Pith. sign in

REVIEW 3 major objections 6 minor 161 references

A bank of frozen self-supervised time-series encoders, gated by similarity and structure then Transformer-aggregated, can power multiple digital-twin perception tasks without retraining the encoders.

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 · grok-4.5

2026-07-12 01:21 UTC pith:R6BA5KBP

load-bearing objection Solid industrial systems paper: a reusable frozen-encoder bank with offline gating and Transformer aggregation that works competitively on ETT and a real hydro-generator virtual sensor, with an honest caveat that the gate only helps if the bank already holds the right structure. the 3 major comments →

arxiv 2607.03585 v1 pith:R6BA5KBP submitted 2026-07-03 cs.LG

Modular Foundation Models for Time-Series Perception in Digital Twins

Quang Hung Pham , Ryad Zemouri , Martin Gagnon , Luc Vouligny This is my paper
classification cs.LG
keywords time-series analysisfoundation modelsmixture of expertsself-supervised learningrepresentation learningdigital twinsPHMvirtual sensing
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.

Engineering digital twins need one perception layer that can impute missing sensor values, forecast far ahead, work with few labels, and estimate hard-to-measure quantities such as rotor temperature, yet most models are rebuilt for each task and each machine. This paper claims that a modular foundation model can meet that need: many encoders are first pretrained once, with reconstruction-plus-contrastive self-supervision, on heterogeneous unlabeled series; for any new target a two-stage gate keeps only the most relevant encoders; their outputs are projected into a shared space, aligned, and fused by Transformer self-attention; lightweight heads then solve the actual tasks while the encoders stay frozen. Ablations show that informed gating and adaptive aggregation give the largest gains; on the public ETT suite the system is competitive with strong baselines for imputation and few-shot forecasting and remains usable for long-horizon prediction; a hydro-generator virtual-sensor case study further shows calibrated temperature estimates that transfer across units. The practical claim is that reusable modular representations, rather than end-to-end task-specific networks, can serve as the scalable perception layer for industrial digital twins and hybrid PHM systems.

Core claim

A collection of heterogeneous encoders pretrained once with self-supervised reconstruction and contrastive losses can be frozen, dynamically selected for a new target by data-similarity and structural-correlation gating, projected and Transformer-aggregated into a shared representation, and then adapted with only lightweight heads to support imputation, long-term forecasting, few-shot forecasting, and industrial virtual sensing at competitive accuracy.

What carries the argument

The FM-TSP triple (encoder bank E, gating-plus-aggregation module G, task heads T): two-stage encoder selection by statistical/spectral distance followed by Pearson correlation of input-versus-latent cosine-similarity matrices, linear projection with optional early MMD alignment, and Transformer self-attention over the selected encoder tokens.

Load-bearing premise

The paper assumes that simple statistical and spectral features plus correlation of neighborhood structures are enough to pick which frozen encoders will actually transfer useful representations to an unseen target domain.

What would settle it

On a held-out industrial or ETT-style target whose long-range past–future structure is absent from every encoder in the bank, measure whether the two-stage gate still ranks the truly best encoders first and whether the frozen-plus-light-head pipeline remains competitive with a model trained end-to-end on that target; if ranking fails or accuracy collapses, the modular-transfer claim fails.

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

If this is right

  • A single frozen encoder bank can be shared across many digital-twin assets and tasks, cutting the cost of building and maintaining separate perception models.
  • New sensors or operating regimes can be handled by adding or selecting pretrained encoders rather than retraining the whole stack.
  • Imputation, forecasting, few-shot adaptation, and virtual sensing become interchangeable heads on the same modular backbone.
  • Hybrid PHM pipelines can plug the same perception layer into physics-based models without task-specific redesign.
  • Conditional computation at the encoder level keeps inference cost low even as the bank grows.

Where Pith is reading between the lines

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

  • If the gate’s proxy ranking is weak for forecasting structure, the practical fix is to enlarge the encoder bank with models deliberately pretrained on long-horizon prediction objectives rather than redesigning the aggregation head.
  • The same modular pattern could be applied to anomaly detection or remaining-useful-life heads without changing the frozen representation layer.
  • Because encoders stay frozen, the architecture is naturally suited to continual addition of new assets or sites without catastrophic forgetting of earlier regimes.

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

3 major / 6 minor

Summary. The paper proposes FM-TSP, a modular foundation model for time-series perception in digital twins and PHM. A bank of heterogeneous encoders (MLP/CNN/Transformer) is pretrained with a reconstruction–contrastive objective on diverse unlabeled datasets and then frozen. For a target dataset, a two-stage offline gate ranks encoders by statistical/spectral data similarity and by Pearson correlation of input vs. latent cosine-similarity matrices, selects a top-k subset, projects the latent vectors into a shared space (optionally with early MMD alignment), and aggregates them with a Transformer self-attention module. Lightweight task heads support imputation, long-term forecasting, few-shot forecasting, and probabilistic virtual sensing while only projectors, aggregator, and heads are fine-tuned. Ablations isolate pretraining, gating, projection/alignment, and aggregation; ETT results are competitive with DLinear, LightTS, PatchTST, and iTransformer; a hydro-generator rotor-temperature virtual-sensor case shows practical transfer across units.

Significance. If the modular transfer story holds, the work offers a practical perception layer for industrial digital twins: reusable frozen encoders, conditional computation, multi-task heads, and limited labeled adaptation. Strengths include systematic ablations with multi-metric box plots (MSE, DTW, Jacobian), shared-protocol ETT comparisons, and a real multi-unit industrial virtual-sensor study with calibrated uncertainty. The design is well aligned with PHM constraints (heterogeneous sensors, scarce labels, non-stationarity) and is more modular than typical end-to-end time-series foundation models. The main significance is architectural and empirical rather than theoretical; the contribution is a coherent, deployable recipe rather than a new universal approximator or scaling law.

major comments (3)
  1. §4.3 (H1–H3) and §6.1.5: The central claim that the offline gate reliably activates task-useful frozen encoders is only weakly supported. Ablation Fig. 4 shows top-k beats random/worst-k under the same proxy, but does not test whether high structural-correlation scores predict downstream utility when the library lacks the needed structure. The authors themselves note weaker long-term forecasting because encoders were not pretrained on ETT and may miss past–future dependencies. A load-bearing experiment is missing: correlation of gate scores with task metrics under controlled library incompleteness, or an oracle/task-aware selection baseline. Without this, multi-task competitiveness and the industrial case remain contingent on a sufficiently rich bank rather than on the selection mechanism itself.
  2. §6.1.4–6.1.5 and Figs. 7–9: ETT results are reported only as overall averages across subsets and horizons, without per-dataset/per-horizon tables or standard deviations. Given that long-term forecasting is already acknowledged as weaker than imputation/few-shot, aggregate plots make it hard to judge whether competitiveness is uniform or driven by easier settings. Full numerical tables (MSE/MAE/DTW per ETTh1/h2/m1/m2 and horizon) against the reimplemented baselines are needed for the benchmark claim to be verifiable.
  3. §6.2.3 and Fig. 11: The industrial transfer result is important but incomplete. Coverage drops from ~93.5% on the training TGU to ~65% on the held-out TGU, with systematic overestimation at low peaks. The paper does not quantify how much of this gap is closed by the modular gate versus a single encoder or a non-gated concatenation of the same bank, nor does it report error metrics (e.g., MAE/RMSE) alongside coverage. A controlled comparison on the same multi-unit split is required to substantiate that modular selection, rather than the probabilistic head alone, drives practical virtual-sensing performance.
minor comments (6)
  1. §1 and §2: The Jacobs et al. (1991) and Vats et al. (2024) citations are duplicated in consecutive sentences; clean the redundancy.
  2. Definition 3 and §4.1: Typographical issues (“agregartion”, “τT” indexing, “Noted that”) and inconsistent notation for the aggregation map G should be corrected.
  3. Eq. (17) vs. Table 1: The contrastive weight α is described inconsistently (trade-off with reconstruction vs. “contrastive weight”); clarify the parameterization used in experiments.
  4. §5: Jacobian-norm interpretation is useful but under-specified (which input perturbations, layer, and normalization). A short formal definition would help readers interpret the box plots.
  5. Figures 7–11: Axis labels, metric units, and exact masking/horizon settings are hard to read from the text alone; ensure captions are self-contained.
  6. Related work: Time-MoE, MOIRAI-MoE, and recent time-series foundation models are cited; a clearer positioning of offline encoder-level gating versus token-level sparse MoE would strengthen novelty claims.

Circularity Check

0 steps flagged

No significant circularity: standard empirical ML pipeline with frozen SSL encoders, external ETT splits, and held-out industrial transfer; self-citations supply motivation only.

full rationale

The paper's load-bearing claims are empirical performance numbers (imputation / long-term / few-shot forecasting on ETT; virtual rotor-temperature sensing on a held-out hydro-generator unit). Pretraining uses reconstruction + NT-Xent on heterogeneous non-ETT corpora with encoders then frozen; fine-tuning updates only projectors, Transformer aggregation, and task heads on train splits and reports held-out test metrics against reimplemented external baselines (DLinear, LightTS, PatchTST, iTransformer). Gating (data-similarity via statistical/spectral ϕ plus structural Pearson of cosine matrices under H1–H3) is a design heuristic validated by top-k vs random/worst-k ablations, not a quantity defined from the reported test losses. Self-citations (Jose et al. 2026; Zemouri et al. 2025) motivate modular MoE for hydrogenerator diagnostics but do not supply the ETT or virtual-sensor numbers. No uniqueness theorem, fitted identity renamed as prediction, or self-definitional reduction appears. The authors' own admission that long-term forecasting weakens when the encoder bank lacks past–future structure is an honesty about transfer limits, not circularity. Score 0 is the correct outcome.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 2 invented entities

The work is empirical systems ML. Load-bearing content is architectural choices and selection heuristics rather than physical laws. Free parameters control pretraining trade-offs, gate width, alignment strength, and task heads; axioms are standard SSL/MoE assumptions plus the paper’s three transfer hypotheses for gating; invented entities are the FM-TSP composite and its offline two-stage gate, not new physical objects.

free parameters (6)
  • contrastive–reconstruction weight α
    Traded off in Eq. (17) and Table 1 over {0,0.25,0.5,0.75,1}; controls encoder representation quality used by all downstream claims.
  • number of selected encoders NE / top-k
    Gate retains NE encoders; performance claims depend on this hand-chosen width (ablation contrasts top-k vs random/worst-k).
  • MMD alignment coefficient β and warm-up epochs ne
    Eq. (29) schedule; authors note no principled β selection beyond empirical tuning, yet alignment is part of the claimed pipeline.
  • masking ratio and latent dimensions
    Pretraining diversity factors in Table 1 (masking 0/0.25/0.5; dims 8–64) that shape the frozen bank.
  • shared projection dimension d and Transformer aggregation hyperparameters
    Define the fused space DG; not uniquely determined by theory.
  • virtual-sensor Beta-loss / mean–variance head parameters
    Industrial case head predicts distribution parameters; calibration coverage depends on this likelihood choice and training.
axioms (4)
  • domain assumption Self-supervised reconstruction + contrastive pretraining on heterogeneous unlabeled series yields transferable task-agnostic representations usable when encoders are frozen.
    Core premise of §4.2 and the foundation-model framing; supported by SSL literature but not proved for the industrial regimes studied.
  • ad hoc to paper (H1) Encoders pretrained on distributions closer to the target extract more informative representations; (H2) statistical/spectral embeddings suffice to represent the process; (H3) a good encoder preserves local neighborhood structure across input and latent spaces.
    Explicitly stated in §4.3 as the basis of the two-stage gate; load-bearing for conditional computation claims.
  • domain assumption Mixture-of-experts / conditional computation improves scalability and regime specialization for non-stationary industrial series.
    Motivation in §§1–2 citing Jacobs et al., Time-MoE, industrial MoE diagnostics.
  • domain assumption Standard supervised losses on ETT splits and industrial train/val/test splits are valid proxies for perception quality in digital twins.
    Evaluation design in §6; usual ML benchmark assumption.
invented entities (2)
  • FM-TSP modular foundation model (E, G, T) with frozen multi-architecture encoder bank no independent evidence
    purpose: Unify multi-task time-series perception for digital twins under limited labels.
    Definition 3 and §4; composite architecture rather than a new physical entity. independent_evidence false beyond this paper’s experiments.
  • Two-stage offline encoder gate (data similarity d(Pi,Y) + structural correlation r(Ei,Y)) no independent evidence
    purpose: Select NE relevant frozen encoders without joint retraining.
    §4.3 Eqs. (21)–(24); heuristic selection mechanism introduced for this framework.

pith-pipeline@v1.1.0-grok45 · 22075 in / 3635 out tokens · 31061 ms · 2026-07-12T01:21:54.495680+00:00 · methodology

0 comments
read the original abstract

Engineering Digital Twins and Prognostics and Health Management (PHM) systems rely on robust perception modules to extract actionable information from heterogeneous and non-stationary time-series data. However, most existing approaches remain task-specific, data-hungry, and difficult to integrate into scalable monitoring and decision-making pipelines. Moreover, purely data-driven models often lack robustness and transferability across varying operating conditions. To address these challenges, this paper proposes a modular foundation model for time-series perception based on a collection of pretrained representation encoders. The framework leverages self-supervised learning on heterogeneous datasets to learn transferable and task-agnostic representations, which can be reused across multiple PHM tasks. A gating mechanism is introduced to dynamically select relevant encoders for a given target dataset, enabling conditional computation and adaptive model composition. The selected representations are projected into a shared latent space and aggregated using a Transformer-based self-attention module that explicitly models cross-encoder interactions. The resulting architecture supports multiple downstream tasks, including imputation, long-term forecasting, and few-shot learning, through lightweight task-specific heads, while keeping pretrained encoders frozen during adaptation. Extensive ablation studies demonstrate the complementary roles of self-supervised pretraining, encoder selection, representation alignment, and adaptive aggregation. Experimental results on the ETT benchmark show competitive performance across tasks, while a real-world industrial case study on virtual sensing for hydro-generator rotor temperature highlights the practical relevance of the approach.

Figures

Figures reproduced from arXiv: 2607.03585 by Luc Vouligny, Martin Gagnon, Quang Hung Pham, Ryad Zemouri.

Figure 1
Figure 1. Figure 1: Conceptual framework of the proposed foundation model for time-series processing. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Encoder pretraining with a lightweight decoder to focus learning on the encoder. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of Self-Supervised Pretraining 5.3 Impact of Encoder Gating and Selection The second ablation examines the role of the encoder gating and selection mechanism described in Section IV-C. We evaluate three configurations: 1. Random selection: A subset of encoders is randomly sampled, and this selection process is repeated 10 times to ensure diversity; 11 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: shows that random selection yields high variance and inconsistent performance, whereas worst-k selection leads to the poorest results. The near-zero Jacobian-based values in the latter case indicate that the selected encoders do not learn meaningful representations. In contrast, the strong performance of top-k selection confirms that the proposed gating strategy consistently improves both accuracy and robu… view at source ↗
Figure 5
Figure 5. Figure 5: Role of Projection and Representation Alignment [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Impact of Downstream aggregation All configurations share the same encoder selection, projection, and fine-tuning settings, and differ only in the aggregation strategy. Experimental results, summarized in [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance comparison for imputation downstream task [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance comparison for long-term forecasting task [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Performance comparison for few-shot learning forecasting task [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Temperature estimated by virtual sensor at training TGU. The zoomed-in section is taken from the split test [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Temperature estimated by virtual sensor at test TGU [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] 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

161 extracted references · 23 canonical work pages

  1. [1]

    Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Foundation models for time series analysis: A tutorial and survey , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  2. [2]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Self-supervised learning for time series analysis: Taxonomy, progress, and prospects , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  3. [3]

    arXiv preprint arXiv:2310.10688 , year=

    A decoder-only foundation model for time-series forecasting , author=. arXiv preprint arXiv:2310.10688 , year=

    Pith/arXiv arXiv

  4. [4]

    ACM Computing Surveys (CSUR) , volume=

    How complex is your classification problem? a survey on measuring classification complexity , author=. ACM Computing Surveys (CSUR) , volume=. 2019 , publisher=

    2019

  5. [5]

    Sustainability , volume=

    Few-shot learning approaches for fault diagnosis using vibration data: a comprehensive review , author=. Sustainability , volume=. 2023 , publisher=

    2023

  6. [6]

    PHM Soc Eur Conf

    Diagnosis of tidal turbine vibration data through deep neural networks. PHM Soc Eur Conf. 2016 , author=

    2016

  7. [7]

    Journal of Computational Design and Engineering , volume=

    Vibration data feature extraction and deep learning-based preprocessing method for highly accurate motor fault diagnosis , author=. Journal of Computational Design and Engineering , volume=. 2023 , publisher=

    2023

  8. [8]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Multimodal learning with transformers: A survey , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2023 , publisher=

    2023

  9. [9]

    arXiv preprint arXiv:2406.12034 , year=

    Self-MoE: Towards Compositional Large Language Models with Self-Specialized Experts , author=. arXiv preprint arXiv:2406.12034 , year=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2402.05859 , year=

    Learning to route among specialized experts for zero-shot generalization , author=. arXiv preprint arXiv:2402.05859 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2403.09809 , year=

    Self-Supervised Learning for Time Series: Contrastive or Generative? , author=. arXiv preprint arXiv:2403.09809 , year=

    Pith/arXiv arXiv

  12. [12]

    Information Sciences , volume=

    Multi-task self-supervised time-series representation learning , author=. Information Sciences , volume=. 2024 , publisher=

    2024

  13. [13]

    Proceedings of Neural Information Processing Systems, NeurIPS , year =

    Self-Supervised Contrastive Pre-Training For Time Series via Time-Frequency Consistency , author =. Proceedings of Neural Information Processing Systems, NeurIPS , year =

  14. [14]

    arXiv preprint arXiv:2402.02592 , year=

    Unified training of universal time series forecasting transformers , author=. arXiv preprint arXiv:2402.02592 , year=

    Pith/arXiv arXiv

  15. [15]

    Preprint , year=

    Lag-llama: Towards foundation models for probabilistic time series forecasting , author=. Preprint , year=

  16. [16]

    arXiv preprint arXiv:2403.00131 , year=

    Units: Building a unified time series model , author=. arXiv preprint arXiv:2403.00131 , year=

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2310.03589 , year=

    TimeGPT-1 , author=. arXiv preprint arXiv:2310.03589 , year=

    Pith/arXiv arXiv

  18. [18]

    Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining , pages=

    Pre-training enhanced spatial-temporal graph neural network for multivariate time series forecasting , author=. Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining , pages=. 2022 , note=

    2022

  19. [19]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Ts2vec: Towards universal representation of time series , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2022 , note=

    2022

  20. [20]

    arXiv preprint arXiv:2303.00320 , year=

    Timemae: Self-supervised representations of time series with decoupled masked autoencoders , author=. arXiv preprint arXiv:2303.00320 , year=

    arXiv

  21. [21]

    Advances in Neural Information Processing Systems , volume=

    Self-supervised contrastive pre-training for time series via time-frequency consistency , author=. Advances in Neural Information Processing Systems , volume=. 2022 , note=

    2022

  22. [22]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    A review of generalized zero-shot learning methods , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2022 , publisher=

    2022

  23. [23]

    IEEE transactions on knowledge and data engineering , volume=

    Self-supervised learning: Generative or contrastive , author=. IEEE transactions on knowledge and data engineering , volume=. 2021 , publisher=

    2021

  24. [24]

    arXiv preprint arXiv:2403.07815 , year=

    Chronos: Learning the language of time series , author=. arXiv preprint arXiv:2403.07815 , year=

    Pith/arXiv arXiv

  25. [25]

    2024 , note =

    Nixtla Team , title =. 2024 , note =

    2024

  26. [26]

    Advances in Neural Information Processing Systems , volume=

    Simmtm: A simple pre-training framework for masked time-series modeling , author=. Advances in Neural Information Processing Systems , volume=. 2024 , note=

    2024

  27. [27]

    Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining , pages=

    A transformer-based framework for multivariate time series representation learning , author=. Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining , pages=. 2021 , note=

    2021

  28. [28]

    arXiv preprint arXiv:2106.14112 , year=

    Time-series representation learning via temporal and contextual contrasting , author=. arXiv preprint arXiv:2106.14112 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:1706.08838 , year=

    TimeNet: Pre-trained deep recurrent neural network for time series classification , author=. arXiv preprint arXiv:1706.08838 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:1610.07258 , year=

    Representation learning with deconvolution for multivariate time series classification and visualization , author=. arXiv preprint arXiv:1610.07258 , year=

    Pith/arXiv arXiv

  31. [31]

    ACM Transactions on Intelligent Systems and Technology (TIST) , volume=

    Representation learning on variable length and incomplete wearable-sensory time series , author=. ACM Transactions on Intelligent Systems and Technology (TIST) , volume=. 2022 , publisher=

    2022

  32. [32]

    Representation learning using a multi-branch transformer for industrial time series anomaly detection , author=

  33. [33]

    IEEE Transactions on Neural Networks and Learning Systems , year=

    Self-supervised time series representation learning via cross reconstruction transformer , author=. IEEE Transactions on Neural Networks and Learning Systems , year=

  34. [34]

    arXiv preprint arXiv:2308.01578 , year=

    Unsupervised representation learning for time series: A review , author=. arXiv preprint arXiv:2308.01578 , year=

    Pith/arXiv arXiv

  35. [35]

    7th international Mardin Artuklu scientific research conference , pages=

    Ideal dataset splitting ratios in machine learning algorithms: general concerns for data scientists and data analysts , author=. 7th international Mardin Artuklu scientific research conference , pages=

  36. [36]

    Artificial Intelligence Review , volume=

    Transformer models for text-based emotion detection: a review of BERT-based approaches , author=. Artificial Intelligence Review , volume=. 2021 , publisher=

    2021

  37. [37]

    Exploring the power of ChatGPT: Applications, techniques, and implications , pages=

    What Is ChatGPT? , author=. Exploring the power of ChatGPT: Applications, techniques, and implications , pages=. 2023 , publisher=

    2023

  38. [38]

    2025 , url =

    IBM , title =. 2025 , url =

    2025

  39. [39]

    Proceedings of the 35th AAAI Conference on Artificial Intelligence , year =

    Haoyi Zhou and Shanghang Zhang and Jieqi Peng and Shuai Zhang and Jianxin Li and Hui Xiong and Wancai Zhang , title =. Proceedings of the 35th AAAI Conference on Artificial Intelligence , year =

  40. [40]

    , title =

    Trindade, A. , title =. 2015 , howpublished =

    2015

  41. [41]

    2020 , howpublished =

    Kolle, Olaf , title =. 2020 , howpublished =

    2020

  42. [42]

    California Performance Measurement System (PeMS) , year =

  43. [43]

    Barrage Hydraulique , year =

  44. [44]

    2020 , url =

    Towards Data Science , title =. 2020 , url =

    2020

  45. [46]

    arXiv preprint arXiv:1807.03748 , year=

    Representation learning with contrastive predictive coding , author=. arXiv preprint arXiv:1807.03748 , year=

    Pith/arXiv arXiv

  46. [47]

    2021 , isbn =

    Zerveas, George and Jayaraman, Srideepika and Patel, Dhaval and Bhamidipaty, Anuradha and Eickhoff, Carsten , title =. 2021 , isbn =. doi:10.1145/3447548.3467401 , booktitle =

    doi:10.1145/3447548.3467401 2021

  47. [48]

    Siddharth and Brooks Paige and Jan-Willem van de Meent and Alban Desmaison and Noah D

    N. Siddharth and Brooks Paige and Jan-Willem van de Meent and Alban Desmaison and Noah D. Goodman and Pushmeet Kohli and Frank Wood and Philip H. S. Torr , title =. 2017 , archiveprefix =. 1706.00400 , file =

    Pith/arXiv arXiv 2017

  48. [49]

    2023 , archiveprefix =

    Jonas Pfeiffer and Sebastian Ruder and Ivan Vulić and Edoardo Maria Ponti , title =. 2023 , archiveprefix =. 2302.11529 , file =

    Pith/arXiv arXiv 2023

  49. [50]

    Hudson and Ehsan Adeli and Russ B

    Rishi Bommasani and Drew A. Hudson and Ehsan Adeli and Russ B. Altman and Simran Arora and Sydney von Arx and Michael S. Bernstein and Jeannette Bohg and Antoine Bosselut and Emma Brunskill and Erik Brynjolfsson and Shyamal Buch and Dallas Card and Rodrigo Castellon and Niladri S. Chatterji and Annie S. Chen and Kathleen Creel and Jared Quincy Davis and D...

    Pith/arXiv arXiv 2021

  50. [51]

    2022 , archiveprefix =

    Nasim Rahaman and Martin Weiss and Frederik Träuble and Francesco Locatello and Alexandre Lacoste and Yoshua Bengio and Chris Pal and Li Erran Li and Bernhard Schölkopf , title =. 2022 , archiveprefix =. 2211.02348 , file =

    Pith/arXiv arXiv 2022

  51. [52]

    2022 , archiveprefix =

    Nasim Rahaman and Martin Weiss and Francesco Locatello and Chris Pal and Yoshua Bengio and Bernhard Schölkopf and Li Erran Li and Nicolas Ballas , title =. 2022 , archiveprefix =. 2210.08031 , file =

    Pith/arXiv arXiv 2022

  52. [53]

    CoRR , title =

    Andrew Jaegle and Sebastian Borgeaud and Jean. CoRR , title =. 2021 , volume =. 2107.14795 , eprinttype =

    Pith/arXiv arXiv 2021

  53. [54]

    Towards artificial general intelligence via a multimodal foundation model , year =

    Fei, Nanyi and Lu, Zhiwu and Gao, Yizhao and Yang, Guoxing and Huo, Yuqi and Wen, Jingyuan and Lu, Haoyu and Song, Ruihua and Gao, Xin and Xiang, Tao and Sun, Hao and Wen, Ji-Rong , journal =. Towards artificial general intelligence via a multimodal foundation model , year =. doi:10.1038/s41467-022-30761-2 , file =

    work page doi:10.1038/s41467-022-30761-2

  54. [55]

    Dusenberry and Du Phan and Mark Collier and Jie Ren and Kehang Han and Zi Wang and Zelda Mariet and Huiyi Hu and Neil Band and Tim G

    Dustin Tran and Jeremiah Liu and Michael W. Dusenberry and Du Phan and Mark Collier and Jie Ren and Kehang Han and Zi Wang and Zelda Mariet and Huiyi Hu and Neil Band and Tim G. J. Rudner and Karan Singhal and Zachary Nado and Joost van Amersfoort and Andreas Kirsch and Rodolphe Jenatton and Nithum Thain and Honglin Yuan and Kelly Buchanan and Kevin Murph...

    Pith/arXiv arXiv 2022

  55. [56]

    2020 , archiveprefix =

    Michael Crawshaw , title =. 2020 , archiveprefix =. 2009.09796 , file =

    Pith/arXiv arXiv 2020

  56. [57]

    Multi-Task Learning for Dense Prediction Tasks: A Survey , year =

    Vandenhende, Simon and Georgoulis, Stamatios and Van Gansbeke, Wouter and Proesmans, Marc and Dai, Dengxin and Van Gool, Luc , journal =. Multi-Task Learning for Dense Prediction Tasks: A Survey , year =. doi:10.1109/TPAMI.2021.3054719 , file =

    doi:10.1109/tpami.2021.3054719 2021

  57. [58]

    and Mann, Matthias , journal =

    Zeng, Wen-Feng and Zhou, Xie-Xuan and Willems, Sander and Ammar, Constantin and Wahle, Maria and Bludau, Isabell and Voytik, Eugenia and Strauss, Maximillian T. and Mann, Matthias , journal =. 2022 , issn =. doi:10.1038/s41467-022-34904-3 , file =

    work page doi:10.1038/s41467-022-34904-3 2022

  58. [59]

    Advances in Neural Information Processing Systems , title =

    Swamy, Vinitra and Satayeva, Malika and Frej, Jibril and Bossy, Thierry and Vogels, Thijs and Jaggi, Martin and K\". Advances in Neural Information Processing Systems , title =. 2023 , editor =

    2023

  59. [60]

    2024 , archiveprefix =

    Byeongjun Park and Sangmin Woo and Hyojun Go and Jin-Young Kim and Changick Kim , title =. 2024 , archiveprefix =. 2310.07138 , file =

    Pith/arXiv arXiv 2024

  60. [61]

    2023 , archiveprefix =

    Mohammed Muqeeth and Haokun Liu and Colin Raffel , title =. 2023 , archiveprefix =. 2306.03745 , file =

    Pith/arXiv arXiv 2023

  61. [62]

    Salman Asif , title =

    Yash Garg and Nebiyou Yismaw and Rakib Hyder and Ashley Prater-Bennette and M. Salman Asif , title =. 2023 , archiveprefix =. 2310.06124 , file =

    arXiv 2023

  62. [63]

    2023 , month = Dec, type =

    Sun, Haozhe , school =. 2023 , month = Dec, type =

    2023

  63. [64]

    CoRR , title =

    Bernhard Sch. CoRR , title =. 2021 , volume =. 2102.11107 , eprinttype =

    Pith/arXiv arXiv 2021

  64. [65]

    Inductive Biases for Deep Learning of Higher-Level Cognition , year =

    Anirudh Goyal and Yoshua Bengio , journal =. Inductive Biases for Deep Learning of Higher-Level Cognition , year =. 2011.15091 , eprinttype =

    Pith/arXiv arXiv 2011

  65. [66]

    2023 , archiveprefix =

    Clifton Poth and Hannah Sterz and Indraneil Paul and Sukannya Purkayastha and Leon Engländer and Timo Imhof and Ivan Vulić and Sebastian Ruder and Iryna Gurevych and Jonas Pfeiffer , title =. 2023 , archiveprefix =. 2311.11077 , file =

    Pith/arXiv arXiv 2023

  66. [67]

    Yu and Lichao Sun , title =

    Ce Zhou and Qian Li and Chen Li and Jun Yu and Yixin Liu and Guangjing Wang and Kai Zhang and Cheng Ji and Qiben Yan and Lifang He and Hao Peng and Jianxin Li and Jia Wu and Ziwei Liu and Pengtao Xie and Caiming Xiong and Jian Pei and Philip S. Yu and Lichao Sun , title =. 2023 , archiveprefix =. 2302.09419 , file =

    Pith/arXiv arXiv 2023

  67. [68]

    CoRR , title =

    Junyi Li and Tianyi Tang and Wayne Xin Zhao and Ji. CoRR , title =. 2021 , volume =. 2105.10311 , eprinttype =

    Pith/arXiv arXiv 2021

  68. [69]

    Data Efficient Masked Language Modeling for Vision and Language , year =

    Yonatan Bitton and Gabriel Stanovsky and Michael Elhadad and Roy Schwartz , journal =. Data Efficient Masked Language Modeling for Vision and Language , year =. 2109.02040 , eprinttype =

    Pith/arXiv arXiv

  69. [70]

    and Healy, Graham and Smeaton, Alan F

    Le-Khac, Phuc H. and Healy, Graham and Smeaton, Alan F. , journal =. Contrastive Representation Learning: A Framework and Review , year =. doi:10.1109/ACCESS.2020.3031549 , file =

    doi:10.1109/access.2020.3031549 2020

  70. [71]

    A Theoretical Analysis of Contrastive Unsupervised Representation Learning , year =

    Saunshi, Nikunj and Plevrakis, Orestis and Arora, Sanjeev and Khodak, Mikhail and Khandeparkar, Hrishikesh , booktitle =. A Theoretical Analysis of Contrastive Unsupervised Representation Learning , year =

  71. [72]

    Martinez and Juan Manuel Gorriz , journal =

    Yu-Dong Zhang and Zhengchao Dong and Shui-Hua Wang and Xiang Yu and Xujing Yao and Qinghua Zhou and Hua Hu and Min Li and Carmen Jiménez-Mesa and Javier Ramirez and Francisco J. Martinez and Juan Manuel Gorriz , journal =. Advances in multimodal data fusion in neuroimaging: Overview, challenges, and novel orientation , year =. doi:https://doi.org/10.1016/...

    work page doi:10.1016/j.inffus.2020.07.006 2020

  72. [73]

    Deep multimodal fusion for semantic image segmentation: A survey , year =

    Yifei Zhang and Désiré Sidibé and Olivier Morel and Fabrice Mériaudeau , journal =. Deep multimodal fusion for semantic image segmentation: A survey , year =. doi:https://doi.org/10.1016/j.imavis.2020.104042 , keywords =

    work page doi:10.1016/j.imavis.2020.104042 2020

  73. [74]

    Attention Bottlenecks for Multimodal Fusion , year =

    Nagrani, Arsha and Yang, Shan and Arnab, Anurag and Jansen, Aren and Schmid, Cordelia and Sun, Chen , booktitle =. Attention Bottlenecks for Multimodal Fusion , year =

  74. [75]

    and Marks, Tim K

    Hori, Chiori and Hori, Takaaki and Lee, Teng-Yok and Zhang, Ziming and Harsham, Bret and Hershey, John R. and Marks, Tim K. and Sumi, Kazuhiko , booktitle =. Attention-Based Multimodal Fusion for Video Description , year =

  75. [76]

    Multi-source heterogeneous data fusion , year =

    Zhang, Lili and Xie, Yuxiang and Xidao, Luan and Zhang, Xin , booktitle =. Multi-source heterogeneous data fusion , year =. doi:10.1109/ICAIBD.2018.8396165 , keywords =

    work page doi:10.1109/icaibd.2018.8396165 2018

  76. [77]

    Lifting the Curse of Multilinguality by Pre-training Modular Transformers , year =

    Pfeiffer, Jonas and Goyal, Naman and Lin, Xi and Li, Xian and Cross, James and Riedel, Sebastian and Artetxe, Mikel , booktitle =. Lifting the Curse of Multilinguality by Pre-training Modular Transformers , year =. doi:10.18653/v1/2022.naacl-main.255 , url =

    doi:10.18653/v1/2022.naacl-main.255 2022

  77. [78]

    Beyond English-Centric Multilingual Machine Translation , year =

    Angela Fan and Shruti Bhosale and Holger Schwenk and Zhiyi Ma and Ahmed El-Kishky and Siddharth Goyal and Mandeep Baines and Onur Celebi and Guillaume Wenzek and Vishrav Chaudhary and Naman Goyal and Tom Birch and Vitaliy Liptchinsky and Sergey Edunov and Michael Auli and Armand Joulin , journal =. Beyond English-Centric Multilingual Machine Translation , year =

  78. [79]

    and Zettlemoyer, Luke , booktitle =

    Gururangan, Suchin and Lewis, Mike and Holtzman, Ari and Smith, Noah A. and Zettlemoyer, Luke , booktitle =. 2022 , address =. doi:10.18653/v1/2022.naacl-main.407 , url =

    doi:10.18653/v1/2022.naacl-main.407 2022

  79. [80]

    Smith and Luke Zettlemoyer , title =

    Margaret Li and Suchin Gururangan and Tim Dettmers and Mike Lewis and Tim Althoff and Noah A. Smith and Luke Zettlemoyer , title =. 2022 , archiveprefix =. 2208.03306 , primaryclass =

    Pith/arXiv arXiv 2022

  80. [81]

    Modular Networks: Learning to Decompose Neural Computation , year =

    Kirsch, Louis and Kunze, Julius and Barber, David , booktitle =. Modular Networks: Learning to Decompose Neural Computation , year =

Showing first 80 references.