Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Self-Supervised Dynamical System Representations for Physiological Time-Series

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The paper claims that self-supervised pretraining for physiological time series should reconstruct one window from the system parameters of another, discarding sample-specific noise, and proves when this recovers exactly the shared system p

desk verdict A solid empirical paper whose theory covers an oracle setting rather than the actual objective; the authors are candid about the gap, and the results are strong enough to warrant a real review. read the letter →

arxiv 2512.00239 v2 pith:FKQGJSHN submitted 2025-11-28 cs.LG stat.ML

Yenho Chen , Maxwell A. Xu , James M. Rehg , Christopher J. Rozell This is my paper
classification cs.LGstat.ML
keywords self-supervisedlearningphysiologicaltimeseriesdynamicalsystemscross-reconstructionsystemparametersrepresentationmaskedautoencodingtransfer
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

The paper argues that for physiological time series, the right self-supervised objective is not a heuristic augmentation but a task derived from how the data are generated: each recording is a dynamical system with shared parameters (the underlying physiological state) and sample-specific details (initial conditions, noise). PULSE pretrains by reconstructing a randomly cropped window from the system parameters of the same window, forcing the encoder to keep only shared information. The theory shows that if one of two same-system samples is fully masked, the minimal set of shared latent variables is exactly the system parameters; partial masking would also recover state variables and confound the representation. On synthetic chaotic systems and four real physiological datasets, the resulting representations separate classes, work with 1% labels, and transfer to related tasks better than contrastive and masked-autoencoding baselines. If correct, this gives a principled alternative to heuristic self-supervised learning for physiological signals.

What carries the argument

The load-bearing object is the cross-reconstruction objective L_PULSE: a system encoder extracts latent parameters Theta_i from a window, an initial-condition encoder estimates a starting state from a randomly chosen crop, and a GRU-plus-linear decoder reconstructs the cropped segment from those two quantities. Because the crop varies randomly, the encoder cannot store initial-condition or noise information without hurting reconstruction, so the only stable information is the system parameters. The theorem is derived by viewing cross-reconstruction as masked autoencoding on the generative DAG; whole-sample masking leaves Theta(s) as the sole shared ancestor of input and target.

What would settle it

Take a known dynamical system with fixed system parameters and generate windows with increasingly large spreads of initial conditions and increasingly strong nonstationary drift. If PULSE's classification accuracy on the system parameters collapses toward chance while cross-reconstruction with true independent same-system pairs stays high, the pseudo-pair approximation is the load-bearing link.

Watch

Extended reading notes

Core claim

The central claim is Theorem 1: under a hierarchical generative model where each time series is produced by latent dynamics with shared system parameters Theta(s), per-sample initial conditions, and noise, the minimal set of latent variables shared between two independently sampled same-system series is Theta(s) if and only if the entire observable of one series is masked. Cross-reconstruction with whole-sample masking therefore recovers exactly the transferable system information, while partial masking also recovers boundary state variables and mixes in sample-specific content. PULSE instantiates this with a system encoder, an initial-condition encoder, and a recurrent-state decoder, using

Load-bearing premise

The paper itself concedes (Appendix J) that PULSE's random-crop pseudo-pairs are not the independent same-system samples that Theorem 1 requires; if that approximation fails when sample-specific variability or nonstationarity is large, the theory no longer backs the method and only the empirical benchmarks remain.

Editorial extensions

If this is right

  • The same pretrained representation linearly separates clinical classes on ECG, PPG, and EEG without any labels during pretraining.
  • With only 1% or 5% labeled samples, PULSE outperforms contrastive and masked-autoencoding baselines, and most supervised baselines, on all four datasets.
  • Pretrained PULSE representations transfer in-domain (EEG to epilepsy, activity to gesture) better than the baselines.
  • The synthetic experiments validate the theorem's prediction: an oracle using true same-system pairs beats an oracle using random temporal masking, with the gap growing as dynamical noise increases.
  • Partial masking strategies, including standard patch masking, recover boundary state variables in theory, which explains why heuristic masked autoencoding can mix sample-specific content into representations.

Reading between the lines

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

  • The pseudo-pair principle generalizes: any self-supervised objective that pairs a sample with a randomly reinitialized reconstruction target may act as a transferable-information bottleneck, not just for biosignals but for any time series with stereotyped dynamics.
  • The theory implies a practical diagnostic: when a physiological dataset has strong nonstationarity or large between-sample variability, random-crop pseudo-pairs should be replaced by better estimates of independent same-system samples; the paper's Appendix J predicts performance should degrade first in that regime.
  • If the results replicate across more clinical datasets, pretraining on raw unlabeled monitoring data could reduce annotation needs for tasks like arrhythmia and sleep staging, since the objective does not depend on task labels or hand-designed augmentations.
  • The masking theorem suggests a design rule for masked modeling more broadly: to isolate system-level information, mask entire samples rather than patches, because patch masking theoretically recovers state variables at mask boundaries.
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

4 major / 5 minor

Summary. The paper proposes PULSE, a self-supervised pretraining objective for physiological time series. The authors model a dataset with a hierarchical dynamical-systems generative model in which samples from the same latent system share system parameters Θ^(s), while initial conditions and noise are sample-specific. They argue that useful representations should preserve Θ^(s) and discard sample-specific information, and they instantiate this via a cross-reconstruction objective (L_Cross, Eq. 2) with an associated system encoder f_sys and initial-condition encoder f_init. Because class labels for similar samples are unavailable, the practical objective L_PULSE (Eq. 3) constructs pseudo-pairs from random crops of the same window. The paper proves Theorem 1 characterizing when the minimal latent set shared between masked and unmasked observables equals Θ^(s), claims this supports L_Cross and, by approximation, PULSE, validates the qualitative prediction on synthetic Lorenz/Thomas/Hindmarsh-Rose systems, and reports strong empirical results on HAR, PPG, ECG, and EEG for linear probing, semi-supervised learning, and transfer.

Significance. The empirical contribution is substantial: PULSE achieves consistent gains over recent contrastive, MAE, and sequential-VAE baselines across four diverse physiological datasets, with careful controls (shared encoder for CL and SVAE baselines, cross-validation, ablations, and transfer experiments). If the theoretical claim were fully supported, this would be a valuable step toward principled, generative-model-guided SSL for physiological signals. The paper's strengths include the breadth and consistency of the experimental evaluation, the inclusion of cross-validation results, and the synthetic experiments that test the theory's qualitative predictions. However, as detailed below, the theory as written does not cover the actual PULSE objective, and even the oracle L_Cross is not exactly the full-masking scheme analyzed in Theorem 1 because the target series is observed by f_init. The central conceptual claim is therefore not yet established, and the paper currently overstates what the theorem proves. The contribution remains defensible as a heuristic method with strong empirical support, but the theoretical framing needs substantial repair or reframing.

major comments (4)
  1. [§3.2, Eq. (2); §3.3] The paper states that L_Cross 'corresponds exactly' to the full-sample masking scheme of Theorem 1, with m_j=0. But Eq. (2) defines X_{j,t0} = [f_init(Y_j)]_{t0}, so Y_j is observed by f_init and is not fully masked. The masking view in §3.3 removes Y_j from the input, yet the objective uses Y_j to infer the initial condition. Therefore Theorem 1, which characterizes the shared latent variables between fully masked and fully observed observables in the graphical model, does not apply to L_Cross as written. The claimed exact correspondence is contradicted by the definition of f_init. This is load-bearing because it undermines the theoretical support for even the oracle objective. Please either reformulate L_Cross so the target is not observed by any encoder, or extend the theory to cover the two-encoder setting where the target contributes initial-condition information.
  2. [§3.2, Eq. (3); Appendix I; Appendix J] Theorem 1 is proved for independent samples (Y_i, Y_j) with one sample fully masked, but L_PULSE uses pseudo-pairs (Y_i, Y_{i,t0:W}) where the target is a deterministic subset of the input. Since f_sys observes the full input including the target region, it can copy sample-specific local signal information into θ_{i,t_k}; the one-dimensional constraint and adaptive pooling limit capacity but do not provide an information-theoretic guarantee that shared system parameters are recovered. Appendix I argues by analogy ('PULSE is designed to emulate'), and Appendix J explicitly concedes 'the gap between the practical PULSE algorithm and the theoretically ideal cross-reconstruction setting.' Thus Theorem 1 provides no formal support for the actual optimized objective. The central claim of the paper needs either a proof for the pseudo-pair objective, a quantitative leakage bound, or an explicit
  3. [Assumption 1 and Eq. (1)] The generative model in Eq. (1) includes additive observation noise ϵ_{i,t_k} and transition noise ν_{i,t_k}, so the functions g_y and g_x are stochastic and not invertible in the usual sense. Assumption 1(ii) requires each g_k to be invertible, and the proof of Theorem 1 invokes Theorem 1 of Kong and Zhang (2023), which is developed for deterministic invertible generative functions. The paper does not verify that the stochastic SSM satisfies the conditions of that theorem, nor does it state a modified definition of invertibility for noisy maps. This leaves a gap in the proof of Theorem 1. Please clarify the intended meaning of invertibility and how the result extends to the noisy setting used in the paper.
  4. [Theorem 1] The theorem is in large part a consequence of the graphical model: Θ^(s) is by definition the only latent variable shared across the two samples, so the 'if' direction follows directly from the definition of the generative process. The nontrivial content is in the partial-masking cases, which the proof handles correctly. However, the paper should avoid the stronger phrasing that the theorem shows PULSE 'recovers system information' during training. The theorem characterizes the minimal shared latent set in the generative model; it does not address whether a trainable encoder minimizing L_Cross or L_PULSE with finite data, capacity constraints, and nonconvex optimization will actually recover that set. The link between the information-theoretic characterization and the learned representation should be stated more cautiously.
minor comments (5)
  1. [§3.3, Theorem 1] The phrase 'all observables from one series is fully masked' is ambiguous. The parenthetical (m_i=0, m_j=1) clarifies that exactly one series is fully masked and the other fully observed, but the text should say this explicitly; as written, a reader could read it as 'there exists some fully masked series,' which is not sufficient (a fully masked Y_i plus a partially masked Y_j would not yield C={Θ^(s)}).
  2. [Table 2, §5.1] In the HAR row of Table 2, PULSE's accuracy (93.27) and AUPRC (98.10) are lower than REBAR (95.35 and 98.91). The text says PULSE is 'competitive' on HAR, but the abstract and introduction claim 'consistent performance improvements.' Please qualify the claim or point to the cross-validation results in Table 6, where PULSE is the top performer on HAR.
  3. [Eq. (3)] There is an unmatched parenthesis in the norm: ∥ eY_{i,t_k} − g_y(g_x(X_{i,t_{k-1}}, Θ_{i,t_k})∥^2 should be ∥ eY_{i,t_k} − g_y(g_x(X_{i,t_{k-1}}, Θ_{i,t_k}))∥^2.
  4. [§5.3, Table 4] The term 'In-Domain Transfer Learning' is confusing: the experiment transfers from EEG to Epilepsy and from HAR to Gesture, i.e., across datasets within the same modality. Consider renaming to 'cross-dataset transfer within sensor modality.'
  5. [References and Appendix E] There are several reference/citation artifacts: 'EN40214341417 Lorenz' in the bibliography, 'V on K¨ugelgen' with TeX markup, and inconsistent author naming for the DSVAE work (Li and Mandt vs. Yingzhen and Mandt). Please clean these up in the final version.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity: Theorem 1 is partly definitional (the recovery target is defined as the only cross-sample shared variable), but the masking analysis and benchmarks are independent.

  1. self definitional [Definition 3.1 (Sec. 3.1) and Theorem 1 (Sec. 3.3)]
    "Two time-series Y_i and Y_j are similar if they are generated by the same system parameters Θ^(s), such that both indices satisfy i, j∈ I_s. ... Theorem 1. ... the minimal set of latent variables shared is the system parameters Θ^(s) if and only if all observables from one series is fully masked (i.e., m_i,1:W,1:M = 0 and m_j,1:W,1:M = 1)."

    The target of recovery (Θ^(s)) is introduced in Eq. 1 / Def. 3.1 as the only latent variable shared across samples of the same system. The full-mask branch of Theorem 1 therefore restates the model's definition of 'shared': when one entire series is removed, the only common ancestor of the two observables is Θ^(s). This is a modeling convention rather than an independent empirical prediction. The partial-masking branch and the synthetic oracle experiment still carry independent content, so this is a mild definitional aspect, not a wholesale circularity.

full rationale

The derivation chain is largely self-contained. Theorem 1 applies an external graph-theoretic algorithm (Kong and Zhang 2023) to the paper's own generative model. The full-sample masking case essentially restates the model's definition: Θ^(s) is declared to be the only variable shared between samples in Eq. 1 / Def. 3.1, so recovering it under full masking is by construction. This is a definitional aspect of the theoretical claim, but it is not a fitted-input prediction: the partial-masking branch is a genuine path computation, and the synthetic experiment independently tests the prediction by comparing a label-informed positive oracle against a negative oracle (Table 1). The practical objective L_PULSE (Eq. 3) is not covered by Theorem 1 because it uses overlapping same-sample crops rather than independent pairs; the authors explicitly concede this in Appendix J ('the gap between the practical PULSE algorithm and the theoretically ideal cross-reconstruction setting'). That is a stated limitation, not a hidden circularity. No load-bearing self-citation or imported uniqueness theorem appears; Kong and Zhang's theorem is external, and the self-citations are background/baseline references. Real-data results are benchmarked against external SSL methods, so the empirical claims do not reduce to the paper's definitions.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a stated generative model (Eq. 1), a borrowed theorem (Kong & Zhang 2023), an explicit invertibility/DAG assumption, and — most importantly — an unproven heuristic bridge from oracle cross-reconstruction to the practical pseudo-pair objective. No new physical entities are introduced; the “system parameters” are latent variables of the proposed model, and their existence is supported by synthetic experiments with known systems.

free parameters (4)
  • time-varying system parameter dimension = 1
    Θ_{i,t_k}'s time-varying component is constrained to a single dimension to block trivial copying of local signal values (§3.2); the w/o TV-Params ablation (Table 5) shows the component is load-bearing.
  • initial-condition sampling range = t0 ~ Uniform(1, W/2)
    Chosen “to prevent overfitting to short subsequences” (§3.2); the w/o Sample-t0 ablation (fixed t0=1) shows the random range matters.
  • number of pseudo-pair crops per window = up to 4
    The authors report using up to four random samples to estimate the expectation in Eq. 3 (§3.2).
  • adaptive max-pooling span for θ_{i,t_k} = consecutive timesteps (size not specified in text)
    Temporal pooling is applied “to limit the temporal variability of θ_{i,k}” (§3.2); exact span deferred to the code repository, which is unavailable.
assumptions (5)
  • standard math Theorem 1 of Kong & Zhang (2023): MAE training recovers the minimal set of shared latent variables between masked and unmasked observables, identifiable by Algorithm 1.
    Used as the foundation of the proof of Theorem 1 (§3.3, Appendix A); assumed correct though not machine-checked.
  • domain assumption Assumption 1: the generative process of Eq. 1 is a DAG and every g_k is invertible.
    Stated in §3.3; needed for the Kong–Zhang machinery; violated by chaotic real systems, acknowledged and tested empirically.
  • domain assumption Real physiological time series are generated by a shared-parameter dynamical system per semantic class, independent of initial conditions and noise, as in Eq. 1.
    Load-bearing for defining which information counts as “system information”; assumed throughout §3.1 and validated only indirectly by downstream results.
  • ad hoc to paper Random-crop pseudo-pairs (Y_i, Y_{i,t0:W}) approximate independent samples from the same system for the purpose of cross-reconstruction.
    Bridges Theorem 1 (oracle L_Cross) to the practical objective L_PULSE; argued heuristically in Appendix I, no proof given, and the paper itself flags the gap in Appendix J.
  • standard math The GRU decoder and linear projection can represent the true dynamics g_x and observation function g_y given the Θ input.
    Universal-approximation-type assumption for the neural decoders, implicit in Eqs. 2–3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Supervised Dynamical System Representations for Physiological Time-Series." pith.science (2026). https://pith.science/paper/FKQGJSHN

@misc{pith2026251200239,
  author       = {Pith},
  title        = {Pith review of: Self-Supervised Dynamical System Representations for Physiological Time-Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FKQGJSHN}},
  note         = {Machine review of arXiv:2512.00239}
}
read the original abstract

The effectiveness of self-supervised learning (SSL) for physiological time series depends on the ability of a pretraining objective to preserve information about the underlying physiological state while filtering out unrelated noise. However, existing strategies are limited due to reliance on heuristic principles or poorly constrained generative tasks. To address this limitation, we propose a pretraining framework that exploits the information structure of a dynamical systems generative model across multiple time-series. This framework reveals our key insight that class identity can be efficiently captured by extracting information about the generative variables related to the system parameters shared across similar time series samples, while noise unique to individual samples should be discarded. Building on this insight, we propose PULSE, a cross-reconstruction-based pretraining objective for physiological time series datasets that explicitly extracts system information while discarding non-transferrable sample-specific ones. We establish theory that provides sufficient conditions for the system information to be recovered, and empirically validate it using a synthetic dynamical systems experiment. Furthermore, we apply our method to diverse real-world datasets, demonstrating that PULSE learns representations that can broadly distinguish semantic classes, increase label efficiency, and improve transfer learning.

Figures

Figures reproduced from arXiv: 2512.00239 by the authors.

Figure 1
Figure 1. Intuition behind PULSE. A dynamical systems model of a physiological time-series [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Our graphical model of multiple time￾series windows, based on dynamical systems, dis￾tinguishes transferable system information shared across similar time-series from non-transferable information unique to each sample such as initial conditions and process noise. Time-Series Dataset Generative Model. As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. PULSE aims to recover system infor￾mation through an inference process that uses two encoders, fsys to estimate shared parameters of a latent dynamical systems and finit to estimate sample-specific initial conditions. By requiring Θi to support reconstruction of randomly sampled Xi,t0 , we encourage the recovered system infor￾mation to be invariant to the sample-specific ones. However, initial conditions cannot be c… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: We illustrate how these different masking strategies recover different sources of in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: t-SNE visualizations of PULSE representations. [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Sonata: A Hybrid World Model for Inertial Kinematics under Clinical Data Scarcity

    cs.LG 2026-04 unverdicted novelty 6.0 of 10

    Sonata is a small hybrid world model pre-trained to predict future IMU states that outperforms autoregressive baselines on clinical discrimination, fall-risk prediction, and cross-cohort transfer while fitting on-devi...

Reference graph

Works this paper leans on

33 extracted references · 13 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Following prior work (Sparrow, 2012; Kamiya et al., 2024), we fixβ= 8/3ands= 28, and sweepρacross the following 10 values: {28,41,55,69,83,96,110,124,138,152}

    y1(ρ−y 3)−y 2 y1y2 −βy 3 # (8) wherey= [y 1, y2, y3]⊤. Following prior work (Sparrow, 2012; Kamiya et al., 2024), we fixβ= 8/3ands= 28, and sweepρacross the following 10 values: {28,41,55,69,83,96,110,124,138,152}. These values span a range of distinct chaotic regimes. Thomas (Thomas, 1999).The Thomas attractor is a 3D strange attractor that produces cyli...

    2012

  2. [2]

    These results indicate that PULSE is robust to within-dataset variability and performs consistently across datasets with diverse signal characteristics

    Notably, PULSE becomes the top performer on HAR when averaged over cross validation splits, even though it was not in the single-split setting. These results indicate that PULSE is robust to within-dataset variability and performs consistently across datasets with diverse signal characteristics. 20 Under review as a conference paper at ICLR 2026 Metric Si...

    2026

  3. [3]

    This consistency across evaluation settings indicates that PULSE’s representations are stable with respect to how the data is partitioned, which further strengthens our conclusion that PULSE can learn label-efficient representations that generalize across physiological datasets with very different signal and dataset characteristics. Dataset SupervisedSimC...

    2026

  4. [7]

    Med- vista3d: Vision-language modeling for reducing diagnostic errors in 3d ct disease detection, un- derstanding and reporting.arXiv preprint arXiv:2509.03800, 2025b

    Yuheng Li, Yenho Chen, Yuxiang Lai, Jike Zhong, Vanessa Wildman, and Xiaofeng Yang. Med- vista3d: Vision-language modeling for reducing diagnostic errors in 3d ct disease detection, un- derstanding and reporting.arXiv preprint arXiv:2509.03800, 2025b. Jiayang Liu, Lin Zhong, Jehan Wickramasuriya, and Venu Vasudevan. uwave: Accelerometer- based personalize...

    arXiv

  5. [9]

    Weakly-supervised disentanglement without compromises

    12 Under review as a conference paper at ICLR 2026 Francesco Locatello, Ben Poole, Gunnar R¨atsch, Bernhard Sch¨olkopf, Olivier Bachem, and Michael Tschannen. Weakly-supervised disentanglement without compromises. InInternational confer- ence on machine learning, pages 6348–6359. PMLR,

    2026

  6. [11]

    A review on the non- linear dynamical system analysis of electrocardiogram signal.Journal of healthcare engineering, 2018(1):6920420,

    Suraj K Nayak, Arindam Bit, Anilesh Dey, Biswajit Mohapatra, and Kunal Pal. A review on the non- linear dynamical system analysis of electrocardiogram signal.Journal of healthcare engineering, 2018(1):6920420,

    2018

  7. [13]

    Time-varying representations of longitudinal biosignals using self-supervised learning

    Sam Jean Perochon, Salar Abbaspourazad, Joseph Futoma, Andrew Miller, and Guillermo Sapiro. Time-varying representations of longitudinal biosignals using self-supervised learning. In NeurIPS 2024 Workshop: Self-Supervised Learning-Theory and Practice. Sebastian Raschka, Joshua Patterson, and Corey Nolet. Machine learning in python: Main develop- ments and...

    arXiv 2024

  8. [14]

    lfads-torch: A modular and extensible implementation of latent factor analysis via dynamical systems.arXiv preprint arXiv:2309.01230,

    Andrew R Sedler and Chethan Pandarinath. lfads-torch: A modular and extensible implementation of latent factor analysis via dynamical systems.arXiv preprint arXiv:2309.01230,

    arXiv

Show all 33 references
  1. [18]

    Unsupervised representation learning for time series with temporal neighborhood coding.arXiv preprint arXiv:2106.00750,

    Sana Tonekaboni, Danny Eytan, and Anna Goldenberg. Unsupervised representation learning for time series with temporal neighborhood coding.arXiv preprint arXiv:2106.00750,

    arXiv

  2. [19]

    What should not be contrastive in contrastive learning.arXiv preprint arXiv:2008.05659,

    Tete Xiao, Xiaolong Wang, Alexei A Efros, and Trevor Darrell. What should not be contrastive in contrastive learning.arXiv preprint arXiv:2008.05659,

    2008 arXiv

  3. [20]

    Rebar: Retrieval-based reconstruction for time-series contrastive learning.arXiv preprint arXiv:2311.00519,

    Maxwell A Xu, Alexander Moreno, Hui Wei, Benjamin M Marlin, and James M Rehg. Rebar: Retrieval-based reconstruction for time-series contrastive learning.arXiv preprint arXiv:2311.00519,

    arXiv

  4. [23]

    We apply Algorithm 1 in Kong and Zhang (2023) to each case to determine what minimal set of shared latent variables are recovered

    There are three possible types of masked subsequence regions: (1) a mask bordering the left edge, or the beginning of the sample, (2) a mask bordering the right edge, or the end of the sample, and (3) a mask in the middle of the sample that is bordered on the left and right by...

    2023

  5. [27]

    To obtain a representative embedding for each time window, we apply a global max pooling layer to aggregate features across the temporal dimension

    and report the best available performance from prior work or our own experiments. To obtain a representative embedding for each time window, we apply a global max pooling layer to aggregate features across the temporal dimension. Below, we describe each baseline in more detail...

    2020

  6. [28]

    We include this baseline in our work since it’s generative process resembles PULSE and our results show it is a competitive baseline when ap- plied onto physiological time-series

    is a generative model originally developed for sequential data (video and audio) and has not previously been applied to physiological signals. We include this baseline in our work since it’s generative process resembles PULSE and our results show it is a competitive baseline w...

    2015

  7. [29]

    In total, 76,590 distinct subsequences are extracted from 23 recordings, each lasting approximately 9.25 hours and sampled at 250 Hz with two channels

    and the con- vention used in ECG classification studies more broadly (Wagner et al., 2020). In total, 76,590 distinct subsequences are extracted from 23 recordings, each lasting approximately 9.25 hours and sampled at 250 Hz with two channels. To further improve computational ...

    2020

  8. [100]

    The results reported in Table 1 are the average result for all three systems

    For each system and noise level, we measure the classification accuracy averaged over ten random seeds, accounting for both dataset sampling (classification difficulty) and model initialization. The results reported in Table 1 are the average result for all three systems. C BA...

    2022

  9. [200]

    Epilepsy

    The following data are publicly available through the links provided in the repository at https://github.com/mims-harvard/TFC-pretraining. Epilepsy. (Andrzejak et al., 2001)The Epilepsy dataset contains 500 single-channel EEG record- ings, each lasting 23.6 seconds. To minimiz...

    2001

  10. [1963]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

    arXiv

  11. [1984]

    Koopman operator based dynamical sim- ilarity analysis for data-driven quantification of distance between dynamics

    Shunsuke Kamiya, Jun Kitazono, and Masafumi Oizumi. Koopman operator based dynamical sim- ilarity analysis for data-driven quantification of distance between dynamics. InICLR 2024 Work- shop on Representational Alignment,

    2024

  12. [2000]

    contains 39 whole-night electroen- cephalography (EEG) recordings collected using sleep cassettes from 20 healthy subjects. Following the preprocessing protocol of (Chambon et al., 2018), we use two EEG leads (Fpz-Cz and Pz-Oz) to evaluate pretraining, sampled at 100 Hz and se...

    2018

  13. [2001]

    Heart rate variability series is the output of a non-chaotic system driven by dynamical noise.arXiv preprint arXiv:2404.11385,

    M Bianco, A Scarciglia, C Bonanno, and G Valenza. Heart rate variability series is the output of a non-chaotic system driven by dynamical noise.arXiv preprint arXiv:2404.11385,

    arXiv

  14. [2009]

    Guidelines for augmentation selection in contrastive learning for time series classification.arXiv preprint arXiv:2407.09336,

    Ziyu Liu, Azadeh Alavi, Minyi Li, and Xiang Zhang. Guidelines for augmentation selection in contrastive learning for time series classification.arXiv preprint arXiv:2407.09336,

    arXiv

  15. [2011]

    This is the canonical 3D nonlinear attractor used to study chaotic behavior in dynamical systems, with a state-space trajectory that resembles butterfly wings

    and neural activity (Chen et al., 2024; Mudrik et al., 2024; Sussillo et al., 2016). This is the canonical 3D nonlinear attractor used to study chaotic behavior in dynamical systems, with a state-space trajectory that resembles butterfly wings. For this system,M= 3, and the dy...

    2024

  16. [2012]

    Lfads-latent factor analysis via dynamical systems.arXiv preprint arXiv:1608.06315,

    David Sussillo, Rafal Jozefowicz, LF Abbott, and Chethan Pandarinath. Lfads-latent factor analysis via dynamical systems.arXiv preprint arXiv:1608.06315,

    arXiv

  17. [2013]

    Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, Zhenan Sun, Hao Luo, and Dacheng Tao

    doi: 10.3389/fnins.2013.00267. Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, Zhenan Sun, Hao Luo, and Dacheng Tao. A survey on self-supervised learning: Algorithms, applications, and future trends.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):9052–9071,

    2013

  18. [2016]

    Deterministic chaos seen in terms of feedback circuits: Analysis, synthesis,” labyrinth chaos”.International Journal of Bifurcation and Chaos, 9(10):1889–1905,

    Ren´e Thomas. Deterministic chaos seen in terms of feedback circuits: Analysis, synthesis,” labyrinth chaos”.International Journal of Bifurcation and Chaos, 9(10):1889–1905,

    1905

  19. [2018]

    A time series is worth 64 words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730,

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730,

    arXiv

  20. [2019]

    Infinite bifurcations in thomas system.arXiv preprint arXiv:2408.09525,

    13 Under review as a conference paper at ICLR 2026 Idan Sorin and Michael Tulchinsky. Infinite bifurcations in thomas system.arXiv preprint arXiv:2408.09525,

    2026 arXiv

  21. [2020]

    Self-supervised learning to unveil brain dysfunctional signatures in brain disorders: Methods and applications.Health Data Science, 5:0282, 2025a

    Ying Li, Yanwu Yang, Yuchu Chen, Chenfei Ye, and Ting Ma. Self-supervised learning to unveil brain dysfunctional signatures in brain disorders: Methods and applications.Health Data Science, 5:0282, 2025a. Yingzhen Li and Stephan Mandt. Disentangled sequential autoencoder.arXiv...

    arXiv

  22. [2022]

    WY k=1 p(Yn,tk |Xn,tk ) #

    14 Under review as a conference paper at ICLR 2026 APPENDIX A PROOF FORTHEOREM1 Theorem 1.Given two time seriesY i andY j independently sampled from the same system (i.e., Θi =Θ j =Θ (s)) under the generative process defined by Eq. 1 and Assumption 1, the minimal set of latent...

    2026

  23. [2023]

    4 under the full-sample masking scheme, it recoversC={Θ (s)}

    is applied to the graphical model in Eq. 4 under the full-sample masking scheme, it recoversC={Θ (s)}. In the selection stage, Θ(s) is the only parent node connecting both samplesY i andY j. In the pruning stage, nothing is removed sinceCcontains only a single element, implyin...

    2026

  24. [2024]

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

    Mingyue Cheng, Qi Liu, Zhiding Liu, Hao Zhang, Rujiao Zhang, and Enhong Chen. Timemae: Self- supervised representations of time series with decoupled masked autoencoders.arXiv preprint arXiv:2303.00320,

  25. [2025]

    Perturbations of hindmarsh-rose neuron dynamics by fractional operators: Bifurcation, firing and chaotic bursts

    11 Under review as a conference paper at ICLR 2026 Emile Franc Doungmo Goufo, Melusi Khumalo, and Patrick M Tchepmo Djomegni. Perturbations of hindmarsh-rose neuron dynamics by fractional operators: Bifurcation, firing and chaotic bursts. Discrete and Continuous Dynamical Syst...

    2026

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.