Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

From Images to Signals: Are Large Vision Models Useful for Time Series Analysis?

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper establishes that pre-trained large vision models are useful for time series classification, but their best forecasting performance is a periodicity-dependent effect tied to specific models, imaging methods, and pre-trained…

desk verdict This is the most comprehensive audit of image-pretrained LVMs for time series yet, and the TSC results are credible; but the RQ8 period-bias analysis is confounded by the fixed-size resizing pipeline. read the letter →

arxiv 2505.24030 v2 pith:TVEY5INM submitted 2025-05-29 cs.LG cs.AIcs.CV

Ziming Zhao , ChengAo Shen , Hanghang Tong , Dongjin Song , Zhigang Deng , Qingsong Wen , Jingchao Ni This is my paper
classification cs.LGcs.AIcs.CV
keywords largevisionmodelstimeseriesclassificationforecastingimagingmaskedautoencodersGramianangularfieldunivariateheatmapperiodicitybias
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Are image-pretrained large vision models (LVMs) useful for time series? This paper runs the first broad comparison—4 LVMs, 8 ways to render a series as an image, 10 classification and 8 forecasting benchmarks, 26 baselines—and answers yes for classification, with caveats for forecasting. On classification, a ViT fed Gramian Angular Field images beats 18 baseline methods on average accuracy, and both zero-shot and lightly fine-tuned LVMs outperform training from scratch, so the pre-training transfers real pattern knowledge. On forecasting, a masked-autoencoder LVM with univariate heatmap (UVH) imaging wins most settings, but only when the series has a strong dominant period: the forecaster effectively stitches past periods together, and the edge disappears on weakly periodic data. The paper's central conclusion is that LVMs are a promising semantic feature extractor for time series but not yet a general numerical forecaster.

What carries the argument

The load-bearing machinery is the imaging-and-reconstruction pipeline, especially the univariate heatmap (UVH) used for forecasting. UVH takes a univariate series, finds the dominant period L by FFT, pads the series to a multiple of L, and stacks length-L segments into a 2D image; GAF, by contrast, encodes pairwise temporal correlations into a matrix and is the best imaging choice for classification. The forecasting pipeline then feeds this image to a masked autoencoder, masks the patches covering the forecast horizon, and lets the pre-trained decoder reconstruct them. The paper's Lemma 1 is the clarifying identity: for a perfectly periodic series divided into length (i/k)L segments, the smallest number of segments before any segment reappears is n = k/GCD(i,k), and this n predicts the difficulty curve of the forecaster. That identity is what turns the observed period-bias into a testable mechanism rather than a qualitative impression.

What would settle it

On a strongly periodic dataset such as ETTh1, sweep the UVH row length from L/6 to 2L while keeping the LVM and decoder fixed; the paper predicts a sharp MSE minimum at L and 2L and an M-shaped difficulty curve following n = k/GCD(i,k). If the error curve is flat across all row lengths, or if the minimum lands at a non-periodic row length, the period-bias mechanism is not the reason for the forecasting gains.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that LVM usefulness splits by task level. For high-level classification, imaged time series carry recognizable visual structure: Gramian Angular Field with a supervised ViT gives the best average accuracy among 19 compared methods, and temporal-order perturbations degrade LVMs more than simpler models, showing the model genuinely uses ordering. For low-level forecasting, the best configuration is a self-supervised masked autoencoder (MAE) paired with UVH imaging and its pre-trained reconstruction decoder; the decoder turns out to matter more than the encoder, because it was trained to reproduce continuous pixel values. The same mechanism creates an inductive bias: UVH stacks the series into period-length rows, and the forecaster performs best when the segment length equals the FFT-detected period L or 2L, effectively forecasting by combining past segments. Hence the paper states that current best LVM forecasters are limited to specific LVM types and imaging methods, biased toward forecasting periods, and unable to exploit very long look-back windows.

Load-bearing premise

The argument assumes that turning a time series into a fixed-size image—resizing with bilinear interpolation, duplicating channels, and normalizing—preserves enough temporal order for the LVM to exploit, and that the single FFT-detected period L used to build UVH images faithfully describes both the look-back and forecast windows.

Editorial extensions

If this is right

  • For time series classification, imaging plus a pre-trained vision transformer is a competitive recipe: GAF with ViT matches or beats strong baselines, and even zero-shot LVMs beat training from scratch.
  • For forecasting, self-supervised LVMs outperform supervised ones, and the pre-trained decoder contributes more than the encoder, so future LVM forecasters should retain or redesign the reconstruction head rather than discard it.
  • LVM-based forecasting inherits a periodicity bias: on datasets with a clear dominant period (ETT, Traffic) MAE with UVH is strong, while on weakly periodic series (Weather, Illness) the advantage shrinks or vanishes.
  • Longer look-back windows help only up to roughly 1000 time steps, after which performance plateaus or declines, implying current fixed-size image inputs cap how much history an LVM can use.
  • Compared with the best tested baselines, LVMs are more expensive to train and slower at forecasting inference, so their gains currently trade compute for accuracy.

Reading between the lines

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

  • If the period-bias diagnosis is right, the apparent forecasting ability of LVM encoders may be mostly a learned image-interpolation effect: a direct test is to run the same masked-reconstruction pipeline on aperiodic synthetic series and check whether error rises sharply as row length deviates from the FFT period.
  • The fixed 224x224 input size means long series are compressed by bilinear resampling, so the look-back plateau near 1000 steps may be an artifact of pixel budget rather than an intrinsic LVM capacity limit; models that adaptively patch or tile long series could recover the lost context.
  • The decoder-dominance result suggests masked-image modeling's pixel-reconstruction objective, rather than image semantics, is the active ingredient; pretraining a small ViT on synthetic time-series-like images without natural-image semantics could replicate the forecasting gains and isolate the mechanism.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper presents a large-scale empirical study of whether image-pretrained large vision models (LVMs) are useful for time series classification (TSC) and forecasting (TSF). It compares 4 LVMs (ViT, Swin, MAE, SimMIM), 8 imaging methods, 18 datasets, and 26 baselines. The main findings are that GAF-imaged ViT/MAE achieve state-of-the-art TSC accuracy; that self-supervised LVMs with UVH imaging and reconstruction decoders are competitive for TSF on strongly periodic datasets but are limited to specific LVM/imaging combinations; that they exhibit a bias toward forecasting with the FFT-derived period L; and that they cannot effectively use look-back windows beyond about 1000 steps. Extensive ablations (RQ1–RQ9) probe the contributions of pre-training, architecture, imaging method, fine-tuning strategy, and temporal order.

Significance. If the findings hold, this is the first systematic evaluation of LVMs across both high-level and low-level time series tasks, and it provides actionable guidance for practitioners (e.g., use GAF for classification, use self-supervised LVMs with UVH and pre-trained decoders for forecasting periodic series). The paper is unusually thorough in scope—eight imaging methods, four LVM families, and two tasks—and the LVM results are repeated three times with standard deviations reported (Tables 10 and 12), which is a strength. The claim that the best LVM forecasters exhibit a period bias is an important cautionary result for the increasingly popular 'time series as images' approach. However, the forecasting-specific conclusions rest on experiments whose confounds need to be resolved before the key claims are fully established.

major comments (3)
  1. [§4.4, Figs. 8/15, with §3 Input Alignment] The RQ8 period-bias experiment is confounded by the fixed-size image resizing step. UVH produces an image of size L × ⌊T/L⌋, and the input alignment described in §3 then resizes every image to 224×224 (MAE) or 192×192 (SimMIM) using bilinear interpolation. For ETTh1 with L=24 and a ~432-step window, the native image is only about 24×18 pixels; bilinear upsampling expands each source pixel into a roughly 10×12 patch. Varying the segment length changes both the aspect ratio and the phase alignment of rows: at L or 2L the rows are phase-aligned and the resized image is horizontally striped, while at non-multiple lengths adjacent rows start at different phases, creating diagonal discontinuities that bilinear interpolation blurs. Any reconstruction model, not necessarily an LVM, would be expected to find the phase-aligned case easier. The n-values from Lemma 1 are computed from the same assumed period L, so the 'M-shape' in Fig. 8 cannot be attributed to LVM-specific behavior without a control. I request a control experiment using a non-LVM model (e.g., the W/O-LVM or LVM2ATTN ablation from RQ4) on the same UVH imaging and resizing pipeline; if a similar M-shape appears, the period-bias claim in the abstract and §4.4 is not established.
  2. [§4.2, Table 2 and Fig. 2] The headline comparisons against non-LVM baselines are not supported by significance testing. In Table 2, MAE outperforms the best non-LVM baselines by small margins (e.g., ETTh1 average MSE 0.409 vs. Time-LLM 0.418 and GPT4TS 0.418), and the baselines are point estimates reproduced from other papers without any reported variance. The TSC margin is also narrow: ViT's 75.3% average accuracy is only 1.3 percentage points above GPT4TS's 74.0%, while ViT's per-dataset standard deviations in Table 10 range up to 1.5 points. Given that the central claim is that 'LVMs are useful for time series,' a McNemar test for classification and a paired test (or at least a comparison of repeated-run distributions) for forecasting should be reported. Without this, the superiority claims are fragile, particularly on forecasting where the model ranking reverses across prediction horizons in several datasets (e.g., Table 11, ETTh1 720-step row).
  3. [§4.4 RQ9 and Table 24] The conclusion that LVMs have 'limited ability to utilize long look-back windows' is itself an artifact of the fixed-size input requirement, as the authors acknowledge in the RQ9 discussion. Because every imaged time series is uniformly resized to the LVM's fixed input size, longer look-back windows are compressed to the same number of pixels, which distorts pixel values and loses temporal detail. This is a property of the LVM input interface, not necessarily of the LVM's sequence modeling capacity. The paper should either separate these two factors (e.g., by using a resize-free patchification or a variable-resolution input) or explicitly reframe RQ9 as measuring the input-alignment bottleneck rather than the LVM's ability to use long context. As written, the abstract's claim of 'limited ability to utilize long look-back windows' conflates the two.
minor comments (5)
  1. [Abstract and throughout] The abstract contains the typo 'truely' (should be 'truly'); please proofread the entire manuscript for similar spelling errors (e.g., 'Illiness' in Table 7 and elsewhere should be 'Illness'; 'greated' in Appendix C should be 'greatest'; 'mult-head' in Appendix B.6 should be 'multi-head'; 'Refromer' in Appendix A.2 should be 'Reformer').
  2. [Table 2 and Fig. 2] The notation '# Wins' is ambiguous: for Table 2 it is reported as 9 for MAE, but the text says MAE outperforms non-LVM baselines in 9/16 cases. Please clarify whether wins are counted per dataset-metric cell, per dataset, or per dataset-prediction length, and ensure the caption and text are consistent.
  3. [Fig. 4 (CD diagrams)] The critical difference diagrams in Fig. 4 would benefit from explicit captions stating which statistical test and significance level were used to draw the CD bars; as printed, the bars are not self-explanatory.
  4. [§4.4 RQ8, Fig. 8] The green 'n (hypothesis)' line is not defined in the figure caption; please state that it is the normalized n from Lemma 1, and clarify how the min-max normalization was applied so readers can reproduce the overlay.
  5. [Appendix A.5, Table 8] The selection of the segment length L uses FFT on a fixed 1152-step window, but the main experiments in Table 2 use a look-back of 336 steps. Please state explicitly whether L is recomputed per look-back window or fixed globally; if fixed, explain the rationale, since this affects the interpretation of the UVH results at different horizons.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: headline results are external-benchmark comparisons; only a minor non-load-bearing self-citation to the authors' survey is present.

full rationale

The central TSC and TSF conclusions rest on comparisons against external baselines (PatchTST, GPT4TS, TimesNet, etc.) on standard benchmarks, with LVMs fine-tuned and evaluated under standard protocols, so the main claims are not circular. The only notable self-citation is reference [40], a survey by overlapping authors, cited for the 8 imaging methods and for pointers to the LVM-for-time-series literature; however, the imaging methods themselves are attributable to independent sources (GAF to Wang and Oates [48], UVH to TimesNet [52] and VisionTS [5]), and the citation supplies no uniqueness theorem or ansatz that forces the paper's conclusions. RQ8's 'n (hypothesis)' is computed from the same FFT-derived period L that defines the UVH segment grid, and the fixed-size bilinear resizing of Section 3 means changing the segment length also changes image geometry; this is a real confound for the 'forecasting-period bias' conclusion. But it is a validity threat, not a circular reduction: the MSE values are genuine model outputs, n is a parameter-free number-theoretic quantity from Lemma 1, and the M-shape agreement is an empirical correlation rather than an equation that equates the hypothesis to the result. No fitted parameter is renamed as a prediction, and no Eq. X = Eq. Y by construction is exhibited, so under the hard-evidence rule no circular step is established.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on three families of hidden choices: FFT-selected segment lengths that make the imaging method periodic-aligned; post hoc selection of imaging methods and fine-tuning strategies; and reliance on baseline numbers from prior papers rather than independent re-runs. No new physical or architectural entities are introduced.

free parameters (3)
  • UVH segment length L (per dataset) = 24 (ETTh1, ETTh2, Traffic, Electricity), 96 (ETTm1, ETTm2), 144 (Weather), 52 (Illness)
    Chosen via FFT as the dominant period of the look-back window (Table 8). The forecasting results and the RQ8 period-bias analysis directly depend on this choice; a different L would change which segment lengths are 'optimal'.
  • Best imaging method per task = GAF for TSC, UVH for TSF
    Selected from the results on the same benchmark test sets (Section 4.2), so the headline LVM numbers reflect the best-case imaging method, not an average across imaging methods. This is post hoc model selection.
  • Fine-tuning strategy for TSF = norm-only (d)
    The conclusion that LVMs are useful for forecasting is based on the best-performing fine-tuning scheme (Table 18); other schemes underperform or match training from scratch.
assumptions (5)
  • domain assumption FFT-based period detection on the look-back window identifies the true period(s) relevant for forecasting.
    Used to set UVH segment length L (Section A.5, Table 8) and to build the RQ8 difficulty hypothesis. If the FFT peak does not reflect the actual periodicity of the process, the period-bias conclusion breaks.
  • domain assumption ImageNet-pretrained LVMs can be repurposed for time series by treating imaged series as images; the patch and attention machinery transfers.
    This is the premise of the whole study (Section 3), supported empirically but assumed when interpreting results as evidence about LVMs rather than about the imaging pipeline.
  • domain assumption Baseline results reproduced from prior papers (Zeng et al., Tan et al., Chen et al.) are directly comparable to the LVM results.
    Used throughout Section 4; the paper does not rerun baselines, so any protocol mismatch affects all comparisons.
  • domain assumption Variate independence: each channel forecasted separately is adequate for multivariate forecasting.
    Adopted in Section 3 for TSF frameworks; a channel-interaction-aware model might change the relative ranking.
  • standard math Lemma 1: standard divisibility and arithmetic (k divides n*i implies k' divides n).
    Proof in Appendix C is correct for perfectly periodic series; its application to real data requires the periodicity assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Images to Signals: Are Large Vision Models Useful for Time Series Analysis?." pith.science (2026). https://pith.science/paper/TVEY5INM

@misc{pith2026250524030,
  author       = {Pith},
  title        = {Pith review of: From Images to Signals: Are Large Vision Models Useful for Time Series Analysis?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TVEY5INM}},
  note         = {Machine review of arXiv:2505.24030}
}
read the original abstract

Transformer-based models have gained increasing attention in time series research, driving interest in Large Language Models (LLMs) and foundation models for time series analysis. As the field moves toward multi-modality, Large Vision Models (LVMs) are emerging as a promising direction. In the past, the effectiveness of Transformer and LLMs in time series has been debated. When it comes to LVMs, a similar question arises: are LVMs truely useful for time series analysis? To address it, we design and conduct the first principled study involving 4 LVMs, 8 imaging methods, 18 datasets and 26 baselines across both high-level (classification) and low-level (forecasting) tasks, with extensive ablation analysis. Our findings indicate LVMs are indeed useful for time series classification but face challenges in forecasting. Although effective, the contemporary best LVM forecasters are limited to specific types of LVMs and imaging methods, exhibit a bias toward forecasting periods, and have limited ability to utilize long look-back windows. We hope our findings could serve as a cornerstone for future research on LVM- and multimodal-based solutions to different time series tasks.

Figures

Figures reproduced from arXiv: 2505.24030 by the authors.

Figure 1
Figure 1. An overview of (a) different imaging methods, (b) LVM-based time series classification, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Model comparison in TSC. The re￾sults are averaged over 10 UEA datasets. See [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of 4 LVMs on TSC (accuracy) and TSF (MSE). [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: Average rank of different imaging methods in (a) TSC [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Key components in LVMs’ Transformer block [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Inference time vs. performance of compared methods on TSC (accuracy) using UWaveGes￾ture, SpokenArabicDigits, and TSF (MSE) using ETTh1, Weather. Full results are in Appendix B.8. 15.5% 9.7% 6.4% 18.1% 1.6% 1.4% 4.8% 11.3% (a) MAE ETTh1 ETTm1 Illness Weather 15.9% 11.0…
Figure 8
Figure 8. Figure 8: Forecasting performance of MAE w.r.t. varying segment length used in UVH imaging. n (green) estimates the difficulty of forecasting. estimated by the average runtime per test sample [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 10
Figure 10. Figure 10: TSF performance (MSE) of MAE with varying look-back window (or context) lengths. [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 9
Figure 9. Figure 9: An illustration of UVH. Interestingly, following the UVH imaging method, we can estimate the difficulty of TSF for MAE by using the segment length. Basically, the difficulty highly correlates with how long a segment can reoccur, measured by the number of segments betwe…
Figure 11
Figure 11. Figure 11: Illustration of LVM’s ablation models. (a) is the model [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: presents the accuracy and inference efficiency comparison between LVMs and the two best-performing baselines on TSC task [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: MSE vs. inference time of the compared methods on TSF benchmark datasets. Green marker stands for LVM, Red marker stands for PatchTST and Blue marker stands for GPT4TS. 0 2 4 0.380 0.410 MAE ETTh1 0 3 6 0.330 0.345 ETTm1 0.5 1.0 Inference Time (ms) 0.187 0.203 MAE Wea…
Figure 14
Figure 14. Figure 14: MAE vs. inference time of the compared methods on TSF benchmark datasets. Green marker stands for LVM, Red marker stands for PatchTST and Blue marker stands for GPT4TS. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_14.png]
Figure 15
Figure 15. Figure 15: provides the forecasting performance of an LVM (i.e., MAE) in terms of metrics MAE w.r.t. segment length that varies from 1 6 L to 12 6 L. The LVM generally achieves the best performance when segment length is a multiple of the period, i.e. L or 2L, which is caused by…
Figure 16
Figure 16. Figure 16: Examples of GAF images on the first channel of multivariate time series with 152 time [PITH_FULL_IMAGE:figures/full_fig_p028_16.png]
Figure 17
Figure 17. Figure 17: Examples of GAF images on the first channel of multivariate time series with 336 time [PITH_FULL_IMAGE:figures/full_fig_p029_17.png]
Figure 18
Figure 18. Figure 18: Visualization of LVM’s inductive bias during TSF on a random example from the ETTh1 [PITH_FULL_IMAGE:figures/full_fig_p030_18.png]
Figure 19
Figure 19. Figure 19: Visualization of LVM’s inductive bias during TSF on a random example from the Traffic [PITH_FULL_IMAGE:figures/full_fig_p030_19.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. VAN-AD: Visual Masked Autoencoder with Normalizing Flow For Time Series Anomaly Detection

    cs.LG 2026-03 unverdicted novelty 6.0 of 10

    VAN-AD adapts a pretrained visual MAE with distribution mapping and normalizing flow modules to detect anomalies in time series data more effectively across different datasets.

  2. CLIPTime: Time-Aware Multimodal Representation Learning from Images and Text

    cs.CV 2025-08 reject novelty 2.0 of 10

    CLIPTime adds a classification head and a transformer-style regression head to CLIP embeddings, hitting 98.7% accuracy on synthetic fungi but with weak timestamp predictions, especially for spores.

Reference graph

Works this paper leans on

68 extracted references · 52 canonical work pages · cited by 2 Pith papers

  1. [1]

    Ardid, D

    A. Ardid, D. Dempsey, C. Caudron, S. Cronin, B. Kennedy, T. Girona, D. Roman, C. Miller, S. Potter, O. D. Lamb, et al. Ergodic seismic precursors and transfer learning for short term eruption forecasting at data scarce volcanoes. Nat. Commun., 16(1):1758, 2025

    work page 2025

  2. [2]

    Bagnall, H

    A. Bagnall, H. A. Dau, J. Lines, M. Flynn, J. Large, A. Bostrom, P. Southam, and E. Keogh. The uea multivariate time series classification archive, 2018. arXiv preprint arXiv:1811.00075, 2018

    arXiv 2018

  3. [3]

    H. Bao, L. Dong, S. Piao, et al. Beit: Bert pre-training of image transformers. In ICLR, 2022

    work page 2022

  4. [4]

    J. O. Caro, A. H. de Oliveira Fonseca, S. A. Rizvi, M. Rosati, C. Averill, J. L. Cross, P. Mittal, E. Zappala, R. M. Dhodapkar, C. Abdallah, et al. Brainlm: A foundation model for brain activity recordings. In ICLR, 2024

    work page 2024

  5. [5]

    M. Chen, L. Shen, Z. Li, X. J. Wang, J. Sun, and C. Liu. VisionTS: Visual masked autoencoders are free-lunch zero-shot time series forecasters. In ICML, 2025

    work page 2025

  6. [6]

    Chen and C

    T. Chen and C. Guestrin. XGboost: A scalable tree boosting system. In SIGKDD, 2016

    work page 2016

  7. [7]

    Dempster, F

    A. Dempster, F. Petitjean, and G. I. Webb. ROCKET: Exceptionally fast and accurate time series classification using random convolutional kernels. Data Min. Knowl., 34(5):1454–1495, 2020

    work page 2020

  8. [8]

    Z. Dong, R. Li, Y . Wu, T. T. Nguyen, J. Chong, F. Ji, N. Tong, C. Chen, and J. H. Zhou. Brain-JEPA: Brain dynamics foundation model with gradient positioning and spatiotemporal masking. In NeurIPS, 2024

    work page 2024

Show all 68 references
  1. [9]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021

    2021

  2. [10]

    Franceschi, A

    J.-Y . Franceschi, A. Dieuleveut, and M. Jaggi. Unsupervised scalable representation learning for multivariate time series. In NeurIPS, 2019

    2019

  3. [11]

    Gong, Y .-A

    Y . Gong, Y .-A. Chung, and J. Glass. AST: Audio spectrogram transformer.Interspeech, 2021

    2021

  4. [12]

    Gong, C.-I

    Y . Gong, C.-I. Lai, Y .-A. Chung, and J. Glass. SSAST: Self-supervised audio spectrogram transformer. In AAAI, 2022

    2022

  5. [13]

    A. Gu, K. Goel, and C. Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021

    2021 arXiv

  6. [14]

    S. Han, X. Hu, H. Huang, M. Jiang, and Y . Zhao. ADBench: Anomaly detection benchmark. In NeurIPS, 2022

    2022

  7. [15]

    K. He, X. Chen, S. Xie, Y . Li, P. Dollár, and R. Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022

    2022

  8. [16]

    Huang, X

    Z. Huang, X. Shi, C. Zhang, Q. Wang, K. C. Cheung, H. Qin, J. Dai, and H. Li. FlowFormer: A transformer architecture for optical flow. In ECCV, 2022

    2022

  9. [17]

    Jiang, K

    Y . Jiang, K. Ning, Z. Pan, X. Shen, J. Ni, W. Yu, A. Schneider, H. Chen, Y . Nevmyvaka, and D. Song. Multi-modal time series analysis: A tutorial and survey. arXiv preprint arXiv:2503.13709, 2025

    2025 arXiv

  10. [18]

    Jiang, Z

    Y . Jiang, Z. Pan, X. Zhang, S. Garg, A. Schneider, Y . Nevmyvaka, and D. Song. Empowering time series analysis with large language models: a survey. In IJCAI, 2024. 10

    2024

  11. [19]

    M. Jin, S. Wang, L. Ma, Z. Chu, J. Y . Zhang, X. Shi, P.-Y . Chen, Y . Liang, Y .-F. Li, S. Pan, et al. Time-LLM: Time series forecasting by reprogramming large language models. arXiv preprint arXiv:2310.01728, 2023

    2023 arXiv

  12. [20]

    M. Jin, Y . Zhang, W. Chen, K. Zhang, Y . Liang, B. Yang, J. Wang, S. Pan, and Q. Wen. Position: What can large language models tell us about time series analysis. In ICML, 2024

    2024

  13. [21]

    Karami, M.-A

    H. Karami, M.-A. Hartley, D. Atienza, and A. Ionescu. TimEHR: Image-based time series generation for electronic health records. arXiv preprint arXiv:2402.06318, 2024

    2024 arXiv

  14. [22]

    W. Kim, B. Son, and I. Kim. ViLT: Vision-and-language transformer without convolution or region supervision. In ICML, 2021

    2021

  15. [23]

    Kitaev, Ł

    N. Kitaev, Ł. Kaiser, and A. Levskaya. Reformer: The efficient transformer. arXiv preprint arXiv:2001.04451, 2020

    2001 arXiv

  16. [24]

    Koprinska, D

    I. Koprinska, D. Wu, and Z. Wang. Convolutional neural networks for energy time series forecasting. In IJCNN, 2018

    2018

  17. [25]

    Lai, W.-C

    G. Lai, W.-C. Chang, Y . Yang, and H. Liu. Modeling long-and short-term temporal patterns with deep neural networks. In SIGIR, 2018

    2018

  18. [26]

    Lenc and A

    K. Lenc and A. Vedaldi. Understanding image representations by measuring their equivariance and equivalence. In CVPR, 2015

    2015

  19. [27]

    X. Li, Y . Kang, and F. Li. Forecasting with time series imaging.Expert Syst. Appl., 160:113680, 2020

    2020

  20. [28]

    X. Li, V . Metsis, H. Wang, and A. H. H. Ngu. TTS-GAN: A transformer-based time-series generative adversarial network. In AIME, 2022

    2022

  21. [29]

    Z. Li, S. Li, and X. Yan. Time series as images: Vision transformer for irregularly sampled time series. In NeurIPS, 2023

    2023

  22. [30]

    S. Lin, W. Lin, W. Wu, H. Chen, and J. Yang. SparseTSF: Modeling long-term time series forecasting with 1k parameters. In ICML, 2024

    2024

  23. [31]

    H. Liu, H. Kamarthi, Z. Zhao, S. Xu, S. Wang, Q. Wen, T. Hartvigsen, F. Wang, and B. A. Prakash. How can time series analysis benefit from multiple modalities? A survey and outlook. arXiv preprint arXiv:2503.11835, 2025

    2025

  24. [32]

    H. Liu, C. Li, Q. Wu, and Y . J. Lee. Visual instruction tuning. In NeurIPS, 2023

    2023

  25. [33]

    P. Liu, H. Guo, T. Dai, N. Li, J. Bao, X. Ren, Y . Jiang, and S.-T. Xia. CALF: Aligning LLMs for time series forecasting via cross-modal fine-tuning. In AAAI, 2025

    2025

  26. [34]

    S. Liu, H. Yu, C. Liao, J. Li, W. Lin, A. X. Liu, and S. Dustdar. Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting. In ICLR, 2022

    2022

  27. [35]

    Y . Liu, H. Wu, J. Wang, and M. Long. Non-stationary transformers: Exploring the stationarity in time series forecasting. In NeurIPS, 2022

    2022

  28. [36]

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021

    2021

  29. [37]

    X. Ma, Z. Dai, Z. He, J. Ma, Y . Wang, and Y . Wang. Learning traffic as images: A deep convolutional neural network for large-scale transportation network speed prediction. Sensors, 17(4):818, 2017

    2017

  30. [38]

    M. A. Morid, O. R. L. Sheng, and J. Dunbar. Time series prediction using deep learning methods in healthcare. ACM TMIS, 14(1):1–29, 2023

    2023

  31. [39]

    Namura, Y

    N. Namura, Y . Ichikawa, et al. Training-free time-series anomaly detection: Leveraging image foundation models. arXiv preprint arXiv:2408.14756, 2024. 11

    2024 arXiv

  32. [40]

    J. Ni, Z. Zhao, C. Shen, H. Tong, D. Song, W. Cheng, D. Luo, and H. Chen. Harnessing vision models for time series analysis: A survey. In IJCAI, 2025

    2025

  33. [41]

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. In ICLR, 2023

    2023

  34. [42]

    Prithyani, M

    V . Prithyani, M. Mohammed, R. Gadgil, R. Buitrago, V . Jain, and A. Chadha. On the feasibility of vision-language models for time-series classification. arXiv:2412.17304, 2024

    2024 arXiv

  35. [43]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021

    2021

  36. [44]

    M. Tan, M. Merrill, V . Gupta, T. Althoff, and T. Hartvigsen. Are language models actually useful for time series forecasting? In NeurIPS, 2024

    2024

  37. [45]

    Touvron, M

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021

    2021

  38. [46]

    Trindade

    A. Trindade. Electricity Load Diagrams. UCI Machine Learning Repository, 2015

    2015

  39. [47]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need. In NeurIPS, 2017

    2017

  40. [48]

    Wang and T

    Z. Wang and T. Oates. Imaging time-series to improve classification and imputation. In IJCAI, 2015

    2015

  41. [49]

    Q. Wen, T. Zhou, C. Zhang, W. Chen, Z. Ma, J. Yan, and L. Sun. Transformers in time series: A survey. In IJCAI, 2023

    2023

  42. [50]

    Wimmer and N

    C. Wimmer and N. Rekabsaz. Leveraging vision-language models for granular market change prediction. arXiv:2301.10166, 2023

    2023 arXiv

  43. [51]

    G. Woo, C. Liu, D. Sahoo, A. Kumar, and S. Hoi. ETSformer: Exponential smoothing transformers for time-series forecasting. arXiv preprint arXiv:2202.01381, 2022

    2022 arXiv

  44. [52]

    H. Wu, T. Hu, Y . Liu, H. Zhou, J. Wang, and M. Long. TimesNet: Temporal 2d-variation modeling for general time series analysis. In ICLR, 2023

    2023

  45. [53]

    H. Wu, J. Xu, J. Wang, and M. Long. Autoformer: Decomposition transformers with auto- correlation for long-term series forecasting. In NeurIPS, 2021

    2021

  46. [54]

    Z. Xie, Z. Zhang, Y . Cao, Y . Lin, J. Bao, Z. Yao, Q. Dai, and H. Hu. SimMIM: A simple framework for masked image modeling. In CVPR, 2022

    2022

  47. [55]

    L. Yang, Y . Wang, X. Fan, I. Cohen, J. Chen, Y . Zhao, and Z. Zhang. ViTime: A visual intelligence-based foundation model for time series forecasting. arXiv preprint arXiv:2407.07311, 2024

    2024

  48. [56]

    A. Zeng, M. Chen, L. Zhang, and Q. Xu. Are transformers effective for time series forecasting? In AAAI, 2023

    2023

  49. [57]

    Z. Zeng, R. Kaur, S. Siddagangappa, et al. From pixels to predictions: Spectrogram and vision transformer for better time series forecasting. In ICAIF, 2023

    2023

  50. [58]

    Zerveas, S

    G. Zerveas, S. Jayaraman, D. Patel, A. Bhamidipaty, and C. Eickhoff. A transformer-based framework for multivariate time series representation learning. In SIGKDD, 2021

    2021

  51. [59]

    Zhang, D

    C. Zhang, D. Song, Y . Chen, X. Feng, C. Lumezanu, W. Cheng, J. Ni, B. Zong, H. Chen, and N. V . Chawla. A deep neural network for unsupervised anomaly detection and diagnosis in multivariate time series data. In AAAI, 2019

    2019

  52. [60]

    Zhang, Y

    T. Zhang, Y . Zhang, W. Cao, J. Bian, X. Yi, S. Zheng, and J. Li. Less is more: Fast multivariate time series forecasting with light sampling-oriented MLP structures. arxiv. arXiv preprint arXiv:2207.01186, 2022. 12

    2022 arXiv

  53. [61]

    Zhang, R

    X. Zhang, R. R. Chowdhury, R. K. Gupta, and J. Shang. Large language models for time series: A survey. In IJCAI, 2024

    2024

  54. [62]

    Zhong, W

    S. Zhong, W. Ruan, M. Jin, H. Li, Q. Wen, and Y . Liang. Time-VLM: Exploring multimodal vision-language models for augmented time series forecasting.arXiv preprint arXiv:2502.04395, 2025

    2025 arXiv

  55. [63]

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In AAAI, 2021

    2021

  56. [64]

    T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin. FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting. In ICML, 2022

    2022

  57. [65]

    T. Zhou, P. Niu, L. Sun, R. Jin, et al. One fits all: Power general time series analysis by pretrained LM. In NeurIPS, 2023

    2023

  58. [66]

    Zhou and R

    Z. Zhou and R. Yu. Can LLMs understand time series anomalies? In ICLR, 2025

    2025

  59. [67]

    Dataset Size

    J. Zhuang, L. Yan, Z. Zhang, R. Wang, J. Zhang, and Y . Gu. See it, think it, sorted: Large multi- modal models are few-shot time series anomaly analyzers. arXiv preprint arXiv:2411.02465, 2024. 13 A Experimental Setup A.1 Benchmarks Time Series Classification. For TSC, follow...

    2024 arXiv

  60. [68]

    forecasting periods

    The training runs up to a maximum of 30 epochs on the training set. Early stopping is applied after 8 consecutive epochs of no improvement is observed on the validation set. For TSF task, we use Adam optimizer with learning rate 0.0001. For ETT and Illness datasets, the batch ...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.