HierCVAE: Hierarchical Attention-Driven Conditional Variational Autoencoders for Multi-Scale Temporal Modeling

Yao Wu

REVIEW 5 major objections 5 minor 21 references

HierCVAE: Hierarchical Attention-Driven Conditional Variational Autoencoders for Multi-Scale Temporal Modeling

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A hierarchical conditional VAE claims to cut energy prediction error by 15–40% while outputting calibrated uncertainty intervals.

desk verdict Nice architecture stack, but the headline accuracy claim is contradicted by the paper's own Table 2; theoretical bounds are vacuous and calibration is unmeasured. read the letter →

arxiv 2508.18922 v1 pith:NOMOZ4M5 submitted 2025-08-26 cs.LG cs.AI

Yao Wu This is my paper · ORCID

classification cs.LGcs.AI

keywords hierarchicalattentionconditionalvariationalautoencoderprobabilistictimeseriesforecastinguncertaintyquantificationmulti-scaletemporalmodelingenergyconsumptionResFormerlatentspace

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 proposes HierCVAE, a generative architecture for temporal modeling that couples a three-tier hierarchical attention mechanism (local, global, cross-temporal) with a conditional variational autoencoder and explicit uncertainty-prediction heads. The paper's aim is to show that one network can simultaneously capture short-term fluctuations, long-term dependencies, and their interactions while producing calibrated confidence intervals, rather than choosing between point accuracy and uncertainty. On three-zone energy consumption data, the paper reports 15–40% accuracy gains over strong baselines and a 96.4% Zone 1 MSE improvement over the best baseline. If true, this matters because grid management and other high-stakes forecasting settings need both accurate point predictions and trustworthy intervals for decisions.

What carries the argument

The load-bearing machinery is (1) a multi-modal condition encoder combining BiLSTM sequence features, four statistical moments, and first-difference convolution trends; (2) a hierarchical attention stack with learnable local masking, global attention, and cross-temporal state-to-history attention, fused by adaptive gating; (3) ResFormer blocks—residual multi-head self-attention layers—placed inside the CVAE latent space; and (4) an uncertainty-aware multi-task loss whose heteroskedastic Gaussian term, L_robust = (1/(2σ²))||x_{t+1} − x̂_{t+1}||² + (1/2)log σ², makes the model output both mean and variance. The gating weights let the model pick which temporal scale dominates, and the variance

What would settle it

Take the same model and data pipeline but transform the target to induce heavy tails (e.g., log-normal demand spikes) and measure ECE/PICP on held-out zones. If calibration degrades sharply under non-Gaussian noise while point accuracy stays high, the 'well-calibrated uncertainty' claim is carried by the Gaussian variance head, not by the architecture. A simpler numeric check: recompute Zone 1 MSE on standardized targets; if the 96.4% gap collapses, the headline improvement is a scaling effect.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that conditioning a CVAE on a multi-modal state description—sequential, statistical-moment, and trend features—and refining the sampled latent with residual transformer (ResFormer) layers lets a single model reconstruct, forecast, and quantify uncertainty at once. The paper claims this combination outperforms state-of-the-art baselines by 15–40% in prediction accuracy while maintaining well-calibrated uncertainty estimates. The headline evidence is a Zone 1 MSE of 0.64M versus 17.78M for the best baseline CatBoost, a 96.4% reduction, together with lower Wasserstein and Kolmogorov–Smirnov distributional scores across zones. The paper also states repr

Load-bearing premise

The predictive distribution is assumed to be Gaussian (mean from the prediction head, variance from the uncertainty head); if the true conditional distribution of energy demand is not Gaussian, the claimed 'well-calibrated uncertainty' has no foundation.

Editorial extensions

If this is right

Long-horizon forecasts can inherit probabilistic intervals through an autoregressive uncertainty-propagation rule that accumulates variance over steps.
Operational users can act on both the point forecast and its calibrated interval, which is what load-shedding and pricing decisions require.
Adaptive fusion weights mean the model can emphasize local fluctuations under volatile conditions and global trends under stable ones, without retraining.
The multi-modal condition encoder gives the model a statistical-trend read on the recent window, not just its sequential order, which should help when distributions shift.

Reading between the lines

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

Editorial extension: the paper's generalizability claim covers traffic, finance, and weather, but the experiments only exercise energy; applying the same architecture to those domains is a direct test the paper leaves open.
Editorial extension: ablating the cross-temporal attention head and the statistical-moment branch separately would reveal which component carries the reported accuracy gain; the paper does not provide such ablations.
Editorial extension: recomputing Zone 1 MSE on standardized targets would show whether the 96.4% gain is a property of the model or of the zone's scale; the paper does not report normalized errors.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper is an architecture proposal that stacks known components – three attention masks, LSTM/moment/trend conditioning, ResFormer blocks in the latent space – into a single CVAE pipeline. That combination seems to be new, and the authors describe the modules clearly. The idea of conditioning a CVAE on multi-scale statistical descriptors is reasonable. If the experiments were trustworthy, this would be a useful engineering contribution.

They aren't. The headline claim – 'consistently outperforms SOTA by 15–40%' – is not supported by Table 2. Computing relative MSE improvements from best baselines gives 96%, 64%, 77% across zones, not 15–40%. In Zone 3, MAPE and SMAPE are worse than FlowCVAE (33.65% vs 11.18%, 25.58% vs 11.51%). So the central claim is internally contradicted by the paper's own numbers. The abstract and conclusion also say 15–40% while Section 4.2 says 96.4% for Zone 1, which is another inconsistency. This isn't a minor sloppiness; it's the load-bearing empirical claim.

The theoretical 'guarantees' in Section 3.7 are vacuous. Equation 3.7.1 uses constants C1, C2, and epsilon_approx that are never defined, so the inequality holds trivially. The prediction bound restates aleatoric vs epistemic decomposition. The convergence result is standard SGD and adds nothing.

The calibration claim is also unsupported. Table 1 lists ECE and PICP as metrics, but Table 2 never reports them. The loss in 3.5.2 is a heteroskedastic Gaussian NLL, which only gives calibrated intervals if the conditional density is Gaussian – an assumption not justified for energy demand. The reader's weakest assumption is exactly this, and it lands.

To give credit: the architecture is described in enough detail to reproduce, and the authors do include several baselines (CVAE, FlowCVAE, gradient boosting, linear) rather than cherry-picking only weak ones. The distributional metrics (Wasserstein, KS, skewness) are a nice addition. But the empirical analysis is not honest to its own data.

Who is this for? Someone studying how overclaimed results manifest in ML papers, or someone looking for an example of why abstract numbers need to be checked against tables. Not for someone seeking a reliable forecasting model.

Recommendation: a serious referee could still be useful to confirm the flaws and to see whether the architecture has any merit after re-analysis, so I'd send it to review rather than desk reject. But the central claim is self-contradictory, so acceptance should be out of the question as submitted.

Referee Report

5 major / 5 minor

Summary. The paper proposes HierCVAE, a conditional variational autoencoder for temporal modeling that combines hierarchical multi-scale attention (local, global, cross-temporal), multi-modal condition encoding (LSTM, statistical moments, trend convolution), ResFormer blocks in the latent space, and multi-task output heads for reconstruction, point prediction, and heteroskedastic uncertainty. The authors claim 15–40% consistent improvements over state-of-the-art methods and well-calibrated uncertainty estimates on energy consumption data. Section 3.7 offers theoretical bounds on reconstruction error, prediction error, and convergence. Experiments are reported on three zones of an energy dataset with several baselines, but calibration metrics promised in Table 1 are absent from Table 2, and the reported results partly contradict the headline accuracy claims.

Significance. If the claims were supported, the contribution would be significant: an architecture that simultaneously captures multi-scale dependencies and provides calibrated predictive uncertainty is of clear practical interest. The proposed components are coherent and the multi-task loss design is reasonable. However, the central empirical claims are internally inconsistent with the paper's own Table 2, the uncertainty-calibration claim is not backed by any reported calibration metric, and the theoretical bounds in Section 3.7 are vacuous as stated. The paper's significance as a “new paradigm” with “theoretical guarantees” is therefore not established by the current manuscript. The architecture might merit a more modest presentation, but the advertised results cannot be accepted as reported.

major comments (5)

[Abstract and Section 4.2.1, Table 2] The headline claim of a consistent 15–40% improvement over state-of-the-art is contradicted by the paper's own numbers. Relative MSE improvements against the best baseline are 96.4% (Zone 1), 64.3% (Zone 2), and 76.9% (Zone 3), all outside the stated 15–40% interval. More seriously, in Zone 2 and Zone 3 HierCVAE is worse than FlowCVAE on MAPE and SMAPE: Zone 3 MAPE is 33.65% versus 11.18%, and SMAPE is 25.58% versus 11.51%. The text in “Zone 2 Analysis” and “Zone 3 Analysis” concedes that FlowCVAE leads in most prediction-accuracy metrics. Thus the abstract's “consistently outperforms state-of-the-art methods by 15–40% in prediction accuracy” is not supported by the reported experiments.
[Section 3.5.2 and Table 1/Table 2] Well-calibrated uncertainty is a central advertised contribution, but no calibration evidence is presented. Table 1 lists ECE and PICP as evaluation metrics, yet Table 2 contains no such columns. The uncertainty head uses a heteroskedastic Gaussian NLL (L_robust), which only provides calibrated estimates if the Gaussian predictive assumption is appropriate. No distributional diagnostics, calibration curves, PICP, or ECE values are reported, so the claim of 'well-calibrated uncertainty estimates' is unsupported.
[Section 3.7.1 and Section 3.7.2] The theoretical bounds are vacuous as stated. Equation 3.7.1 bounds reconstruction error by C1*DKL + C2*L_smooth + epsilon_approx, where C1, C2, and epsilon_approx are unspecified and no proof is given. Since epsilon_approx can absorb any finite left-hand side, the inequality imposes no constraint. Similarly, Equation 3.7.2 introduces delta_model as an undefined residual term; the inequality is tautological. The discussion claims that ResFormer blocks reduce epsilon_approx and that hierarchical attention reduces delta_model, but these are assertions without formal content. This section should be removed or replaced with real, non-vacuous statements.
[Section 4.3.1 and Section 5] The paper overclaims generalizability and deployment. Section 4.3.1 states 'Consistent improvements across diverse domains (traffic, finance, weather, energy)' although only one energy dataset is evaluated. The conclusion asserts 'Real-world deployment in smart grid systems validates our approach's practical impact,' but no deployment study or field trial is described anywhere. These statements are not supported by the manuscript's evidence.
[Section 4.1.3 and Section 4.2.1] Experimental reporting is insufficient to assess the extraordinary gains claimed. No hyperparameters, data splits, training settings, seeds, or confidence intervals are provided. For example, MAE drops from thousands (e.g., 3876.69) to 19.40 in Zone 1, a 99.5% change, but no reproducibility details are given to rule out normalization, scaling, or leakage artifacts. At minimum, standard deviations across multiple runs and a detailed experimental configuration are needed before such results can be evaluated.

minor comments (5)

[Section 3.3.1] The local attention mask M^(local) is described as learnable, but no mechanism or constraint is given. Clarify how the mask is parameterized and whether it is differentiable.
[Section 3.6.4] The unified loss says the lambda weights are 'learned during training,' but Section 4.3.2 mentions 'careful tuning of weighting parameters.' Please disambiguate: are the lambdas optimized by gradient descent or set by hand?
[Table 2] Units are inconsistent: MSE values are shown with an 'M' suffix (presumably millions), while MAE is given as a raw number. State units explicitly for all metrics. Also, R2 is written as 'R2' in the table and 'R²' in the text; unify notation.
[Section 2.1] The related work cites transformer forecasters (Informer, Autoformer, FEDformer), but none of these are included as baselines. Since the abstract claims state-of-the-art improvement, the omission of standard transformer forecasting baselines is notable; at least discuss why they are not compared.
[References] Reference formatting is inconsistent: some ICLR / NeurIPS entries lack page numbers or venue details, and the paper's title in the header is rendered as 'HierCV AE' with a space. A final copyediting pass is recommended.

Circularity Check

2 steps flagged · score 6.0 of 10

The §3.7 theoretical bounds are tautological: the prediction-error 'bound' defines δ_model as the residual, and the reconstruction 'bound' leaves C1, C2, and ε_approx unspecified so it cannot fail. The empirical Table 2 comparison is not circular, though the abstract's 15–40% claim is internally inconsistent with the table.

self definitional [Section 3.7.1, Representation Learning Bound]
"Under mild regularity conditions on the data distribution and model architecture, our enhanced CV AE satisfies the reconstruction bound: E[∥xt − ˆxt∥2 2] ≤ C1 · DKL(qϕ∥p) + C2 · Lsmooth + ϵapprox where C1, C2 are constants depending on model capacity, and ϵapprox represents the approximation error reduced by ResFormer enhancement."

The bound is not derived from any specified C1, C2, or ϵapprox. Since ϵapprox is introduced as 'the approximation error' rather than given an independent value or bound, the inequality is satisfied by defining ϵapprox as the residual (LHS − C1·DKL − C2·Lsmooth). No reconstruction error is excluded, so the claimed guarantee is an identity/definition rather than a theorem.
self definitional [Section 3.7.2, Temporal Prediction Error Bound]
"The prediction error is bounded by the sum of irreducible uncertainty and model approximation error: ∥xt+1 − ˆxt+1∥2 2 ≤ ∥xt+1 − E[xt+1|Ht]∥2 2 + δmodel where δmodel decreases with the expressiveness of our multi-modal conditioning and hierarchical attention mechanisms."

This is the standard excess-risk decomposition. δmodel is not independently estimated or bounded; it is exactly the difference between the left side and the conditional-variance term. Since the conditional mean minimizes squared error, the inequality holds automatically once δmodel is defined as that difference. The 'bound' is therefore stipulated by definition, and the asserted decrease of δmodel with expressiveness is an unproved claim appended to a tautology.

full rationale

The empirical benchmark comparison is not circular: Table 2 reports HierCVAE's MSE/MAE against external baselines, and the zone-wise improvements are direct arithmetic from those reported numbers. No fitted parameter is renamed as a prediction in the architecture, and no load-bearing self-citation chain appears. However, the theoretical analysis that the paper invokes as a guarantee is circular in the technical sense. §3.7.2's prediction-error 'bound' is an identity because δ_model is defined as the difference between the two sides, and §3.7.1's reconstruction 'bound' is unfalsifiable because C1, C2, and ε_approx are unspecified, allowing ε_approx to absorb any residual. These two 'bounds' cannot fail and therefore do not support the claimed theoretical guarantees. Separately, the abstract's headline '15–40% improvement' is internally inconsistent with the paper's own Table 2: relative MSE improvements over the best baselines are 96.4% (Zone 1), 64.3% (Zone 2), and 76.9% (Zone 3). Also, the claimed 'well-calibrated uncertainty estimates' are not backed by the ECE/PICP columns promised in Table 1 but omitted from Table 2. These are correctness/consistency concerns rather than construction-circularity. The paper itself admits in §4.3.2 that 'stronger theoretical guarantees for multi-scale attention mechanisms and uncertainty calibration under distribution shift remain open research questions,' which reinforces that the §3.7 bounds are not the stated guarantees. Overall score 6: two central theoretical 'predictions/bounds' reduce by construction, while the empirical claims themselves are not circular.

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

No new physical or ontological entities are introduced; the architecture uses standard attention, LSTM, conv, residual blocks, and VAE machinery. The free parameters are the unspecified theoretical constants and learned loss weights. The key axioms are the Gaussian likelihood assumption, the unverified SGD convergence assumptions, and the unstated regularity conditions that make the theoretical bounds vacuous.

free parameters (4)

KL weight beta
Appears in L_recon (Section 3.5.1) as a multiplier on the KL divergence; no value or tuning procedure is reported.
Loss balancing weights lambda_pred, lambda_robust, lambda_smooth, lambda_attn = learned
Section 3.6.4 says the lambda parameters are learned during training to balance objectives; their values, constraints, and validation selection are not specified.
C1, C2, epsilon_approx in reconstruction bound
Section 3.7.1 introduces these as constants depending on model capacity; no values or derivation are given, so the bound is vacuous.
delta_model in prediction bound
Section 3.7.2 defines prediction error as irreducible error plus delta_model; delta_model is not characterized, making the inequality an identity.

assumptions (4)

domain assumption Conditional predictive distribution is Gaussian with mean x_hat_{t+1} and variance sigma_pred^2
Used to write the robust loss L_robust in Section 3.5.2; if the true distribution is not Gaussian, the calibration claims do not follow.
domain assumption Standard SGD convergence assumptions (smoothness, Lipschitz continuity, bounded gradient noise) hold for the whole multi-objective network
Invoked in Section 3.7.3 via Bottou et al.; no verification that this specific architecture with attention, LSTM, VAE sampling, and multi-task losses satisfies them.
ad hoc to paper Unstated 'mild regularity conditions' on data distribution and model architecture for the reconstruction bound
Section 3.7.1 states the bound holds under mild regularity conditions, but the conditions are never listed.
ad hoc to paper ResFormer blocks reduce approximation error epsilon_approx
Section 3.7.1 relies on epsilon_approx being reduced by ResFormer enhancement; no theorem or experiment isolates this effect.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HierCVAE: Hierarchical Attention-Driven Conditional Variational Autoencoders for Multi-Scale Temporal Modeling." pith.science (2026). https://pith.science/paper/NOMOZ4M5

@misc{pith2026250818922,
  author       = {Pith},
  title        = {Pith review of: HierCVAE: Hierarchical Attention-Driven Conditional Variational Autoencoders for Multi-Scale Temporal Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NOMOZ4M5}},
  note         = {Machine review of arXiv:2508.18922}
}

read the original abstract

Temporal modeling in complex systems requires capturing dependencies across multiple time scales while managing inherent uncertainties. We propose HierCVAE, a novel architecture that integrates hierarchical attention mechanisms with conditional variational autoencoders to address these challenges. HierCVAE employs a three-tier attention structure (local, global, cross-temporal) combined with multi-modal condition encoding to capture temporal, statistical, and trend information. The approach incorporates ResFormer blocks in the latent space and provides explicit uncertainty quantification via prediction heads. Through evaluations on energy consumption datasets, HierCVAE demonstrates a 15-40% improvement in prediction accuracy and superior uncertainty calibration compared to state-of-the-art methods, excelling in long-term forecasting and complex multi-variate dependencies.

Figures

Figures reproduced from arXiv: 2508.18922 by the authors.

**Figure 2.** Hierarchical Attention Mechanism. The system applies three types of attention: (a) Local [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** ResFormer-Enhanced Latent Space. The initial latent representation [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 12 canonical work pages

[1]

Tactis: Transformer-attentional copulas for time series

Arjun Ashok, Nicholas Rhinehart, Feras Beainy, and Kris M Kitani. Tactis: Transformer-attentional copulas for time series. In International Conference on Machine Learning, pages 1013--1051, 2022

work page 2022
[2]

Weight uncertainty in neural networks

Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural networks. In International Conference on Machine Learning, pages 1613--1622, 2015

work page 2015
[3]

Optimization methods for large-scale machine learning

L \'e on Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning. SIAM review, 60 0 (2): 0 223--311, 2018

2018
[4]

Timevae: A variational auto-encoder for multivariate time series generation

Sneha Desai, Said Ouala, Kais Skandrani, Jorge Llorca, and Ronan Fablet. Timevae: A variational auto-encoder for multivariate time series generation. arXiv preprint arXiv:2111.08095, 2021

arXiv 2021
[5]

Gp-vae: Deep probabilistic time series imputation

Vincent Fortuin, Dmitry Baranchuk, Gunnar R \"a tsch, and Stephan Mandt. Gp-vae: Deep probabilistic time series imputation. In International Conference on Artificial Intelligence and Statistics, pages 1651--1661, 2020

work page 2020
[6]

Dropout as a bayesian approximation: Representing model uncertainty in deep learning

Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In International Conference on Machine Learning, pages 1050--1059, 2016

work page 2016
[7]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770--778, 2016

2016
[8]

Long short-term memory

Sepp Hochreiter and J \"u rgen Schmidhuber. Long short-term memory. Neural computation, 9 0 (8): 0 1735--1780, 1997

1997

Show all 21 references

[9]

Auto-encoding variational bayes

Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations, 2014

2014
[10]

Simple and scalable predictive uncertainty estimation using deep ensembles

Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Systems, volume 30, 2017

2017
[11]

A time series is worth 64 words: Long-term forecasting with transformers

Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. In International Conference on Learning Representations, 2023

2023
[12]

Estimating the mean and variance of the target probability distribution

David A Nix and Andreas S Weigend. Estimating the mean and variance of the target probability distribution. In Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94), volume 1, pages 55--60. IEEE, 1994. doi:10.1109/ICNN.1994.374138

1994
[13]

Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting

Kashif Rasul, Calvin Seward, Ingmar Schuster, and Roland Vollgraf. Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. In International Conference on Machine Learning, pages 8857--8868, 2021

2021
[14]

Variational inference with normalizing flows

Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International Conference on Machine Learning, pages 1530--1538, 2015

2015
[15]

Learning structured output representation using deep conditional generative models

Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems, pages 3483--3491, 2015

2015
[16]

Csdi: Conditional score-based diffusion models for probabilistic time series imputation

Yusuke Tashiro, Jiaming Song, Yang Song, and Stefano Ermon. Csdi: Conditional score-based diffusion models for probabilistic time series imputation. In Advances in Neural Information Processing Systems, volume 34, pages 24804--24816, 2021

2021
[17]

Attention is all you need

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998--6008, 2017

2017
[18]

Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting

Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In Advances in Neural Information Processing Systems, volume 34, pages 22419--22430, 2021

2021
[19]

Timesnet: Temporal 2d-variation modeling for general time series analysis

Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. Timesnet: Temporal 2d-variation modeling for general time series analysis. In International Conference on Learning Representations, 2023

2023
[20]

Informer: Beyond efficient transformer for long sequence time-series forecasting

Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 11106--11115, 2021

2021
[21]

Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting

Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In International Conference on Machine Learning, pages 27268--27286, 2022

2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Tactis: Transformer-attentional copulas for time series

Arjun Ashok, Nicholas Rhinehart, Feras Beainy, and Kris M Kitani. Tactis: Transformer-attentional copulas for time series. In International Conference on Machine Learning, pages 1013--1051, 2022

work page 2022

[2] [2]

Weight uncertainty in neural networks

Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural networks. In International Conference on Machine Learning, pages 1613--1622, 2015

work page 2015

[3] [3]

Optimization methods for large-scale machine learning

L \'e on Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning. SIAM review, 60 0 (2): 0 223--311, 2018

2018

[4] [4]

Timevae: A variational auto-encoder for multivariate time series generation

Sneha Desai, Said Ouala, Kais Skandrani, Jorge Llorca, and Ronan Fablet. Timevae: A variational auto-encoder for multivariate time series generation. arXiv preprint arXiv:2111.08095, 2021

arXiv 2021

[5] [5]

Gp-vae: Deep probabilistic time series imputation

Vincent Fortuin, Dmitry Baranchuk, Gunnar R \"a tsch, and Stephan Mandt. Gp-vae: Deep probabilistic time series imputation. In International Conference on Artificial Intelligence and Statistics, pages 1651--1661, 2020

work page 2020

[6] [6]

Dropout as a bayesian approximation: Representing model uncertainty in deep learning

Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In International Conference on Machine Learning, pages 1050--1059, 2016

work page 2016

[7] [7]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770--778, 2016

2016

[8] [8]

Long short-term memory

Sepp Hochreiter and J \"u rgen Schmidhuber. Long short-term memory. Neural computation, 9 0 (8): 0 1735--1780, 1997

1997