acceptodds
Withdrawn from ICLR 2027

Takens-JEPA: Scientific Inference Across Sampling Intervals

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Physical processes are often recorded at different time intervals, and a model trained at one interval can become unreliable when that interval changes. We ask whether a self-supervised representation can support scientific inference when the sampling interval at deployment differs from that used for downstream training. Takens’ embedding theorem suggests an answer: histories that end at the same mo- ment describe the same underlying state even when their samples are spaced differ- ently. Takens-JEPA turns this correspondence into an auxiliary ranking objective for a joint-embedding predictive architecture, drawing together histories that share an endpoint and separating them from histories with other endpoints. It needs no new data, since the extra histories are subsampled from stored trajectories, and adds no inference cost. On Active Matter, Rayleigh-Bénard convection, and Shear Flow from The Well, we fit parameter-estimation and endpoint-state reconstruc- tion heads to frozen representations at one interval and evaluate them unchanged at coarser, partly unseen intervals. On Active Matter, Takens-JEPA roughly halves the predictive baseline’s parameter-estimation error at every shifted interval and cuts endpoint-state error by a third; on Shear Flow it lowers parameter-estimation error under shift by up to 62% at the cost of a small loss at the training interval; gains on Rayleigh-Bénard are smaller. A control that applies the predictive objec- tive at several intervals recovers part of this gain: Takens-JEPA is more accurate at the training interval and the nearest unseen interval, matches the control at the coarsest interval, and uses about one-third less training compute.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
self-supervised learning, joint-embedding predictive architecture, representation learning, physical dynamics, dynamical systems, Takens embedding theorem, delay coordinates, temporal sampling rate, distribution shift, scientific machine learning, parameter estimation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_aPwaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper extends ST-JEPA by introducing an auxiliary ranking loss between observation histories sampled at different cadences but sharing the same endpoint. A cadence-conditioned projector is used to establish correspondences between these histories. Experiments show that Takens-JEPA effectively improves cross-cadence transfer performance over ST-JEPA.

Critical strengths
  1. The evaluation problem is clearly defined and meaningful. Cross-cadence transfer for scientific inference is a concrete problem with practical relevance, and learning correspondences between endpoint-aligned histories through an auxiliary ranking loss is a reasonable design.

  2. The paper presents experiments across multiple physical systems and downstream tasks, together with several ablations. The results provide convincing positive evidence for the effectiveness of the proposed method over ST-JEPA in several settings.

Critical weaknesses
  1. My main concern comes from the Multi-Cadence ST-JEPA ablation. Simply applying predictive pretraining at multiple cadences already recovers a large portion of the cross-cadence improvements. On Active Matter parameter estimation, Multi-Cadence ST-JEPA achieves NMSE of 0.1608 versus 0.1403 for Takens-JEPA at τ = 5, and 0.1973 versus 0.1947 at τ = 6. Although Takens-JEPA performs better at τ = 3 and τ = 4 and uses less training compute, these results raise concerns about whether the Takens-inspired endpoint-aligned ranking objective provides sufficiently distinctive, consistent, and generalizable benefits. Moreover, this comparison is only conducted on Active Matter parameter estimation.

  2. Appendix D shows that adding more cadence relations does not necessarily improve performance. In particular, the chained relation set {(3, 6), (6, 9)} outperforms the configuration containing all three forward relations {(3, 6), (3, 9), (6, 9)}. I find this an interesting and somewhat counterintuitive observation that is directly related to the core design of the proposed method. However, the paper does not provide a deeper analysis or explanation of this phenomenon, which I consider a noticeable gap.

  3. Minor presentation issue: The figures have noticeable presentation issues. Figures 2 and 3 are raster images with limited clarity, and some text and annotations are difficult to read. In Figure 3, some annotations extend beyond the panel boundaries, and the overall layout and visual presentation could be improved.

Questions for authors
  1. Regarding Multi-Cadence ST-JEPA, could the authors provide comparisons on the other two datasets and on the endpoint-state reconstruction task? These results would help clarify whether the proposed endpoint-aligned ranking objective provides additional benefits beyond simply exposing the model to multiple sampling cadences.

  2. How should we understand the results in Appendix D, where adding more cadence relations does not necessarily improve performance?

Reviewer_34QoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce Takens-JEPA with the motivation to handle different sampling intervals of dynamical systems. This is useful if for example when the frequency at which the system gets measured changes. Their motivation comes from Takens' embedding theorem which states that under suitable mathematical conditions a sufficiently rich history of observations can uniquely represent the underlying state of a dynamical system. The authors attempt to achieve this invariance by extending ST-JEPA (Qu et al) with an auxiliary loss term that encourages the model to embed different observations of the same system similarly.

Critical strengths

S1:The motivation based on Takens theorem is interesting.

S2: The results show substantial improvements over ST-JEPA on Active Matter and Shear Flow, particularly under changes in sampling cadence.

Critical weaknesses

W1: The directed cadence relations used in the main experiments were selected based on Active Matter test performance at the evaluation cadences (Appendix D). This introduces test-set-informed model selection and weakens the generalization claims, particularly for Active Matter.

W2: The novelty seems quite limited. It's basically just an additional auxiliary loss to ST-JEPA and its not surprising that the model performs better under this setting since it has been trained for it. Their Multi-Cadence ST-JEPA baseline also seems to capture most of these performance gains already.

W3: Positive pairs share the same endpoint frame and potentially several additional frames. The model could therefore identify matching histories through shared spatial content rather than learning the dynamical correspondence motivated by Takens' theorem. The experiments do not isolate these possibilities.

W4: All experiments use uniformly spaced temporal observations with integer sampling intervals and synchronized endpoints. The method is not evaluated on irregularly sampled trajectories, missing observations, or asynchronous measurements, which are common in real-world scientific data. This limits the practical generality of the proposed approach.

Questions for authors

See my weaknesses.

Reviewer_a2BcOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Takens-JEPA, an auxiliary margin-ranking objective for spatiotemporal JEPA pretraining. Motivated by Takens embedding theorem, delay-chart representations sharing an endpoint but built at different sampling strides are pulled together relative to permuted batch negatives. On three datasets from The Well, heads trained at τ=3 on frozen encoders are evaluated unchanged at coarser cadences. The method reduces parameter-estimation and endpoint-reconstruction error over the ST-JEPA baseline, most notably on Active Matter, at no extra inference cost.

Critical strengths
  1. The method is simple, well motivated, and adds no inference cost, since the auxiliary projector is discarded after pretraining. The cadence-shift evaluation protocol is clearly defined, and the paper designs detailed experiments.

  2. The idea of relating same-endpoint histories across strides is a clean, theory-motivated use of Takens' theorem, and the finding that relational alignment beats simply training on more cadences (at one-third less compute) is an interesting empirical observation.

Critical weaknesses
  1. Some hyperparameters were selected on the test cadences, which is acknowledged in the limitations. The key Multi-Cadence control covers only Active Matter parameter estimation. It remains to be clarified whether this enables a fair comparison.

  2. The paper claims the ranking objective makes the encoder cadence-invariant, but Appendix J shows the same-state structure is almost entirely confined to the discarded projector, with very low margin accuracy in the retained encoder space, contradicting the paper's own claim and calling for clarification.

  3. The advantages are evident only on Active Matter, while the claimed strong performance is not observed on Rayleigh-Bénard or Shear Flow, which calls for further analysis.

  4. All cadences are synthesized by subsampling the same archived trajectories, yet the natural remedy of interpolating the coarse inputs back to τ=3 is neither tested nor discussed. Related sampling-rate-equivariant time-series work (e.g., TE-ESN(IJCAI 2021), Warpformer(KDD 2023), FlowState(ICML 2026)) is also missing from the related work.

Questions for authors

See Critical Weaknesses.

Reviewer_XuctOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Takens-JEPA, which studies the following question: Can a model, trained on one time interval, transfer at inference time to the same task at a different time interval? To address this, the authors use Takens embedding theorem, which says, in the author’s words “histories that end at the same moment describe the same underlying state even when their samples are spaced differently.“, implying that such a transferable model should indeed be possible.The authors propose incorporating this as an auxiliary objective in JEPA training. Using various datasets from The Well (Active Matter, Rayleigh-Bénard, Shear Flow), they demonstrate gains on shifted intervals.

Critical strengths

The problem setting is quite interesting. JEPAs at different time intervals have been a topic of study in the context of hierarchical JEPA, but the question of whether a single model can understand different time intervals has not been seriously explored, and such a model could be useful in terms of robustness, and could itself perhaps be used in a hierarchical framework. This is actually a broader motivation than the author’s framing, which focuses on physical processes, but regardless the problem is well motivated.

As its main comparison, the paper uses ST-JEPA (Qu et al. 2026), which is an appropriate and recent baseline. Figure 2, and tables 1,2, and 3, shows solid improvements over the baseline, and section 4.3 contains a necessary ablation.

Critical weaknesses

As we see in table 4, in some cases, when we use the same interval for training and inference, Takens-JEPA does worse than ST-JEPA, a very serious limitation that the paper brushes over, and even overclaims about. For example the “small loss” on Shear Flow at the training interval is actually a 33 percent increase.

The comparisons also do not take into account unfair computation, as shown in Table 10. Takens-JEPA uses double the compute, so fair experiments would train for twice as long on the baseline.

Several experiments are reported as single runs (line 689 for example). Its unclear why this is done inconsistently, and needs to be fixed.

The evaluations are also pretty limited, and additional settings beyond these 3 would greatly strengthen the results

Questions for authors

Can you run fair compute comparisons? I suspect this will affect the results Can you explain why some results are reported on a single seed? Can you redraft the overclaims, especially in the abstract, and in light of this, address the significance of worse performance on the training interval? Can you add 3 more datasets to ensure the results are robust?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.