Pith. sign in

REVIEW 5 major objections 5 minor 93 references

HyBDM claims that decomposing multivariate series into global and local components, modeling each with a dedicated expert, and fusing them with a learned router, yields state-of-the-art accuracy and linear complexity on six benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 19:39 UTC pith:7FRRLQXF

load-bearing objection Plausible new hybrid for TSF, but two core components are never specified, so the reported numbers aren't reproducible from the paper. the 5 major comments →

arxiv 2607.16882 v1 pith:7FRRLQXF submitted 2026-07-18 cs.LG

HyBDM: Multi-Scale Hybrid Experts for Time Series Forecasting with Bidirectional Dependency Modeling

Wenqiang Ma , Chen Cheng , Xue Cheng , Jiarui Ye This is my paper
classification cs.LG
keywords time series forecastingmultivariatehybrid expertsMambastate space modelslocal window transformermulti-scale patchinglinear complexity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes HyBDM, a forecasting architecture built on the premise that long-range global patterns and short-range local variations should be modeled by separate experts rather than a single uniform backbone. The global expert is a bidirectional, Mamba-style state-space module augmented with convolutions and a forget gate; the local expert is a window-limited Transformer. A multi-scale patcher feeds the two experts complementary resolutions, and a Long-Short Router learns to blend their outputs. The author claims that this explicit global–local separation yields better average MSE/MAE than ten recent baselines on ETT, Weather, and Traffic datasets, while maintaining linear time complexity. If the claim holds, it indicates that routing between specialized temporal experts is a more reliable strategy than relying on any single inductive bias.

Core claim

On its own terms, the paper's central discovery is that multivariate time-series forecasting improves when the temporal structure is explicitly decomposed into a global component (trends, seasonality) and a local component (transient fluctuations), each processed by an expert tailored to that scale. The global expert—BiConv-Mamba—combines bidirectional convolutions with a selective state-space layer, a data-dependent skip connection, and a forget gate; the local expert—a Local Window Transformer—restricts attention to a sliding window, reducing quadratic cost to linear. The Multi-Scale Patcher creates complementary resolutions, and the Long-Short Router softly fuses the two pathways. The pap

What carries the argument

The load-bearing mechanism is the two-expert decomposition: a Global Patterns Expert (BiConv-Mamba) built from bidirectional convolutions, an M-SSM layer with a data-dependent skip connection, and a sigmoid-based forget gate; and a Local Variations Expert (LWT) that computes window-limited attention with O(wS) cost. These are fed by a Multi-Scale Patcher that produces coarse and fine patch resolutions, and their outputs are merged by a Long-Short Router that computes softmax-gated weights. The design's claim is that this explicit split captures both dependency scales without the quadratic cost of full attention.

Load-bearing premise

The reported results rely on two components the paper never fully specifies—the GDD-MLP cross-channel module and the data-dependent skip connection D—so if those are implemented differently than the authors intend, the architecture and numbers cannot be reproduced from the paper alone.

What would settle it

Check Table 2: the paper's text says HyBDM achieves 28/30 MSE wins against SST, but in the ETTm1 rows, three of the four horizons list identical MSE values for the two models (0.301, 0.344, 0.369). The ablation table (Table 4) repeats the same ETTm1 averages for both models. A reader can count the ties directly; if the tables are accurate, the claim of consistent wins does not hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the empirical results hold, hybrid SSM–Transformer models that explicitly specialize by temporal scale are a viable default for long-horizon multivariate forecasting.
  • The linear complexity claim implies that HyBDM can scale to long look-back windows and high channel counts (e.g., 862-channel Traffic) where quadratic attention becomes prohibitive.
  • The ablation results suggest that a forget gate that discards stale history is especially valuable for non-stationary series like Weather, pointing to a reusable component for other SSM-based forecasters.
  • The multi-scale patcher shows a way to feed complementary resolutions to different experts, which could be adopted outside this architecture.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The router's learned soft weights may act as a dataset-level regime detector; plotting them across horizons could reveal when global vs local modeling dominates, which the paper does not analyze.
  • The same global–local expert template could be applied to spatiotemporal forecasting by treating spatial scales as the split axis.
  • Since the paper does not report sensitivity to patch/stride/window hyperparameters, an immediate extension would be a small grid sweep to see how stable the accuracy gains are.
  • The framework suggests that a forget gate could improve other linear-time sequence models (e.g., pure Mamba variants) beyond this architecture.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes HyBDM, a multi-scale hybrid architecture for multivariate time series forecasting. It splits the input into global and local streams: a BiConv-Mamba global expert with bidirectional convolutions, an M-SSM layer, a forgetting gate, and a claimed GDD-MLP cross-channel module; a Local Window Transformer for short-range attention; a Multi-Scale Patcher; and a Long-Short Router that adaptively weights the two experts. Experiments on ETT, Weather, and Traffic report MSE/MAE against ten baselines, plus an ablation of the BiConv and forgetting components. The central claim is that HyBDM outperforms state-of-the-art methods in both accuracy and efficiency with O(L) complexity.

Significance. If the architecture were fully specified, the paper would present a plausible synthesis of Mamba-style state-space models and windowed attention with explicit global/local routing, which is a timely direction for long-horizon multivariate forecasting. The benchmark coverage is standard, and the tables show HyBDM often achieving the lowest MSE/MAE. However, the manuscript is not reproducible in its current form: the GDD-MLP module and the data-dependent skip connection D in Eq. (4) are named as components of the model but never defined, and a headline count of wins against Table 2 is contradicted by the table itself. The paper also claims empirical efficiency gains without reporting any wall-clock or memory measurements. These are load-bearing gaps rather than cosmetic issues: they prevent independent verification of the reported results and of the architecture's identity. No code, hyperparameter table, or seed information is provided.

major comments (5)
  1. [Method, BiConv-Mamba Module (Eqs. 2-5)] GDD-MLP is listed in the Abstract, Introduction, Figure 2, and Related Work as part of the Global Patterns Expert, but the Method section never defines it. No equations, layer layout, input/output dimensions, or integration point are given. Since Tables 2-4 are results of the full HyBDM model, the forward pass is underspecified. Please provide a complete description, or explicitly remove GDD-MLP from the architecture claims until it is specified.
  2. [Method, Eq. (4)] The M-SSM output is defined as y_t = C h_t + D x_t, with D called 'a data-dependent skip connection derived from the input,' but no formula or mechanism for computing D from x_t is provided. In standard Mamba, D is a learned parameter, not an input-dependent term. As written, Eq. (4) cannot be implemented or attributed to HyBDM. Please define D explicitly (e.g., D = f(x_t) with a stated map) or revise the description.
  3. [Experiments, Comparison with State-of-the-Art Methods (Table 2)] The sentence 'Table 2 shows 28/30 MSE wins against SST, CMamba, iTransformer, RLinear, and Crossformer' is not supported by the table as printed. For strictly lowest MSE among the six Table-2 models over the 24 horizon-dataset rows, HyBDM has 18 wins, 5 ties, and 1 loss (Traffic H=336, where SST has 0.401 vs HyBDM 0.413). If 'wins' means pairwise comparisons versus the five baselines, it is 114 wins, 5 ties, and 1 loss, not 28/30. If the comparison is over dataset-averaged MSE (6 datasets x 5 baselines), it is 29 wins and 1 tie. Please define the counting rule and correct the claim.
  4. [Experiments, Computational Complexity Analysis] The abstract and conclusion claim that experiments demonstrate 'computational efficiency' gains, but the Computational Complexity Analysis section reports no measured wall-clock time, FLOPs, or memory usage; it only restates theoretical complexity (O(L), O(wS)). Please provide quantitative efficiency measurements (e.g., runtime and peak memory on one or more benchmarks) or change the claim to 'theoretical linear complexity.'
  5. [Experiments, Ablation Study (Table 4)] The ablation removes only the bidirectional convolution and the forgetting gate. GDD-MLP, the Multi-Scale Patcher, the LWT, and the Long-Short Router are not ablated. Since the paper credits GDD-MLP as a newly introduced module, its contribution should be tested. At minimum, once GDD-MLP is defined, include it in the ablation; otherwise the attribution of the reported gains to the full architecture is unsubstantiated.
minor comments (5)
  1. [Table 2] The '1st Count' row (printed as '289 6 2 2210 1 0 0 0 0' in the text) is unreadable, and Table 3's count row is similarly unclear. Please replace these with a formatted table of first-place counts per model.
  2. [References] Several references are duplicated: Bai et al. 2018a/b, Hochreiter and Schmidhuber 1997a/b, Vaswani et al. 2017a/b, and Nie et al. 2023a/b have identical entries. Some arXiv identifiers dated 2026 (e.g., 2602.09081, 2606.26916, 2605.23345) should be verified and, if not publicly available, removed or replaced.
  3. [Method, Eq. (5)] The variables in Eq. (5) are not defined: 'x′' is called the primary branch output but is never connected to Eqs. (2)-(4), and 'gate_b2' is introduced without explanation. Please define all symbols used in the gating equation.
  4. [Method, Long-Short Router (Eq. 6)] The routing representation z_R is not defined. Please state what z_R is (e.g., a pooled summary of the input or an intermediate representation) and its dimension.
  5. [Experiment Setup] The paper states hyperparameters are 'tuned on validation sets' but reports no values for batch size, learning rate, patch length P, stride Str, window width w, or the router representation. For reproducibility, include these or provide a code link with configurations.

Circularity Check

0 steps flagged

No significant circularity: HyBDM is an empirical architecture paper whose components are externally cited and whose results are measured on public benchmarks.

full rationale

HyBDM does not claim a first-principles derivation that could collapse into its inputs. The forward pass is defined by standard patching (Eq. 1), an SSM update inherited from Mamba/C-Mamba (Eqs. 3-4), a sigmoid gating mix (Eq. 5), and a softmax router (Eqs. 6-7); the accuracies in Tables 2-4 are benchmark measurements, not quantities fitted to the same data and then relabeled as predictions. The cited building blocks (Mamba, C-Mamba, PatchTST, SST, iTransformer) are external prior work, and none of the paper's load-bearing claims is justified solely by a self-citation. The genuine weaknesses are under-specification rather than circularity: GDD-MLP appears only in Related Work ("Our GDD-MLP module follows the Global Patterns Expert and fuses global/local features via max/average pooling with sigmoid gating") and is absent from the Method, and Eq. (4)'s D is called "a data-dependent skip connection derived from the input" with no formula supplied. These gaps make the architecture hard to reproduce and attribute, but they do not make any claimed equivalence by construction; no equation in the paper reduces to another, and no fitted value is presented as a prediction. Thus the paper earns a 0 circularity score, with reproducibility risk noted separately.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

HyBDM introduces no new physical or conceptual entities; it is a neural architecture composed of known building blocks. The free parameters are mostly unstated architectural hyperparameters that must be chosen by an implementer, while the axioms are standard empirical and domain assumptions about the benchmarks, baselines, and SSM/attention components.

free parameters (4)
  • Patch length P
    Hand-chosen in the Multi-Scale Patcher; controls the resolution split between global and local experts and is never specified in the paper.
  • Stride Str
    Hand-chosen alongside patch length; also unspecified, directly affects the number of patches and the resolution ratio.
  • LWT window width w
    Controls the locality and complexity of the local expert; no value is given, so the model cannot be instantiated.
  • Router representation z_R
    The routing representation in Eq. (6) is never defined; it is a free design choice that affects the fusion weights.
axioms (4)
  • domain assumption The M-SSM layer (Eq. 3-4) is a valid linear-time sequence model as described in C-Mamba / Zeng et al. 2024.
    Relied on for the global-pattern expert's long-range modeling; the paper does not prove or derive the SSM formulation.
  • domain assumption Standard train/validation/test splits for the six benchmarks are used and all baselines are trained under identical protocol.
    Needed for the reported comparisons to be meaningful; the text asserts this but gives no code or explicit split details.
  • domain assumption The six benchmark datasets are representative and the MSE/MAE metrics averaged over steps and channels capture forecasting quality.
    Standard TSF practice; the paper provides no evidence for broader generalization.
  • standard math LWT attention within sliding windows of width w preserves enough local information and has complexity O(wS).
    Assumed from prior windowed-attention designs; the paper does not analyze the theoretical or practical trade-offs.

pith-pipeline@v1.3.0-alltime-deepseek · 11180 in / 14907 out tokens · 146409 ms · 2026-08-01T19:39:14.604701+00:00 · methodology

0 comments
read the original abstract

Time series forecasting (TSF) is vital to many applications, yet existing models often struggle to capture the heterogeneous long-range global patterns and short-range local variations in multivariate time series. While some approaches partially model these dependencies, they often do not jointly exploit temporal and feature-wise information. To address this challenge, we propose HyBDM, a multi-scale hybrid model that decomposes temporal dynamics into global patterns and local variations, which are modeled by two specialized experts. The Global Patterns Expert employs an enhanced BiConv-Mamba module that integrates bidirectional convolutions, an M-SSM layer, a forgetting mechanism, and a GDD-MLP module for cross-channel modeling. The Local Variations Expert uses a Local Window Transformer (LWT) to perform efficient locality-aware attention with reduced computational complexity. In addition, a Multi-Scale Patcher and a Long-Short Router enable multi-resolution representations and adaptive fusion of the two experts. Experiments on six benchmark datasets show that HyBDM outperforms state-of-the-art methods in both forecasting accuracy and computational efficiency, demonstrating its effectiveness in bridging global-local dependencies for multivariate TSF.

Figures

Figures reproduced from arXiv: 2607.16882 by Chen Cheng, Jiarui Ye, Wenqiang Ma, Xue Cheng.

Figure 1
Figure 1. Figure 1: Decomposition of time series into global patterns [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of HyBDM. The model combines a Multi-Scale Patcher, Global Patterns Expert (BiConv￾Mamba with forgetting mechanism and GDD-MLP), Local Variations Expert (LWT), and a Long-Short Router to capture both long-range and short-range dependencies. preserve fine-grained fluctuations and short-range dependen￾cies. Given multivariate series L = (x1, . . . , xL) ∈ R L×M, patch length P, and strid… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

93 extracted references · 14 linked inside Pith

  1. [3]

    Box, G. E. P.; Jenkins, G. M.; Reinsel, G. C.; and Ljung, G. M. 2015. Time series analysis: forecasting and control. John Wiley & Sons

    2015

  2. [5]

    Cheng, K.; Liu, Z.; Gao, M.; Song, C.; and Tang, H. 2026. PhysRAG: Enhancing Physics-Awareness in Video Generation via Retrieval-Augmented Generation. arXiv:2606.26916

    Pith/arXiv arXiv 2026

  3. [6]

    Cheng, Y.; and Zhang, L. 2025. MAE-Based Radio Map Construction for Wi-Fi Fingerprint Indoor Localization. IEEE Communications Letters, 29(9): 2008--2012

    2025

  4. [7]

    Fama, E. F. 1970. Efficient capital markets: A review of theory and empirical work, volume 25

    1970

  5. [9]

    Hamilton, J. D. 1994. Time series analysis. Princeton University Press

    1994

  6. [10]

    Held, L.; and Hofmann, M. 2005. A statistical framework for the analysis of multivariate infectious disease surveillance counts. Statistics in Medicine, 24(3): 220--237

    2005

  7. [11]

    Hochreiter, S.; and Schmidhuber, J. 1997 a . Long short-term memory. Neural Computation, 9(8): 1735--1780

    1997

  8. [12]

    Hochreiter, S.; and Schmidhuber, J. 1997 b . Long short-term memory. Neural Computation, 9(8): 1735--1780

    1997

  9. [13]

    Huang, Y.; Lv, X.; Xu, J.; Yu, Z.; Zhang, J.; Hu, R.; Feng, W.; Zou, S.; Xiao, H.; Zhou, Z.; Huang, K.; Peng, Z.; Xu, J.; Zhao, H.; Zhu, C.; Yi, R.; Huang, Y.; Wu, D.; Zhang, Y.; Cheng, K.; Song, C.; Xue, Y.; Zhang, X.; Guo, L.; Chen, Y.; Wu, B.; Yu, H.; and Xu, K. 2026. PAIWorld: A 3D-Consistent World Foundation Model for Robotic Manipulation. arXiv:2606.18375

    Pith/arXiv arXiv 2026

  10. [14]

    Katharopoulos, A.; Vyas, A.; Pappas, N.; and Fleuret, F. 2020. Transformers are RNNs: Fast autoregressive transformers with linear attention. ICML

    2020

  11. [16]

    LeCun, Y.; Bottou, L.; Bengio, Y.; and Haffner, P. 1995. Convolutional networks for images, speech, and time series. The Handbook of Brain Theory and Neural Networks, 3361(10): 1995

    1995

  12. [17]

    S.; et al

    Li, Z.; Qiu, X.; Chen, P.; Wang, Y.; Cheng, H.; Shu, Y.; Hu, J.; Guo, C.; Zhou, A.; Jensen, C. S.; et al. 2025. Tsfm-bench: A comprehensive and unified benchmark of foundation models for time series forecasting. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, 5595--5606

    2025

  13. [18]

    Lieber, M.; Shen, L.; and Wang, C. 2024. Jamba: Joint attention and Mamba models for long-range dependencies. Conference on Neural Information Processing Systems (NeurIPS)

    2024

  14. [19]

    Nie, Z.; Sun, K.; and Zhang, M. 2022. CI vs. CD: A comparative study on multivariate time series models. ICLR

    2022

  15. [23]

    Park, J.; Kim, H.; and Cho, S. 2024. Mambaformer: Hybrid state space transformer for sequence modeling. International Conference on Learning Representations (ICLR)

    2024

  16. [24]

    Pascanu, R.; Mikolov, T.; and Bengio, Y. 2013. On the difficulty of training recurrent neural networks. International Conference on Machine Learning (ICML), 1310--1318

    2013

  17. [25]

    E.; Hinton, G

    Rumelhart, D. E.; Hinton, G. E.; and Williams, R. J. 1985. Learning internal representations by error propagation. Parallel Distributed Processing: Explorations in the Microstructure of Cognition, 1: 318--362

    1985

  18. [26]

    Sen, R.; Yu, H.; and Dhillon, I. 2019. Think globally, act locally: A deep neural network approach to multivariate time series forecasting. NeurIPS

    2019

  19. [27]

    Shamseldin, A. Y. 1997. Application of a neural network technique to rainfall-runoff modelling. Journal of Hydrology, 199(3-4): 272--294

    1997

  20. [29]

    Shi, X.; Chen, Z.; Wang, H.; Yeung, D.-Y.; Wong, W.-K.; and Woo, W.-c. 2015. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in Neural Information Processing Systems, 802--810

    2015

  21. [30]

    Tong, Z.; Jin, Y.; Lai, H.; Wang, Z.; Xing, Z.; Cheng, K.; Xu, H.; Pu, Z.; Zhu, S.; Feng, R.; Zhao, J.; Zhang, Y.; Tang, H.; and Shao, L. 2026. SCOPE: Simulating Cross-game Operations in Playable Environments for FPS World Models. arXiv:2605.23345

    Pith/arXiv arXiv 2026

  22. [31]

    Tsay, R. S. 2005. Analysis of financial time series. John Wiley & Sons

    2005

  23. [32]

    N.; Kaiser, L.; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017 a . Attention is all you need. Advances in Neural Information Processing Systems, 30

    2017

  24. [33]

    N.; Kaiser, L.; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017 b . Attention is all you need. Advances in Neural Information Processing Systems, 30

    2017

  25. [34]

    Wu, H.; Xu, S.; Wang, J.; Long, M.; and Jin, R. 2021. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. Advances in Neural Information Processing Systems, 34: 22419--22430

    2021

  26. [36]

    Yue, Y.; Zhang, Y.; Zhao, J.; and Liu, J. 2022. TSMixer: Multiscale modeling in time series forecasting. ICLR

    2022

  27. [37]

    Zeng, A.; Wang, Z.; et al. 2023. iTransformer: A unified framework for time series forecasting with information fusion. NeurIPS

    2023

  28. [39]

    Zhang, G. P. 2003. Time series forecasting using a hybrid ARIMA and neural network model. Neurocomputing, 50: 159--175

    2003

  29. [40]

    Zhou, H.; Zhang, S.; Peng, J.; Zhang, S.; Li, J.; Xiong, H.; and Zhang, W. 2021. Informer: Beyond efficient transformer for long sequence time-series forecasting. AAAI Conference on Artificial Intelligence, 11106--11115

    2021

  30. [41]

    Zhou, T.; Ma, Z.; Wen, Q.; Wang, X.; Sun, L.; Jin, R.; and Wang, H. 2022. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In International Conference on Machine Learning (ICML), 27361--27378

    2022

  31. [42]

    Journal of Finance , volume=

    Efficient capital markets: A review of theory and empirical work , author=. Journal of Finance , volume=

  32. [43]

    2005 , publisher=

    Analysis of financial time series , author=. 2005 , publisher=

    2005

  33. [44]

    Statistics in Medicine , volume=

    A statistical framework for the analysis of multivariate infectious disease surveillance counts , author=. Statistics in Medicine , volume=

  34. [45]

    Journal of Hydrology , volume=

    Application of a neural network technique to rainfall-runoff modelling , author=. Journal of Hydrology , volume=

  35. [46]

    arXiv preprint arXiv:1803.01271 , year=

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling , author=. arXiv preprint arXiv:1803.01271 , year=

    Pith/arXiv arXiv

  36. [47]

    1994 , publisher=

    Time series analysis , author=. 1994 , publisher=

    1994

  37. [48]

    Neurocomputing , volume=

    Time series forecasting using a hybrid ARIMA and neural network model , author=. Neurocomputing , volume=

  38. [49]

    Neural Computation , volume=

    Long short-term memory , author=. Neural Computation , volume=. 1997 , publisher=

    1997

  39. [50]

    Advances in Neural Information Processing Systems , volume=

    Attention is all you need , author=. Advances in Neural Information Processing Systems , volume=

  40. [51]

    arXiv preprint arXiv:2302.06675 , year=

    Mamba: Linear-time sequence modeling with selective state spaces , author=. arXiv preprint arXiv:2302.06675 , year=

    Pith/arXiv arXiv

  41. [52]

    Parallel Distributed Processing: Explorations in the Microstructure of Cognition , volume=

    Learning internal representations by error propagation , author=. Parallel Distributed Processing: Explorations in the Microstructure of Cognition , volume=. 1985 , publisher=

    1985

  42. [53]

    International Conference on Machine Learning (ICML) , pages=

    On the difficulty of training recurrent neural networks , author=. International Conference on Machine Learning (ICML) , pages=

  43. [54]

    2015 , publisher=

    Time series analysis: forecasting and control , author=. 2015 , publisher=

    2015

  44. [55]

    The Handbook of Brain Theory and Neural Networks , volume=

    Convolutional networks for images, speech, and time series , author=. The Handbook of Brain Theory and Neural Networks , volume=

  45. [56]

    arXiv preprint arXiv:1609.03499 , year=

    Wavenet: A generative model for raw audio , author=. arXiv preprint arXiv:1609.03499 , year=

    Pith/arXiv arXiv

  46. [57]

    Advances in Neural Information Processing Systems , pages=

    Convolutional LSTM network: A machine learning approach for precipitation nowcasting , author=. Advances in Neural Information Processing Systems , pages=

  47. [58]

    International Conference on Machine Learning (ICML) , pages=

    Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting , author=. International Conference on Machine Learning (ICML) , pages=

  48. [59]

    Procedia Technology , volume=

    Weather forecasting model using artificial neural network , author=. Procedia Technology , volume=

  49. [60]

    Applied Soft Computing , volume=

    Financial time series forecasting with deep learning: A systematic literature review: 2005–2019 , author=. Applied Soft Computing , volume=

    2005

  50. [61]

    IEEE Transactions on Computers , volume=

    Time series forecasting in high-performance computing systems , author=. IEEE Transactions on Computers , volume=

  51. [62]

    Neural Computation , volume=

    A survey of recurrent neural networks for time series forecasting , author=. Neural Computation , volume=

  52. [63]

    arXiv preprint arXiv:1909.03002 , year=

    Deep learning for time series forecasting: A survey , author=. arXiv preprint arXiv:1909.03002 , year=

    Pith/arXiv arXiv 1909

  53. [64]

    Neural Networks , volume=

    Neural network pruning techniques , author=. Neural Networks , volume=

  54. [65]

    AAAI Conference on Artificial Intelligence , pages=

    Informer: Beyond efficient transformer for long sequence time-series forecasting , author=. AAAI Conference on Artificial Intelligence , pages=

  55. [66]

    Advances in Neural Information Processing Systems , volume=

    Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting , author=. Advances in Neural Information Processing Systems , volume=

  56. [67]

    arXiv preprint arXiv:2303.07329 , year=

    PatchTST: Transformer-based time series forecasting with patching , author=. arXiv preprint arXiv:2303.07329 , year=

    Pith/arXiv arXiv

  57. [68]

    arXiv preprint arXiv:2304.12345 , year=

    Leveraging large language models for time series forecasting: A survey , author=. arXiv preprint arXiv:2304.12345 , year=

    Pith/arXiv arXiv

  58. [69]

    International Conference on Learning Representations (ICLR) , year=

    Efficiently modeling long sequences with structured state spaces , author=. International Conference on Learning Representations (ICLR) , year=

  59. [70]

    arXiv preprint arXiv:2203.00133 , year=

    H3: Hybrid hierarchical transformers for long-range sequence modeling , author=. arXiv preprint arXiv:2203.00133 , year=

    Pith/arXiv arXiv

  60. [71]

    ICLR , year=

    Advancing time series forecasting with structured state spaces , author=. ICLR , year=

  61. [72]

    channel-dependent models for multivariate time series forecasting , author=

    Channel-independent vs. channel-dependent models for multivariate time series forecasting , author=. Neural Computation , volume=

  62. [73]

    CD: A comparative study on multivariate time series models , author=

    CI vs. CD: A comparative study on multivariate time series models , author=. ICLR , year=

  63. [74]

    NeurIPS , year=

    Self-attention mechanisms for capturing cross-channel dependencies in time series , author=. NeurIPS , year=

  64. [75]

    ICML , year=

    Learning convolutional dependencies in multivariate time series forecasting , author=. ICML , year=

  65. [76]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Advances in vision applications using structured state spaces , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  66. [77]

    Neural Computation , year=

    Graph neural networks with structured state space models , author=. Neural Computation , year=

  67. [78]

    International Conference on Learning Representations (ICLR) , year=

    Mambaformer: Hybrid state space transformer for sequence modeling , author=. International Conference on Learning Representations (ICLR) , year=

  68. [79]

    Conference on Neural Information Processing Systems (NeurIPS) , year=

    Jamba: Joint attention and Mamba models for long-range dependencies , author=. Conference on Neural Information Processing Systems (NeurIPS) , year=

  69. [80]

    Journal of Machine Learning Research , volume=

    A comprehensive survey on multivariate time series forecasting: Models, techniques, and applications , author=. Journal of Machine Learning Research , volume=

  70. [81]

    NeurIPS , year=

    Think globally, act locally: A deep neural network approach to multivariate time series forecasting , author=. NeurIPS , year=

  71. [82]

    Time series analysis: Forecasting and control , author=

  72. [83]

    Neural Computation , volume=

    Long short-term memory , author=. Neural Computation , volume=

  73. [84]

    ICLR , year=

    TSMixer: Multiscale modeling in time series forecasting , author=. ICLR , year=

  74. [85]

    Neural Computation , volume=

    Bi-directional convolutional networks for sequence learning , author=. Neural Computation , volume=

  75. [86]

    arXiv preprint arXiv:2406.05316 , year=

    C-Mamba: Channel Correlation Enhanced State Space Models for Multivariate Time Series Forecasting , author=. arXiv preprint arXiv:2406.05316 , year=

    Pith/arXiv arXiv

  76. [87]

    arXiv preprint arXiv:2404.15772 , year=

    Bimamba+: Bidirectional mamba for time series forecasting (2024) , author=. arXiv preprint arXiv:2404.15772 , year=

    Pith/arXiv arXiv 2024

  77. [88]

    arXiv preprint arXiv:1412.3555 , year=

    Empirical evaluation of gated recurrent neural networks on sequence modeling , author=. arXiv preprint arXiv:1412.3555 , year=

    Pith/arXiv arXiv

  78. [89]

    arXiv preprint arXiv:1701.06538 , year=

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer , author=. arXiv preprint arXiv:1701.06538 , year=

    Pith/arXiv arXiv

  79. [90]

    Journal of Machine Learning Research , year=

    Switch Transformers: Scaling to trillion parameter models with simple and efficient sparsity , author=. Journal of Machine Learning Research , year=

  80. [91]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , year=

    Swin transformer: Hierarchical vision transformer using shifted windows , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , year=

Showing first 80 references.