Pith. sign in

REVIEW 1 major objections 5 minor 43 references

SJEPA: Learning Elegant Latent Dynamics with Hybrid Symbolic-Neural Predictors

T0 review · 1 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Joint symbolic–neural learning discovers latent dynamics 5.6 times simpler than post-hoc symbolic regression.

desk verdict SJEPA has a genuinely new idea and sound theory, but its headline rollout comparison is confounded by a poorly fitting affine decoder and needs revision before the empirical claims are convincing. read the letter →

arxiv 2608.04060 v1 pith:4EO7JZQM submitted 2026-08-04 cs.LG cs.AI

Yongchao Huang This is my paper · ORCID
classification cs.LGcs.AI
keywords joint-embeddingpredictivearchitecturesymbolicregressionlatentdynamicsoperatorcompressionrepresentationcollapsehybridsymbolic-neuralmodelgoverningequationdiscoveryself-supervisedlearning
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 proposes SJEPA, a reconstruction-free joint-embedding predictive architecture whose latent transition is a hybrid of a compact symbolic law and a regularized neural correction. Its central claim is that operator compression—minimizing the complexity of the symbolic-plus-correction transition subject to predictive adequacy and non-collapse constraints—selects latent coordinates whose induced dynamics are simpler and remain adequately predictive. In a controlled pendulum task, jointly learning representation and symbolic dynamics reduced mean weighted symbolic complexity from 26.0 to 4.67 and physical-state test rollout error from 1.229 to 0.467 relative to fitting symbols post hoc to frozen predictive coordinates. The paper also shows that without representation constraints, unconstrained operator compression drives the representation to collapse to a near-constant state, making the transition trivially simple. A second experiment shows that correction regularization keeps a representable symbolic mechanism intact under grammar misspecification while directing the neural component to residual dynamics.

What carries the argument

The load-bearing object is operator compression realized through the induced-dynamics complexity functional. Given encoders, $C_{\mathrm{dyn}}(E_\theta,E_{\bar\theta}) = \inf_{E,\alpha,\phi}\{\Omega(E)+\lambda_c R_{\mathrm{corr}}(\phi) : L_{\mathrm{pred}} \leq \delta_{\mathrm{pred}}\}$ asks how simple the latent transition can be once coordinates are fixed; the outer problem $\min_{(\theta,\bar\theta)\in\Theta_{\mathrm{repr}}} C_{\mathrm{dyn}}$ selects coordinates whose dynamics are easiest to describe. The hybrid predictor $H=F+c_\phi$ combines a typed symbolic expression with structural complexity $\Omega(E)$ and a neural correction regularized by $R_{\mathrm{corr}}(\phi)=\mathbb{E}[\|c_\phi\|_2^2]+\tau_\phi\|\phi\|_2^2$, which controls how much predictive structure is delegated to the correction. Four propositions carry the argument: predictive coordinates are non-identifiable under bijections; unconstrained compression has a collapsed global minimizer; the symbolic–neural decomposition is non-identifiable; and correction regularization shrinks the correction toward the conditional-mean residual by the factor $(1+\lambda)^{-1}$. The practical objective is a scalarized constrained problem training one-step prediction, rollout, representation quality, symbolic parsimony, and correction control jointly.

What would settle it

Run SJEPA and the post-hoc symbolic baseline on a system with known simple latent dynamics under an observation map where the simple coordinates are nonlinearly mixed into observations, then evaluate rollouts through an invertible nonlinear decoder rather than an affine probe; if SJEPA's complexity and rollout advantages vanish under the correct alignment, the operator-compression claim would be refuted. Alternatively, a direct search for the collapsed solution in the unconstrained one-step objective, checking whether every seed reaches a near-constant representation with identity dynamics, would confirm or refute the collapse-shortcut prediction.

Watch

Extended reading notes

Core claim

The paper's core discovery is that the complexity of the induced transition operator can serve as a criterion for learning predictive representations: among informative, non-collapsed coordinate systems, SJEPA selects coordinates whose evolution admits the simplest adequate symbolic law. Formally, for fixed encoders the induced-dynamics complexity $C_{\mathrm{dyn}}(E_\theta, E_{\bar\theta})$ is the minimum symbolic-plus-correction complexity needed to meet a predictive tolerance, and representation learning minimizes this quantity over admissible encoder pairs. The hybrid predictor $H_{E,\alpha,\phi}(Z_C,\varepsilon)=F_{E,\alpha}(Z_C,\varepsilon)+c_\phi(Z_C,\varepsilon)$ acts on the transition operator rather than partitioning the representation. Predictive coordinates are shown to be non-identifiable up to arbitrary bijections, so operator complexity supplies the extra criterion; unconstrained compression is shown to admit a degenerate global solution where both encoders map everything to one point and an identity transition predicts perfectly. The empirical results support the operator-compression claim: joint learning discovers oscillator-like equations across seeds, lowers rollout divergence, and the unregularized diagnostic reaches the predicted collapse shortcut.

Load-bearing premise

The main load-bearing premise is that the affine map fitted from training latents to the physical state is a fair way to compare rollout quality across models, even though for the jointly learned SJEPA coordinates that affine probe explains only about 29% of test variance; if that alignment assumption fails, the reported rollout improvement over post-hoc fitting is not a clean comparison of latent-dynamics quality.

Editorial extensions

If this is right

  • If SJEPA's principle holds, latent coordinates learned under operator compression will expose inspectable governing equations rather than black-box maps, with comparable or better recursive stability than post-hoc symbolic fits.
  • The collapse shortcut implies that any method that adds an equation-simplicity penalty to a JEPA-style objective must pair it with explicit non-collapse constraints; otherwise the simplest 'law' is the identity on a near-constant representation.
  • Correction regularization yields a controllable symbolic–neural allocation: the symbolic component retains representable mechanisms and the neural component tracks residual dynamics, with the trade-off visible in the normalized correction-energy ratio.
  • A complete symbolic grammar can remove the need for a neural correction in controlled settings, recovering the true vector field to within about 7% coefficient error.
  • Long-horizon accuracy is a separate requirement from one-step adequacy: rollout error grows through the learned transition's Lipschitz constant, so compactness alone does not guarantee non-divergence.

Reading between the lines

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

  • The paper does not claim that operator complexity is a general model-selection principle, but the same constrained objective could plausibly regularize latent-dynamics models beyond JEPA whenever inspectable equations are desired.
  • An untested extension: replacing the affine alignment probe in Experiment 1 with a nonlinear invertible decoder would show whether the rollout advantage of SJEPA persists under a fairer coordinate comparison.
  • The non-identifiability result implies that reported symbolic complexity is convention-dependent; publishing the grammar, coefficient threshold, and latent normalization alongside the discovered equation would make complexity comparisons reproducible across methods.
  • The action-conditioned planning interface is provided but not evaluated; a natural next step is a control task where the discovered law reveals which latent directions are action-sensitive.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 5 minor

Summary. SJEPA introduces a joint-embedding predictive architecture in which the latent transition is a hybrid of a compact symbolic law and a regularized neural correction. The paper's central principle is operator compression: among informative, non-collapsed predictive representations, select coordinates whose induced dynamics have minimal symbolic-plus-correction complexity subject to predictive adequacy. The theory formalizes induced-dynamics complexity, proves that predictive coordinates and the symbolic/correction decomposition are non-identifiable from prediction loss alone, identifies a collapse shortcut created by unconstrained operator compression, derives the optimal correction under quadratic regularization, and bounds rollout error under uniform one-step adequacy. Two controlled pendulum experiments are reported: Experiment 1 compares jointly learned symbolic coordinates against post-hoc symbolic regression on frozen Neural JEPA coordinates, reporting lower symbolic complexity (mean Ω 4.67 vs 26.0) and lower physical-state rollout MSE; Experiment 2 studies grammar misspecification and the effect of correction regularization on symbolic–neural allocation. The paper is careful to state limitations and to separate predictive accuracy from parsimony claims.

Significance. If the central empirical comparison were clean, SJEPA would be a valuable contribution: it turns transition complexity into an explicit learning signal inside a reconstruction-free JEPA, with clearly stated propositions and proofs that are elementary but appropriate. The controlled experimental design, disjoint train/validation/test/OOD splits, per-seed reporting, and explicit acknowledgement of scope are strengths. The paper is also unusually honest about the trade-off between symbolic parsimony and raw predictive accuracy. However, the main empirical claim of lower physical-state rollout error currently rests on an affine decoder that is nearly uninformative for the jointly learned coordinates, so the headline comparison is not yet established in observable coordinates. This makes the significance conditional on a revised evaluation.

major comments (1)
  1. [Section 8.1, Table 4, Appendix D.3, Table 12] The central empirical comparison—joint SJEPA reducing physical-state rollout MSE from 1.229 to 0.467 relative to post-hoc symbolic regression—is confounded by the affine alignment map used for evaluation. Table 12 reports test affine-probe R² = 0.293±0.154 for SJEPA, versus 0.814±0.063 for Neural JEPA/post-hoc coordinates. Because the affine decoder explains only about 29% of the variance of SJEPA's test latents, the reported physical-state rollout errors for SJEPA are dominated by decoder misspecification rather than by the quality of the learned latent transition. The manuscript acknowledges this lower alignment in Section 8.1 but continues to use affine-aligned rollout as the head-to-head metric in Table 4 and in the abstract/conclusion claims of 'lower physical-state rollout error.' This is load-bearing because the physical-state rollout comparison is the quantitative basis for the claim that joint learning discovers more adequate dynamics, not merely simpler latent equations. Please either (a) evaluate rollout error through a matched-capacity nonlinear decoder trained to minimize physical-state MSE for all models, (b) use a coordinate-invariant evaluation such as mapping rolled-out latents back to observation space through the known forward observation model, or (c) reframe the empirical claim as concerning complexity of induced latent dynamics only, treating physical-state rollout numbers as illustrative and clearly subordinate to the affine-probe caveat. As written, the claimed cross-model rollout improvement is not a clean comparison of latent dynamics quality.
minor comments (5)
  1. [Section 9, first paragraph] The sentence discussing the two coupled questions ends with a stray '13?' marker that appears to be a leftover reference callout; please remove it.
  2. [Section 8.2 and Appendix D] Equation references such as '(Eq.64)' and '(Eq.81)' mix textual labels with equation numbers; please use one consistent numbering convention throughout.
  3. [References] The entries 'Posner et al. 2026a' and 'Posner et al. 2026b' have identical titles and identical URLs; if they are the same work, list it once, and if they are different works, provide distinct titles or identifiers.
  4. [Section 8.1, Table 3 and Table 5] The terminology for the unregularized one-step condition is inconsistent: Table 3 calls it 'one-step collapse diagnostic' while the text and Table 5 refer to a 'no-RIB' condition. Please align the terminology.
  5. [Table 4] Because latent MSE is coordinate-dependent, consider adding a footnote or a 'within-coordinate' qualifier directly in the table header to discourage cross-model comparison of this column.

Circularity Check

1 steps flagged · score 2.0 of 10

Complexity-reduction claim is partly by construction because symbolic complexity is the optimized objective; rollout/OOD comparisons and the theoretical results remain independent.

  1. self definitional [Section 8.1 (Table 4); Eq. (29); Appendix D.3]
    "Joint coordinate search reduces mean symbolic complexity from 26.0 to 4.67, a factor of approximately 5.6. ... LtrainSJEPA = Lpred + λrLroll + λIBRIB + λsΩtrain(E, α) + λcRcorr(ϕ) ... among candidates whose validation risks lie within a predefined relative tolerance, the lower-complexity symbolic law is selected."

    The headline metric Ω reported in Table 4 is the same symbolic-complexity term that appears in the training objective Eq. (29) as λsΩtrain, and the checkpoint-selection rule in Appendix D.3 explicitly prefers lower discrete symbolic complexity when validation risks are within 2%. The measured reduction from 26.0 to 4.67 is therefore substantially a restatement of the objective being minimized and of the selection criterion, rather than an independent empirical prediction of the operator-compression theory. The claim is not fully circular because the post-hoc baseline also minimizes Ω on frozen coordinates and the rollout/OOD comparisons use external physical-state metrics, which are not part of the complexity definition; those parts retain independent content.

full rationale

The paper's theoretical propositions (coordinate non-identifiability, the collapse shortcut, corrected-allocation formula, rollout bound, and MAP equivalence) are derived from stated assumptions in the appendices and do not import their conclusions from the training objective. The self-citations to the author's VJEPA and information-bottleneck work are related-work or modular options and are not load-bearing for the central empirical claims; the experiments instantiate RIB with VICReg, not with the cited predictive bottleneck. The only mild circularity is that the symbolic-complexity improvement in Experiment 1 is partly by construction, since Ω is the optimized objective and is also used as a checkpoint tie-breaker. However, the rollout and OOD divergence comparisons use physical-state metrics that are external to the complexity objective, and the comparison against post-hoc symbolic regression on frozen coordinates gives the simplicity result empirical content. The affine-decoder concern raised about the physical-state rollout evaluation is a validity/interpretation issue, not a circularity: the affine map is fitted on training latents only and does not encode SJEPA's advantage by construction. Overall, the derivation chain is largely self-contained, so a low score is appropriate.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claims rest on a residual ODE latent model, a hand-chosen symbolic grammar and complexity weighting, a specific VICReg-style regularizer, and a heuristic alternating optimizer. The theories themselves are self-contained, but the empirical successes depend on these domain assumptions; the paper is transparent about most of them.

free parameters (6)
  • Primitive complexity weights w_type(v) (Eq. 4) = e.g., (0.5,1,1,2,2,2,2,1.5,1.5,2) in Exp. 1; weight 3 for p|p| in Exp. 2
    Defines the symbolic parsimony metric and is used both in the training objective and in the reported complexity comparisons; a different weighting scheme would change the magnitude of the claimed 5.6x simplification.
  • Regularization weights (lambda_s, lambda_c, lambda_IB, lambda_r, rollout weight 0.35) = lambda_s=2.5e-4 (Exp.1), 1.5e-4 (Exp.2); lambda_c=1.5e-2 (Exp.2); see Table 11
    Control the trade-off between predictive fidelity, symbolic parsimony, correction reliance, and representation quality; chosen by the authors and validated manually rather than by a principled selection method.
  • Symbolic library and grammar masks = 10-term library in Exp. 1; output-specific masks {p}, {sin(q)}, {p|p|} in Exp. 2
    The central empirical successes depend on the library containing the true terms; the paper explicitly studies misspecification in Exp. 2, and the complete grammar is a strong oracle.
  • Coefficient thresholds and reporting threshold = STLSQ threshold 0.035, zeroing threshold 0.0175, reporting threshold 0.05
    These thresholds determine which terms appear in the reported equations and therefore affect the claimed complexity values and the per-seed equations.
  • Affine alignment map from latents to physical state = Fitted on training latents per model; SJEPA probe R2=0.293 vs 0.814 for Neural JEPA
    Used to compute the physical-state rollout metrics that support the 'lower rollout error' claim; the map is a post-hoc fitted decoder and its quality differs substantially across models.
  • Latent dimension (d_z=2) = 2
    The representation dimensionality is chosen by hand; the results are specific to a 2D latent space matching the pendulum's state dimensionality.
assumptions (5)
  • domain assumption The latent transition is well-approximated by the residual forward-Euler scheme z_{t+1}=z_t+Delta_t[F(z)+c(z)] with variable time steps.
    Used to define the predictor in Eq. (46); if the learned latent dynamics are not captured by this ODE-plus-Euler approximation, the discovered symbolic laws may be integrator artifacts.
  • domain assumption The selected symbolic grammar contains the true mechanism for the complete-grammar condition, and the complexity weights reflect a meaningful notion of simplicity.
    Experiment 2's recovery of the drag law requires the p|p| primitive; the theory offers no guarantee of discovery when the grammar is incomplete, which the paper acknowledges.
  • domain assumption The VICReg-style representation regularizer ensures informative, non-collapsed, and predictive latent coordinates.
    The collapse analysis (Prop 4.5) shows constraints are necessary, but the paper provides no proof that the specific variance, covariance, and invariance terms preserve the information needed for symbolic dynamics; it is validated only on the pendulum experiment.
  • domain assumption Alternating gradient-based optimization can find encoders whose induced dynamics have low operator complexity; no global convergence guarantee exists.
    Algorithm 1 alternates dynamics search and space search; the paper notes the problem is nonconvex and the method is susceptible to local optima.
  • domain assumption The latent state is a sufficient statistic for future prediction (Markov assumption) and observations are sampled i.i.d. from a fixed distribution.
    The JEPA objective assumes the context embedding carries all predictive information; in partially observed or non-stationary settings this can fail, as the paper notes in the limitations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SJEPA: Learning Elegant Latent Dynamics with Hybrid Symbolic-Neural Predictors." pith.science (2026). https://pith.science/paper/4EO7JZQM

@misc{pith2026260804060,
  author       = {Pith},
  title        = {Pith review of: SJEPA: Learning Elegant Latent Dynamics with Hybrid Symbolic-Neural Predictors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4EO7JZQM}},
  note         = {Machine review of arXiv:2608.04060}
}
read the original abstract

Joint-embedding predictive architectures learn abstract states by predicting target embeddings from context embeddings, but their transition models are typically opaque neural maps. We introduce SJEPA, a reconstruction-free JEPA framework that learns predictive representations whose induced dynamics admit compact symbolic descriptions. Its hybrid transition combines a symbolic law with a regularised neural correction for dynamics outside the selected grammar. The central principle is to learn the simplest adequate dynamics: representation constraints preserve informative, non-collapsed predictive coordinates, while operator compression favours low-complexity symbolic-neural transitions that remain predictively adequate. We formalise this principle through induced-dynamics complexity, analyse predictive-coordinate non-identifiability, and show that unconstrained operator compression creates a direct shortcut to representation collapse. The framework supports both alternating representation-equation learning and symbolic dynamics fitted to fixed representations. In controlled pendulum experiments, joint learning discovers substantially simpler symbolic dynamics with lower long-horizon rollout error and divergence than post-hoc fitting, while an unconstrained one-step diagnostic realises the predicted collapse shortcut. Under grammar misspecification, correction regularisation preserves the representable symbolic mechanism and directs the neural component towards residual dynamics. The results expose a controllable trade-off among predictive fidelity, representation quality, symbolic parsimony, and symbolic-neural allocation.

Figures

Figures reproduced from arXiv: 2608.04060 by the authors.

Figure 1
Figure 1. SJEPA decomposes the latent transition operator into a compact symbolic governing law and a neural correction for dynamics. The decomposition acts on the transition operator rather than partitioning the latent representation. where FE,α ≡ FE,α is a symbolic expression with structure E and coefficients α, and cϕ is a neural corrector for dynamics not captured adequately by the symbolic grammar. Setting cϕ ≡ 0 recover… view at source ↗
Figure 2
Figure 2. Experiment 1 evaluates operator compression and representation collapse. [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Experiment 2 evaluates controlled grammar mismatch and symbolic–neural allocation. [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Additional qualitative and mechanism-specific diagnostics. [PITH_FULL_IMAGE:figures/full_fig_p039_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 20 canonical work pages

  1. [1]

    2022 , howpublished =

    Yann LeCun , title =. 2022 , howpublished =

    work page 2022

  2. [2]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Mahmoud Assran and Quentin Duval and Ishan Misra and Piotr Bojanowski and Pascal Vincent and Michael Rabbat and Yann LeCun and Nicolas Ballas , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    work page

  3. [3]

    Transactions on Machine Learning Research , issn=

    Revisiting Feature Prediction for Learning Visual Representations from Video , author=. Transactions on Machine Learning Research , issn=. 2024 , url=

    work page 2024

  4. [4]

    2025 , eprint=

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning , author=. 2025 , eprint=

    2025

  5. [5]

    , title =

    Koza, John R. , title =. 1992 , isbn =

    1992

  6. [6]

    Science , volume =

    Michael Schmidt and Hod Lipson , title =. Science , volume =. 2009 , doi =. https://www.science.org/doi/pdf/10.1126/science.1165893 , abstract =

    doi:10.1126/science.1165893 2009

  7. [7]

    Brunton and Joshua L

    Steven L. Brunton and Joshua L. Proctor and J. Nathan Kutz , title =. Proceedings of the National Academy of Sciences , volume =. 2016 , doi =. https://www.pnas.org/doi/pdf/10.1073/pnas.1517384113 , abstract =

    doi:10.1073/pnas.1517384113 2016

  8. [8]

    and Kutz, J

    Kaiser, E. and Kutz, J. N. and Brunton, S. L. , title =. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume =. 2018 , month =. doi:10.1098/rspa.2018.0335 , url =

    arXiv 2018

Show all 43 references
  1. [9]

    Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =

    Cranmer, Miles and Sanchez-Gonzalez, Alvaro and Battaglia, Peter and Xu, Rui and Cranmer, Kyle and Spergel, David and Ho, Shirley , title =. Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =. 2020 , isbn =

    2020

  2. [10]

    International Conference on Learning Representations , year=

    Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients , author=. International Conference on Learning Representations , year=

  3. [11]

    Proceedings of the 38th International Conference on Machine Learning , series =

    Luca Biggio and Tommaso Bendinelli and Alexander Neitz and Aurelien Lucchi and Giambattista Parascandolo , title =. Proceedings of the 38th International Conference on Machine Learning , series =

  4. [12]

    2023 , eprint=

    Interpretable Machine Learning for Science with PySR and SymbolicRegression.jl , author=. 2023 , eprint=

    2023

  5. [13]

    Chen, Ricky T. Q. and Rubanova, Yulia and Bettencourt, Jesse and Duvenaud, David K , booktitle =. Neural Ordinary Differential Equations , url =

  6. [14]

    Recurrent World Models Facilitate Policy Evolution , url =

    Ha, David and Schmidhuber, J\". Recurrent World Models Facilitate Policy Evolution , url =. Advances in Neural Information Processing Systems , editor =

  7. [15]

    PLOS ONE , publisher =

    Koopman Invariant Subspaces and Finite Linear Representations of Nonlinear Dynamical Systems for Control , year =. PLOS ONE , publisher =. doi:10.1371/journal.pone.0150171 , author =

    doi

  8. [16]

    Nathan and Brunton, Steven L

    Lusch, Bethany and Kutz, J. Nathan and Brunton, Steven L. , title =. Nature Communications , volume =. 2018 , doi =

    2018

  9. [17]

    Pereira and William Bialek , title =

    Naftali Tishby and Fernando C. Pereira and William Bialek , title =. Proceedings of the 37th Annual Allerton Conference on Communication, Control, and Computing , pages =

  10. [18]

    International Conference on Learning Representations , year=

    Deep Variational Information Bottleneck , author=. International Conference on Learning Representations , year=

  11. [19]

    2022 , url=

    Adrien Bardes and Jean Ponce and Yann LeCun , booktitle=. 2022 , url=

    2022

  12. [20]

    Barlow Twins: Self-Supervised Learning via Redundancy Reduction , booktitle =

    Jure Zbontar and Li Jing and Ishan Misra and Yann LeCun and St. Barlow Twins: Self-Supervised Learning via Redundancy Reduction , booktitle =

  13. [21]

    Bootstrap your own latent a new approach to self-supervised learning , year =

    Grill, Jean-Bastien and Strub, Florian and Altch\'. Bootstrap your own latent a new approach to self-supervised learning , year =. Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =

  14. [22]

    International Conference on Learning Representations , year=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. International Conference on Learning Representations , year=

  15. [23]

    Emerging Properties in Self-Supervised Vision Transformers , booktitle =

    Mathilde Caron and Hugo Touvron and Ishan Misra and Herv. Emerging Properties in Self-Supervised Vision Transformers , booktitle =

  16. [24]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    He, Kaiming and Chen, Xinlei and Xie, Saining and Li, Yanghao and Doll\'ar, Piotr and Girshick, Ross , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2022 , pages =

    2022

  17. [25]

    International Conference on Learning Representations , year=

    CLEVRER: Collision Events for Video Representation and Reasoning , author=. International Conference on Learning Representations , year=

  18. [26]

    Bear and Elias Wang and Damian Mrowca and Felix Binder and Hsiao-Yu Fish Tung and R

    Daniel M. Bear and Elias Wang and Damian Mrowca and Felix Binder and Hsiao-Yu Fish Tung and R. T. Pramod and Cameron Holdaway and Sirui Tao and Kevin Smith and Fan-Yun Sun and Fei-Fei Li and Nancy Kanwisher and Joshua B. Tenenbaum and Daniel L. K. Yamins and Judith Fan , title...

  19. [27]

    2018 , eprint=

    DeepMind Control Suite , author=. 2018 , eprint=

    2018

  20. [28]

    PDEBENCH: an extensive benchmark for scientific machine learning , year =

    Takamoto, Makoto and Praditia, Timothy and Leiteritz, Raphael and MacKinlay, Dan and Alesiani, Francesco and Pfl\". PDEBENCH: an extensive benchmark for scientific machine learning , year =. Proceedings of the 36th International Conference on Neural Information Processing Syst...

  21. [29]

    Nathan Kutz and Steven L

    Kathleen Champion and Bethany Lusch and J. Nathan Kutz and Steven L. Brunton , title =. Proceedings of the National Academy of Sciences , volume =. 2019 , doi =. https://www.pnas.org/doi/pdf/10.1073/pnas.1906995116 , abstract =

    2019 doi

  22. [30]

    Mars and Kutz, J

    Yermakov, Alexey and Zoro, David and Gao, L. Mars and Kutz, J. Nathan , title =. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume =. 2026 , month =. doi:10.1098/rsta.2024.0586 , url =

    2026

  23. [31]

    2026 , eprint=

    Extracting Governing Equations from Latent Dynamics via Multi-View Contrastive Learning , author=. 2026 , eprint=

    2026

  24. [32]

    2025 , eprint=

    LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics , author=. 2025 , eprint=

    2025

  25. [33]

    2026 , eprint=

    From Observation to Insight: Mechanistic World Models and the Quest for Autonomous Discovery , author=. 2026 , eprint=

    2026

  26. [34]

    On the Information Bottleneck of

    Huang, Yongchao , year =. On the Information Bottleneck of

  27. [35]

    Knowledge, Rules and Their Embeddings: Two Paths towards Neuro-Symbolic

    Huang, Yongchao and Raza, Hassan , journal =. Knowledge, Rules and Their Embeddings: Two Paths towards Neuro-Symbolic

  28. [36]

    Rasmussen, Carl Edward and Williams, Christopher K. I. , title =. 2005 , month =. doi:10.7551/mitpress/3206.001.0001 , url =

    2005 doi

  29. [37]

    2023 , eprint=

    MC-JEPA: A Joint-Embedding Predictive Architecture for Self-Supervised Learning of Motion and Content Features , author=. 2023 , eprint=

    2023

  30. [38]

    2025 , eprint=

    Audio-JEPA: Joint-Embedding Predictive Architecture for Audio Representation Learning , author=. 2025 , eprint=

    2025

  31. [39]

    IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=

    Point-JEPA: Joint Embedding Predictive Architecture for 3D Point Cloud Self-supervised Learning , author=. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=

  32. [40]

    2024 , eprint=

    3D-JEPA: A Joint Embedding Predictive Architecture for 3D Self-Supervised Representation Learning , author=. 2024 , eprint=

    2024

  33. [41]

    2026 , eprint=

    VL-JEPA: Joint Embedding Predictive Architecture for Vision-language , author=. 2026 , eprint=

    2026

  34. [42]

    arXiv preprint arXiv:2501.14622 , year =

    ACT-JEPA: Novel Joint-Embedding Predictive Architecture for Efficient Policy Representation Learning , author =. arXiv preprint arXiv:2501.14622 , year =

    arXiv

  35. [43]

    2026 , url=

    Yongchao Huang , booktitle=. 2026 , url=

    2026

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.