown for T850 prediction error with versus without AT-inferred LSM, with the true full LSM, or with no LSM at all. Additionally, since the model was trained with two thirds of the LSM together with orography, soil type, latitude/longitude, tisr, and boundary flags, the inferred American LSM could be reconstructed from correlations among these static inputs rather than from an identified hidden causal factor. A quantitative ablation isolating the contribution of the AT-inferred LSM is needed.
  3. [§2.2 and §4] The identifiability of hidden causal factors from prediction-error gradients is an open issue that the paper itself acknowledges in §4 ('future research will need to investigate the robustness of this tendency'), yet the abstract and §1 state that DISTANA+AT 'can reliably derive location-respective hidden causal factors.' The wave experiment only establishes a monotone transform: the inferred values in Figure 3 range from about -0.6 to 2.6 while the true context values range from 0 to 1, so the method does not recover the factor itself, only an order-preserving function of it. The authors should either soften the causal-identification claim to match the evidence or provide a test that distinguishes the true causal factor from other latent variables that also reduce prediction error.
minor comments (6)
  1. [§3.2] The sentence reporting results states that ConvLSTM achieves (8.69±0.87)×10^-4 and DISTANA achieves the identical value (8.69±0.87)×10^-4; this looks like a typo, since the same numbers cannot meaningfully describe two different models.
  2. [§3.4] 'RSME' should be 'RMSE' (root mean squared error).
  3. [§2.2] 'Intern Covariate Drift' should be 'Internal Covariate Shift'.
  4. [§3.3–§3.4] The active-tuning hyperparameters for the WeatherBench LSM-inference experiment are not specified (wave values are c=1, H=30, eta=0.1/0.01); the authors should report the corresponding values for the weather experiments.
  5. [§3.4] The notation 'sLSM' is used without a formal definition; the text says the 'entire LSM vector' is initialized with zeros, but the relation between the LSM input channel, the inferred vector, and the nine static inputs should be stated explicitly.
  6. [Figure 5] The inference quality of the missing LSM is described as 'reasonable' and the latent code as showing a 'decent latitude coding'; a quantitative correlation or similarity metric between the inferred and ground-truth LSM would make the claim more precise.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; wave and WeatherBench external checks are independent of AT's objective, though the weather 'improvement' claim lacks an ablation and identifiability is conceded.

full rationale

The central derivation is not circular. DISTANA and active tuning come from prior work by the same group, but they are used as tools; the paper's key claims are checked against external quantities that are not part of the AT objective. In the wave benchmark, the model infers unseen propagation-speed contexts (0.2 to 1.0) and the inferred values preserve monotone order, while prediction MSE for DISTANA+AT is lower than for DISTANA without AT (3.87e-4 vs 8.69e-4). In WeatherBench, the model is trained on two-thirds of the land-sea mask and the held-out American third is reconstructed from T850 dynamics and compared visually to the true mask. Neither check is forced by the AT loss. The main limitation is explicitly acknowledged in Section 4: the authors state that 'future research will need to investigate the robustness of this tendency' for the claim that hidden factors identify causal influences; this is an identifiability caveat, not a circular step. The abstract's wording that DISTANA 'uses the self inferred information to improve its own future temperature predictions' is not backed by a quantitative ablation in Section 3.4, but it is not forced by construction, since s_LSM is tuned on observed dynamics and could in principle fail on forward rollouts. The self-citations to [13], [17], and related AT work are normal method attribution and are not load-bearing for the external benchmark results. The score of 2 reflects minor self-citation presence and the unsupported abstract wording, not a circular derivation.

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

The central machinery rests on four assumptions: shared universal dynamics across space, constancy of hidden factors, identifiability of those factors from prediction error, and transferability of the learned static-input encoding. The latent s is a fitted free parameter per location; its recovered values are not unique in scale, as the wave experiment shows. No new physical entities are postulated.

free parameters (4)
  • Local latent static context vector s = wave: inferred values range [-0.6, 2.6] vs ground truth [0,1]; weather: clamped to [-1,1]
    Optimized by active tuning to minimize prediction error; this is the central quantity being inferred. The wave experiment shows only a monotone transform of the true context is recovered, so absolute scale is not identifiable.
  • Active tuning hyperparameters (c, H, eta) = c=1, H=10 during training, H=30 during wave testing, eta=0.1 for first three epochs and 0.01 for remaining seven
    Chosen by hand; these control the convergence and quality of the inferred latent state, and the paper does not systematically explore their sensitivity.
  • Static input preprocessing normalization weights W and bias b = Updated via b <- b + W*mu_s; W <- W*sigma_s (Eq. 1)
    Introduced to stabilize AT inference after normalization; this modification changes gradient flow asymmetrically and is essential to the reported convergence.
  • Scheduled sampling schedule = Linear slope of 270 epochs from teacher forcing probability 0.0 to 0.9, with 30 teacher forcing steps per sequence
    Chosen by hand; affects closed-loop training stability and therefore the comparison between models.
assumptions (4)
  • domain assumption Universal causes (e.g. physics) apply homogeneously throughout space and time, so a single set of prediction kernel weights can model all grid cells.
    Invoked in the Introduction and Section 2: DISTANA assumes the sensed dynamics are generated by universal causal principles, and all prediction kernels share identical weights.
  • domain assumption Hidden causal factors are constant over time, locally distinct, and only indirectly observable.
    Section 2.2 states that AT assumes the latent context s is constant and is optimized retrospectively; the Final Discussion summarizes the factors as locally distinct, constant, and universally present.
  • domain assumption Prediction-error gradients are sufficient to recover the true hidden factor rather than an arbitrary latent variable that also reduces prediction error.
    Section 2.2 describes AT as optimizing s to reduce local prediction errors; the authors acknowledge in Section 4 that the robustness of this causal tendency needs future investigation.
  • domain assumption In the weather LSM experiment, the model trained on two-thirds of the land-sea mask learns a generalizable encoding that transfers to the unseen America and Atlantic region.
    Section 3.4: only two thirds of the LSM values were provided during training; inference on the missing third assumes the learned LSM influence on T850 generalizes spatially.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent State Inference in a Spatiotemporal Generative Model." pith.science (2026). https://pith.science/paper/AFY6LTW3

@misc{pith2026200909823,
  author       = {Pith},
  title        = {Pith review of: Latent State Inference in a Spatiotemporal Generative Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AFY6LTW3}},
  note         = {Machine review of arXiv:2009.09823}
}
read the original abstract

Knowledge about the hidden factors that determine particular system dynamics is crucial for both explaining them and pursuing goal-directed interventions. Inferring these factors from time series data without supervision remains an open challenge. Here, we focus on spatiotemporal processes, including wave propagation and weather dynamics, for which we assume that universal causes (e.g. physics) apply throughout space and time. A recently introduced DIstributed SpatioTemporal graph Artificial Neural network Architecture (DISTANA) is used and enhanced to learn such processes, requiring fewer parameters and achieving significantly more accurate predictions compared to temporal convolutional neural networks and other related approaches. We show that DISTANA, when combined with a retrospective latent state inference principle called active tuning, can reliably derive location-respective hidden causal factors. In a current weather prediction benchmark, DISTANA infers our planet's land-sea mask solely by observing temperature dynamics and, meanwhile, uses the self inferred information to improve its own future temperature predictions.

Figures

Figures reproduced from arXiv: 2009.09823 by the authors.

Figure 1
Figure 1. 3 × 3 sensor mesh grid showing the connection scheme of Prediction Kernels (PKs) that model the local dynamical process while communicating laterally. Figure modified from [14]. 2 DISTANA As introduced in [13] and following the naming convention of [27], DISTANA (the DIstributed SpatioTemporal graph Artificial Neural network Architecture) can be described as a spatiotemporal graph neural network (ST-GNN). While GNNs… view at source ↗
Figure 2
Figure 2. Left: two-dimensional wave propagating through a 16 × 16 grid with obstacles. Darker dots in the grid nodes correspond to stronger blocking effect on the wave. Right: wave activity for two exemplary positions in the grid, with fast and slow propagation speeds. 2 952 free parameters to project the 16 × 16 × 1 input (ignoring batch and time dimensions) via the first layer on eight feature maps (resulting in dimensiona… view at source ↗
Figure 3
Figure 3. Top left: ground truth and model outputs at time step 80, which is 30 time steps after the start of closed loop prediction (from left to right: ground truth, ConvLSTM, TCN, DISTANA, DISTANA + AT). Bottom left: ground truth and model outputs over time at position x = 2, y = 6 in the two-dimensional grid. Top right: inferred static context during testing with values in the range [−0.6, 2.6] after 1, 2, 10, 500 iterati… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Predicted temperature (T850) in degree Kelvin for 24, 48 and 72 hours (corre￾sponding to time steps) into the future (closed-loop). The first row shows the ground truth and the second row the network output. However, seeing that the best numerical operational weather p…
Figure 5
Figure 5. Figure 5: Top left: original land-sea mask (LSM). Top center: global LSM inferred during testing after being trained on two thirds of the globe (the model has never seen America’s LSM). Top right: a latent vector which developed during training and encodes LSM information as wel…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 26 canonical work pages

  1. [1]

    An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling

    Bai, S., Kolter, J.Z., Koltun, V.: An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv:1803.01271 (2018)

    work page Pith review arXiv 2018

  2. [2]

    In: Advances in neural information processing systems

    Battaglia, P., Pascanu, R., Lai, M., Rezende, D.J., et al.: Interaction networks for learning about objects, relations and physics. In: Advances in neural information processing systems. pp. 4502–4510 (2016)

    work page 2016

  3. [3]

    Relational inductive biases, deep learning, and graph networks

    Battaglia, P.W., Hamrick, J.B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Ma- linowski, M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., et al.: Relational inductive biases, deep learning, and graph networks. arXiv:1806.01261 (2018)

    work page Pith review arXiv 2018

  4. [4]

    Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks

    Bengio, S., Vinyals, O., Jaitly, N., Shazeer, N.: Scheduled sampling for sequence prediction with recurrent neural networks. arXiv:1506.03099 (2015)

    work page Pith review arXiv 2015

  5. [5]

    Nature 566(7744), 373–377 (2019)

    Boers, N., Goswami, B., Rheinwalt, A., Bookhagen, B., Hoskins, B., Kurths, J.: Complex networks reveal global pattern of extreme-rainfall teleconnections. Nature 566(7744), 373–377 (2019)

    work page 2019

  6. [6]

    IEEE Signal Processing Magazine 34(4), 18–42 (2017)

    Bronstein, M.M., Bruna, J., LeCun, Y., Szlam, A., Vandergheynst, P.: Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine 34(4), 18–42 (2017)

    2017

  7. [7]

    Neural Networks 117, 135–144 (2019)

    Butz, M.V., Bilkey, D., Humaidan, D., Knott, A., Otte, S.: Learning, planning, and control in a monolithic neural event inference architecture. Neural Networks 117, 135–144 (2019)

    work page 2019

  8. [8]

    Artificial Neural Networks and Machine Learning – ICANN 2019 11727, 639–653 (2019)

    Butz, M.V., Menge, T., Humaidan, D., Otte, S.: Inferring event-predictive goal- directed object manipulations in reprise. Artificial Neural Networks and Machine Learning – ICANN 2019 11727, 639–653 (2019)

    work page 2019

Show all 28 references
  1. [9]

    In: Proceedings of the 34th International Conference on Machine Learning-Volume 70

    Dauphin, Y.N., Fan, A., Auli, M., Grangier, D.: Language modeling with gated convolutional networks. In: Proceedings of the 34th International Conference on Machine Learning-Volume 70. pp. 933–941. JMLR. org (2017)

    2017

  2. [10]

    Neural computation 9(8), 1735–1780 (1997)

    Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735–1780 (1997)

    1997

  3. [11]

    In: Bach, F., Blei, D

    Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Bach, F., Blei, D. (eds.) Proceedings of the 32nd International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 37, pp. 448–4...

    2015

  4. [12]

    arXiv:1610.10099 (2016)

    Kalchbrenner, N., Espeholt, L., Simonyan, K., Oord, A.v.d., Graves, A., Kavukcuoglu, K.: Neural machine translation in linear time. arXiv:1610.10099 (2016)

    2016 arXiv

  5. [13]

    arXiv:1912.11141 (12 2019) 12 Karlbauer et al

    Karlbauer, M., Otte, S., Lensch, H.P.A., Scholten, T., Wulfmeyer, V., Butz, M.V.: A distributed neural network architecture for robust non-linear spatio-temporal prediction. arXiv:1912.11141 (12 2019) 12 Karlbauer et al

    1912 arXiv

  6. [14]

    In: International Confer- ence on Artificial Neural Networks

    Karlbauer, M., Otte, S., Lensch, H.P., Scholten, T., Wulfmeyer, V., Butz, M.V.: Inferring, predicting, and denoising causal wave dynamics. In: International Confer- ence on Artificial Neural Networks. pp. 566–577. Springer (2020)

    2020

  7. [15]

    International Conference on Learning Representations (12 2014)

    Kingma, D., Ba, J.: Adam: A method for stochastic optimization. International Conference on Learning Representations (12 2014)

    2014

  8. [16]

    arXiv:1802.04687 (2018)

    Kipf, T., Fetaya, E., Wang, K.C., Welling, M., Zemel, R.: Neural relational inference for interacting systems. arXiv:1802.04687 (2018)

    2018 arXiv

  9. [17]

    arXiv:2010.03958 (2020)

    Otte, S., Karlbauer, M., Butz, M.V.: Active tuning. arXiv:2010.03958 (2020)

    2020 arXiv

  10. [18]

    In: Dy, J., Krause, A

    Rabinowitz, N., Perbet, F., Song, F., Zhang, C., Eslami, S.M.A., Botvinick, M.: Machine theory of mind. In: Dy, J., Krause, A. (eds.) Proceedings of the 35th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 80, pp. 4218–4227. PMLR, S...

    2018

  11. [19]

    arXiv:2002.00469 (2020)

    Rasp, S., Dueben, P.D., Scher, S., Weyn, J.A., Mouatadid, S., Thuerey, N.: Weather- bench: A benchmark dataset for data-driven weather forecasting. arXiv:2002.00469 (2020)

    2020 arXiv

  12. [20]

    In: Advances in Neural Information Processing Systems

    Rodriguez, R.C., Alaniz, S., Akata, Z.: Modeling conceptual understanding in image reference games. In: Advances in Neural Information Processing Systems. pp. 13155–13165 (2019)

    2019

  13. [21]

    arXiv:1806.01242 (2018)

    Sanchez-Gonzalez, A., Heess, N., Springenberg, J.T., Merel, J., Riedmiller, M., Hadsell, R., Battaglia, P.: Graph networks as learnable physics engines for inference and control. arXiv:1806.01242 (2018)

    2018 arXiv

  14. [22]

    In: Advances in neural information processing systems

    Santoro, A., Raposo, D., Barrett, D.G., Malinowski, M., Pascanu, R., Battaglia, P., Lillicrap, T.: A simple neural network module for relational reasoning. In: Advances in neural information processing systems. pp. 4967–4976 (2017)

    2017

  15. [23]

    Adaptive Behavior 19, 295–316 (2011)

    Sugita, Y., Tani, J., Butz, M.V.: Simultaneously emerging braitenberg codes and compositionality. Adaptive Behavior 19, 295–316 (2011)

    2011

  16. [24]

    Neural Networks 17, 1273–1289 (2004)

    Tani, J., Ito, M., Sugita, Y.: Self-organization of distributedly represented multiple behavior schemata in a mirror system: Reviews of robot experiments using rnnpb. Neural Networks 17, 1273–1289 (2004)

    2004

  17. [25]

    arXiv:1802.10353 (2018)

    Van Steenkiste, S., Chang, M., Greff, K., Schmidhuber, J.: Relational neural expectation maximization: Unsupervised discovery of objects and their interactions. arXiv:1802.10353 (2018)

    2018 arXiv

  18. [26]

    arXiv:2003.11927 (2020)

    Weyn, J.A., Durran, D.R., Caruana, R.: Improving data-driven global weather prediction using deep convolutional neural networks on a cubed sphere. arXiv:2003.11927 (2020)

    2020 arXiv

  19. [27]

    arXiv:1901.00596 (2019)

    Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., Yu, P.S.: A comprehensive survey on graph neural networks. arXiv:1901.00596 (2019)

    2019 arXiv

  20. [28]

    In: Advances in neural information processing systems

    Xingjian, S., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K., Woo, W.c.: Convolu- tional LSTM network: A machine learning approach for precipitation nowcasting. In: Advances in neural information processing systems. pp. 802–810 (2015)

    2015

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.