pith. sign in

arxiv: 2511.20577 · v4 · pith:IZTV6GYHnew · submitted 2025-11-25 · 💻 cs.LG

MSTN: A Lightweight and Fast Model for General TimeSeries Analysis

Sumit S Shevtekar , Chandresh K Maurya This is my paper

Pith reviewed 2026-05-21 18:13 UTC · model grok-4.3

classification 💻 cs.LG
keywords time series analysismulti-scale modelingforecastingimputationclassificationlightweight neural networksearly temporal aggregationself-gated fusion
0
0 comments X

The pith

MSTN uses early temporal aggregation with multi-scale convolution, sequence modeling, and self-gated fusion to reach state-of-the-art results on time series tasks while staying lightweight and fast.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents the Multi-scale Temporal Network as a way to handle real-world time series that show non-stationarity, nonlinear dynamics, and patterns at many different speeds. It builds the model around an Early Temporal Aggregation principle that first combines information from multiple time scales before feeding it into later stages. A multi-scale convolutional encoder picks up fine local details, a recurrent or attention module tracks longer dependencies, and a self-gated fusion step with squeeze-excitation reweights the combined features on the fly. This design avoids the rigid fixed-scale choices common in other architectures and keeps the total parameter count low. Readers would care because the approach delivers top performance on imputation, long-term forecasting, and classification while running quickly enough for practical use.

Core claim

MSTN is a hybrid neural architecture grounded in the Early Temporal Aggregation principle. It integrates three components: a multi-scale convolutional encoder that captures fine-grained local structure, a sequence modeling module that learns long-range dependencies through recurrent or attention-based mechanisms, and a self-gated fusion stage that uses squeeze-excitation and a single dense layer to dynamically reweight and fuse multi-scale representations. This enables MSTN to flexibly model temporal patterns spanning milliseconds to extended horizons without the computational cost of long-context models.

What carries the argument

Early Temporal Aggregation principle, which combines multi-scale convolutional encoding, sequence modeling, and self-gated fusion to capture and dynamically balance features across temporal scales before full sequence processing.

Load-bearing premise

The design assumes that the Early Temporal Aggregation principle with its specific multi-scale convolution, sequence modeling, and self-gated fusion will produce generalizable improvements without needing extensive dataset-specific tuning.

What would settle it

A controlled ablation experiment on the same 27 datasets that removes either the multi-scale convolutional branch or the self-gated fusion and measures whether performance drops, stays the same, or improves.

Figures

Figures reproduced from arXiv: 2511.20577 by Chandresh K Maurya, Sumit S Shevtekar.

Figure 1
Figure 1. Figure 1: Proposed MSTN: (a) architectural diagram and (b) signal processing pipeline [PITH_FULL_IMAGE:figures/full_fig_p017_1.png] view at source ↗
read the original abstract

Real-world time series often exhibit strong non-stationarity, complex nonlinear dynamics, and behavior expressed across multiple temporal scales, from rapid local fluctuations to slow-evolving long-range trends. However, many contemporary architectures impose rigid, fixed-scale structural priors-such as patch-based tokenization, predefined receptive fields, or frozen backbone encoders-which can over-regularize temporal dynamics and limit adaptability to abrupt high-magnitude events. To handle this, we introduce the Multi-scale Temporal Network (MSTN), a hybrid neural architecture grounded in an Early Temporal Aggregation principle. MSTN integrates three complementary components: (i) a multi-scale convolutional encoder that captures fine-grained local structure; (ii) a sequence modeling module that learns long-range dependencies through either recurrent or attention-based mechanisms; and (iii) a self-gated fusion stage incorporating squeeze-excitation and a single dense layer to dynamically reweight and fuse multi-scale representations. This design enables MSTN to flexibly model temporal patterns spanning milliseconds to extended horizons, while avoiding the computational burden typically associated with long-context models. Across extensive benchmarks covering imputation, long-term forecasting, classification, and cross-dataset generalization, MSTN achieves state-of-the-art performance, establishing new best results on 21 of 27 datasets, while remaining lightweight (~0.40M params for MSTN-BiLSTM and ~1.06M for MSTN-Transformer) and suitable for low-latency inference (<1 sec, often in milliseconds), resource-constrained deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces the Multi-scale Temporal Network (MSTN), a hybrid architecture grounded in an Early Temporal Aggregation principle. It combines a multi-scale convolutional encoder for local structure, a sequence modeling module (BiLSTM or Transformer) for long-range dependencies, and a self-gated fusion stage with squeeze-excitation to dynamically reweight representations. The model is positioned as lightweight and fast for general time series tasks, with empirical claims of state-of-the-art results on 21 of 27 datasets spanning imputation, long-term forecasting, classification, and cross-dataset generalization.

Significance. If the performance claims hold under rigorous verification, MSTN would provide a practical, resource-efficient alternative for modeling non-stationary multi-scale time series without the overhead of long-context models. The hybrid design and emphasis on low parameter counts (~0.4M–1M) and sub-second inference address real deployment constraints in the field.

major comments (2)
  1. [§5] §5 (Experimental Results): The central claim of new best results on 21 of 27 datasets is not accompanied by an explicit list of baseline methods, number of random seeds, error bars, or statistical significance tests. Without these, it is impossible to determine whether reported gains are robust or sensitive to post-hoc choices.
  2. [§4.3] §4.3 (Self-Gated Fusion): The fusion mechanism is described at a high level but lacks the precise formulation of the squeeze-excitation operation and the single dense layer (e.g., input/output dimensions, activation, or initialization). This detail is load-bearing for reproducibility of the multi-scale reweighting.
minor comments (2)
  1. [§3] The abstract and §3 refer to 'Early Temporal Aggregation' without a concise formal statement or pseudocode; a short boxed definition would improve clarity.
  2. Table captions in the results section should explicitly state the metric (e.g., MAE, accuracy) and whether lower or higher is better for each task.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback on our work. We have prepared point-by-point responses to the major comments and will incorporate revisions to address the concerns raised regarding experimental reporting and technical details for reproducibility.

read point-by-point responses

  1. Referee: [§5] §5 (Experimental Results): The central claim of new best results on 21 of 27 datasets is not accompanied by an explicit list of baseline methods, number of random seeds, error bars, or statistical significance tests. Without these, it is impossible to determine whether reported gains are robust or sensitive to post-hoc choices.

    Authors: We thank the referee for this important comment on the presentation of experimental results. While the manuscript includes a list of baseline methods in the tables and text of §5, we agree that additional details on random seeds, error bars, and statistical tests would improve the assessment of robustness. In the revised version, we will explicitly report the number of random seeds, include error bars in the result tables, and add statistical significance tests to support the performance claims. These changes will be made without altering the reported results. revision: yes

  2. Referee: [§4.3] §4.3 (Self-Gated Fusion): The fusion mechanism is described at a high level but lacks the precise formulation of the squeeze-excitation operation and the single dense layer (e.g., input/output dimensions, activation, or initialization). This detail is load-bearing for reproducibility of the multi-scale reweighting.

    Authors: We appreciate the referee's suggestion for greater precision in describing the self-gated fusion mechanism. We agree that the current high-level description in §4.3 could be enhanced with exact formulations to aid reproducibility. We will revise the manuscript to provide the precise mathematical details of the squeeze-excitation operation and the dense layer, including dimensions, activations, and initialization. This will be added to §4.3. revision: yes

Circularity Check

0 steps flagged

No significant circularity in architecture design or empirical claims

full rationale

The paper introduces MSTN as a new hybrid architecture motivated by handling non-stationarity and multi-scale temporal dynamics via Early Temporal Aggregation, multi-scale convolution, sequence modeling, and self-gated fusion. These are presented as design choices without any equations, predictions, or first-principles derivations that reduce to fitted inputs or self-definitions by construction. Performance claims rest on reported benchmark results across imputation, forecasting, classification, and generalization tasks rather than on self-citation chains, uniqueness theorems from prior author work, or renaming of known patterns. No load-bearing self-referential steps appear in the abstract or described components; the contribution is self-contained as an empirical model proposal evaluated on external datasets.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim depends on standard neural network training assumptions and the unproven effectiveness of the Early Temporal Aggregation principle for general time series; no new physical entities or mathematical axioms beyond domain conventions are introduced.

free parameters (2)
  • number of scales and hidden dimensions
    Architecture hyperparameters such as the number of convolutional scales and channel sizes are chosen and fitted during model development and training.
  • fusion gate parameters
    Weights in the squeeze-excitation and dense fusion layer are learned from data.
axioms (1)
  • domain assumption Early Temporal Aggregation principle enables flexible modeling of multi-scale dynamics without over-regularization
    The abstract states the architecture is grounded in this principle to handle non-stationarity and multiple temporal scales.

pith-pipeline@v0.9.0 · 5794 in / 1622 out tokens · 66275 ms · 2026-05-21T18:13:05.724028+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages · 9 internal anchors

  1. [1]

    M. A. Morid, O. R. L. Sheng, J. Dunbar, Time series prediction using deep learning methods in healthcare 14 (1) (Jan. 2023).doi:10.1145/3531326. URLhttps://doi.org/10.1145/3531326

    work page doi:10.1145/3531326 2023

  2. [2]

    Kadiyala, A

    A. Kadiyala, A. Kumar, Multivariate time series models for prediction of air quality inside a public transportation bus using available software, En- vironmental Progress & Sustainable Energy 33 (2) (2014) 337–341

    work page 2014

  3. [3]

    Gruca, F

    A. Gruca, F. Serva, L. Lliso, P. Rípodas, X. Calbet, P. Herruzo, J. Pihrt, R. Raevskyi, P. Šimánek, M. Choma, et al., Weather4cast at neurips 2022: Super-resolution rain movie prediction under spatio-temporal shifts, in: NeurIPS 2022 Competition Track, PMLR, 2022, pp. 292–313

    work page 2022

  4. [4]

    E. G. Kardakos, M. C. Alexiadis, S. I. Vagropoulos, C. K. Simoglou, P. N. Biskas, A. G. Bakirtzis, Application of time series and artificial neural network models in short-term forecasting of pv power generation, in: 2013 48th International Universities’ Power Engineering Conference (UPEC), 2013, pp. 1–6.doi:10.1109/UPEC.2013.6714975

    work page doi:10.1109/upec.2013.6714975 2013

  5. [5]

    H. Wu, J. Xu, J. Wang, M. Long, Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting (2022).arXiv:2106. 13008. URLhttps://arxiv.org/abs/2106.13008

    work page arXiv 2022

  6. [6]

    S. Zhao, M. Jin, Z. Hou, C. Yang, Z. Li, Q. Wen, Y. Wang, Himtm: Hi- erarchical multi-scale masked time series modeling with self-distillation for long-term forecasting (2024).arXiv:2401.05012. URLhttps://arxiv.org/abs/2401.05012

    work page arXiv 2024

  7. [7]

    H. Wu, T. Hu, Y. Liu, H. Zhou, J. Wang, M. Long, Timesnet: Temporal 2d-variation modeling for general time series analysis (2023).arXiv:2210. 02186. URLhttps://arxiv.org/abs/2210.02186

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  8. [8]

    B. Lim, S. Zohren, Time-series forecasting with deep learning: a sur- vey, Philosophical Transactions of the Royal Society A 379 (2194) (2021) 20200209.doi:10.1098/rsta.2020.0209. 26

    work page doi:10.1098/rsta.2020.0209 2021

  9. [9]

    Y. Nie, N. H. Nguyen, P. Sinthong, J. Kalagnanam, A time series is worth 64 words: Long-term forecasting with transformers (2023).arXiv:2211. 14730. URLhttps://arxiv.org/abs/2211.14730

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  10. [10]

    A. Zeng, M. Chen, L. Zhang, Q. Xu, Are transformers effective for time series forecasting? (2022).arXiv:2205.13504. URLhttps://arxiv.org/abs/2205.13504

    work page arXiv 2022

  11. [11]

    Franceschi, A

    J.-Y. Franceschi, A. Dieuleveut, M. Jaggi, Unsupervised scalable represen- tation learning for multivariate time series, in: Advances in Neural Infor- mation Processing Systems (NeurIPS), 2019, pp. 4652–4663

    work page 2019

  12. [12]

    S. Bai, J. Z. Kolter, V. Koltun, An empirical evaluation of generic convolu- tionalandrecurrentnetworksforsequencemodeling, CoRRabs/1803.01271 (2018). URLhttp://arxiv.org/abs/1803.01271

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  13. [13]

    Long short -term memory,

    S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Compu- tation 9 (8) (1997) 1735–1780.doi:10.1162/neco.1997.9.8.1735

    work page doi:10.1162/neco.1997.9.8.1735 1997

  14. [14]

    Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks

    G. Lai, W.-C. Chang, Y. Yang, H. Liu, Modeling long- and short-term temporal patterns with deep neural networks (2018).arXiv:1703.07015. URLhttps://arxiv.org/abs/1703.07015

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  15. [15]

    Y. He, J. Zhao, Temporal convolutional networks for anomaly detection in time series, Journal of Physics: Conference Series 1213 (4) (2019) 042050. doi:10.1088/1742-6596/1213/4/042050

    work page doi:10.1088/1742-6596/1213/4/042050 2019

  16. [16]

    Attention Is All You Need

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need (2023).arXiv:1706. 03762. URLhttps://arxiv.org/abs/1706.03762

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  17. [17]

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, W. Zhang, Informer: Beyond efficient transformer for long sequence time-series forecasting, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35, 2021, pp. 11106–11115

    work page 2021

  18. [18]

    T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, R. Jin, Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting (2022). arXiv:2201.12740. URLhttps://arxiv.org/abs/2201.12740

    work page arXiv 2022

  19. [19]

    X. Qiu, H. Cheng, X. Wu, J. Hu, C. Guo, B. Yang, A comprehensive survey of deep learning for multivariate time series forecasting: A channel strategy perspective (2025).arXiv:2502.10721. URLhttps://arxiv.org/abs/2502.10721 27

    work page arXiv 2025

  20. [20]

    Chang, W.-Y

    C. Chang, W.-Y. Wang, W.-C. Peng, T.-F. Chen, Llm4ts: Aligning pre- trained llms as data-efficient time-series forecasters, ACM Trans. Intell. Syst. Technol. 16 (3) (Apr. 2025).doi:10.1145/3719207. URLhttps://doi.org/10.1145/3719207

    work page doi:10.1145/3719207 2025

  21. [21]

    M. Jin, S. Wang, L. Ma, Z. Chu, J. Y. Zhang, X. Shi, P.-Y. Chen, Y. Liang, Y.-F. Li, S. Pan, Q. Wen, Time-llm: Time series forecasting by reprogram- ming large language models (2024).arXiv:2310.01728. URLhttps://arxiv.org/abs/2310.01728

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  22. [22]

    Zhang, L

    Y. Zhang, L. Ma, S. Pal, Y. Zhang, M. Coates, Multi-resolution time-series transformer for long-term forecasting (2024).arXiv:2311.04147. URLhttps://arxiv.org/abs/2311.04147

    work page arXiv 2024

  23. [23]

    Han, X.-Y

    L. Han, X.-Y. Chen, H.-J. Ye, D.-C. Zhan, Softs: Efficient multivariate time series forecasting with series-core fusion (2024).arXiv:2404.14197. URLhttps://arxiv.org/abs/2404.14197

    work page arXiv 2024

  24. [24]

    W. Han, T. Zhu, L. Chen, H. Ning, Y. Luo, Y. Wan, Mcformer: Multivari- ate time series forecasting with mixed-channels transformer, IEEE Internet of Things Journal 11 (17) (2024) 28320–28329.doi:10.1109/JIOT.2024. 3401697

    work page doi:10.1109/jiot.2024 2024

  25. [25]

    Alharthi, K

    M. Alharthi, K. Mahmood, S. Patel, A. Mahmood, Emtsf:extraordinary mixture of sota models for time series forecasting (2025).arXiv:2510. 23396. URLhttps://arxiv.org/abs/2510.23396

    work page arXiv 2025

  26. [26]

    T. Zhou, P. Niu, X. Wang, L. Sun, R. Jin, One fits all:power general time series analysis by pretrained lm (2023).arXiv:2302.11939. URLhttps://arxiv.org/abs/2302.11939

    work page arXiv 2023

  27. [27]

    Y. Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, M. Long, itransformer: Invertedtransformersareeffectivefortimeseriesforecasting(2024).arXiv: 2310.06625. URLhttps://arxiv.org/abs/2310.06625

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  28. [28]

    Less Is More: Fast Multivariate Time Series Forecasting with Light Sampling-oriented MLP Structures,

    T. Zhang, Y. Zhang, W. Cao, J. Bian, X. Yi, S. Zheng, J. Li, Less is more: Fast multivariate time series forecasting with light sampling-oriented mlp structures (2022).arXiv:2207.01186. URLhttps://arxiv.org/abs/2207.01186

    work page arXiv 2022

  29. [29]

    Rodegast, et al., Motorcycle collision dataset (2024).doi: 10.18419/darus-3301

    M. Rodegast, et al., Motorcycle collision dataset (2024).doi: 10.18419/darus-3301. URLhttps://darus.uni-stuttgart.de/dataset.xhtml? persistentId=doi:10.18419/darus-3301

    work page doi:10.18419/darus-3301 2024

  30. [30]

    Trindade, ElectricityLoadDiagrams20112014, UCI Machine Learning Repository, DOI: https://doi.org/10.24432/C58C86 (2015)

    A. Trindade, ElectricityLoadDiagrams20112014, UCI Machine Learning Repository, DOI: https://doi.org/10.24432/C58C86 (2015). 28

    work page doi:10.24432/c58c86 2015

  31. [31]

    Köllé, Wetterstation

    O. Köllé, Wetterstation. weather., Technical report and dataset, Max- Planck-Institut für Biogeochemie (BGC Jena), Germany, data freely avail- able athttps://www.bgc-jena.mpg.de/wetter/(2025). URLhttps://www.bgc-jena.mpg.de/wetter/

    work page 2025

  32. [32]

    The UEA multivariate time series classification archive, 2018

    A. Bagnall, H. A. Dau, J. Lines, M. Flynn, J. Large, A. Bostrom, P. Southam, E. Keogh, The uea multivariate time series classification archive, 2018 (2018).arXiv:1811.00075. URLhttps://arxiv.org/abs/1811.00075

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  33. [33]

    dataset on powered two wheelers fall and critical events detection

    A. Boubezoul, F. Dufour, S. Bouaziz, S. Espié, Corrigendum to “dataset on powered two wheelers fall and critical events detection”, Data in Brief 30 (2020) 105577.doi:https://doi.org/10.1016/j.dib.2020.105577. URLhttps://www.sciencedirect.com/science/article/pii/ S2352340920304716

    work page doi:10.1016/j.dib.2020.105577 2020

  34. [34]

    Anguita, Alessandro Ghio, L

    J. Reyes-Ortiz, D. Anguita, A. Ghio, L. Oneto, X. Parra, Human Activity Recognition Using Smartphones, UCI Machine Learning Repository, DOI: https://doi.org/10.24432/C54S4K (2013)

    work page doi:10.24432/c54s4k 2013

  35. [35]

    Reiss, PAMAP2 Physical Activity Monitoring, UCI Machine Learning Repository, DOI: https://doi.org/10.24432/C5NW2H (2012)

    A. Reiss, PAMAP2 Physical Activity Monitoring, UCI Machine Learning Repository, DOI: https://doi.org/10.24432/C5NW2H (2012)

    work page doi:10.24432/c5nw2h 2012

  36. [36]

    O. I. Dissanayake, S. E. McPherson, J. Allyndrée, E. Kennedy, P. Cunning- ham, L. Riaboff, Actbecalf: Accelerometer-based multivariate time-series dataset for calf behavior classification, Data in Brief 60 (2025) 111462. doi:https://doi.org/10.1016/j.dib.2025.111462. URLhttps://www.sciencedirect.com/science/article/pii/ S2352340925001945

    work page doi:10.1016/j.dib.2025.111462 2025

  37. [37]

    Davari, B

    N. Davari, B. Veloso, R. Ribeiro, J. Gama, MetroPT-3 Dataset, UCI Machine Learning Repository, dOI:https://doi.org/10.24432/C5VW3R (2021)

    work page doi:10.24432/c5vw3r 2021

  38. [38]

    Saxena, K

    A. Saxena, K. Goebel, Nasa turbofan engine degradation simulation data set, nASA Ames Prognostics Center of Excellence (2008). URLhttps://www.nasa.gov/intelligent-systems-division/ discovery-and-systems-health/pcoe/pcoe-data-set-repository/

    work page 2008

  39. [39]

    Rodegast, S

    P. Rodegast, S. Maier, J. Kneifl, J. Fehr, On using machine learning algo- rithms for motorcycle collision detection, Discover Applied Sciences 6 (6) (2024) 326

    work page 2024

  40. [40]

    F. Elwy, R. Aburukba, A. R. Al-Ali, A. A. Nabulsi, A. Tarek, A. Ayub, M. Elsayeh, Data-driven safe deliveries: The synergy of iot and machine learning in shared mobility, Future Internet 15 (10) (2023)

    work page 2023

  41. [41]

    D. P. Ismi, S. Panchoo, M. Murinto, K-means clustering based filter feature selection on high dimensional data, International Journal of Advances in 29 Intelligent Informatics 2 (2016) 38–45. URLhttps://api.semanticscholar.org/CorpusID:43897444

    work page 2016

  42. [42]

    Reiss, D

    A. Reiss, D. Stricker, Introducing a new benchmarked dataset for activity monitoring, in: 2012 16th International Symposium on Wearable Comput- ers, 2012, pp. 108–109.doi:10.1109/ISWC.2012.13

    work page doi:10.1109/iswc.2012.13 2012

  43. [43]

    Abhyuday Desai, Cynthia Freeman, Zuhui Wang, and Ian Beaver

    N. Davari, B. Veloso, R. P. Ribeiro, P. M. Pereira, J. Gama, Predictive maintenance based on anomaly detection using deep learning for air pro- duction unit in the railway industry, in: 2021 IEEE 8th International Con- ference on Data Science and Advanced Analytics (DSAA), 2021, pp. 1–10. doi:10.1109/DSAA53316.2021.9564181

    work page doi:10.1109/dsaa53316.2021.9564181 2021

  44. [44]

    G. Woo, C. Liu, D. Sahoo, A. Kumar, S. Hoi, Etsformer: Exponential smoothing transformers for time-series forecasting (2022).arXiv:2202. 01381. URLhttps://arxiv.org/abs/2202.01381

    work page arXiv 2022

  45. [45]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimization (2017). arXiv:1412.6980. 30

    work page internal anchor Pith review Pith/arXiv arXiv 2017