Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

Bounding Distributional Shifts in World Modeling through Novelty Detection

T0 review · 4 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a variational autoencoder (VAE) trained on the same latent states as a visual world model can act as a novelty detector, and that adding its reconstruction loss as a per-action planning cost keeps trajectories in-dist

desk verdict A sensible but thinly validated trick—adding a VAE reconstruction cost to CEM planning—with directionally positive results on three FleX tasks, but the data-efficiency claim is unsupported and the novelty proxy is unvalidated. read the letter →

arxiv 2508.06096 v1 pith:MDBMTWCR submitted 2025-08-08 cs.RO cs.AI

Eric Jing , Abdeslam Boularias This is my paper
classification cs.ROcs.AI
keywords worldmodelsnoveltydetectionvariationalautoencodermodel-basedplanningDINO-WMdistributionalshiftcross-entropymethoddataefficiency
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 tries to establish that a variational autoencoder (VAE), trained on the same latent states as a visual world model, can serve as a novelty detector that bounds distributional shift during planning. The reconstruction loss of each predicted latent state is added as a per-action cost in the CEM trajectory optimizer, steering the planner away from states the world model was not trained on. The authors show this improves goal-reaching, measured by Chamfer distance, over the plain DINO-WM baseline in three simulated manipulation tasks with limited training data. A sympathetic reading: if correct, the method makes model-based planning more data-efficient by making planners robust to the gaps any finite training set leaves in the world model's knowledge.

What carries the argument

A VAE trained on the same latent states as the world model; at inference, each predicted latent state $z_{i+1}$ is passed through the VAE, and the mean-squared error $\text{MSE}(y_{i+1}, z_{i+1})$ is added, weighted by $w$, to the trajectory cost alongside the terminal goal error. The CEM planner then favors trajectories whose predicted states the VAE can reconstruct, i.e., states near the training distribution.

What would settle it

A synthetic environment with a narrow corridor of latent states that the transition model predicts accurately but that were never shown to the VAE during training; if the novelty penalty prevents the planner from taking the corridor and the goal is unreachable, the central claim fails. Alternatively, a scatter plot of VAE reconstruction loss vs. actual world-model prediction error across a test set of trajectories; if the correlation is weak or negative, the claim collapses.

Watch

Extended reading notes

Core claim

The paper claims that adding a variational autoencoder as a novelty detector to the DINO-WM world model, and adding its reconstruction loss as a per-action penalty in the CEM trajectory cost, makes model-based planning robust to imperfect world models trained on limited data. In three simulated manipulation tasks (granular, rope, cloth), the augmented planner (WM-VAE) reaches states closer to the goal, as measured by Chamfer distance, than the plain DINO-WM planner, with the largest gain on cloth (9.228 to 5.372).

Load-bearing premise

The method works only if a low VAE reconstruction loss for a predicted latent state reliably means the world model's prediction is trustworthy; if the VAE's errors do not track the world model's errors, the extra cost term misleads the planner.

Editorial extensions

If this is right

  • Planners using pre-trained visual world models can operate with substantially smaller training datasets when each predicted state is scored by a novelty detector trained on the same data.
  • The per-action reconstruction-loss penalty discourages trajectories that leave the training distribution, reducing the compounding-error problem without changing the world model or the action space.
  • There is a finite optimal weight $w$ for the novelty cost; too small a weight fails to bound shift, and too large a weight drowns the goal signal and makes CEM converge erratically (Table III).
  • The benefit depends on the encoder: with ResNet embeddings the same novelty mechanism gives worse Chamfer distances than with DINOv2 embeddings (Table II).
  • Because the VAE is trained on the same images as the world model, the method is a drop-in addition requiring no extra environment interaction.

Reading between the lines

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

  • The same novelty-cost idea could be bolted onto other trajectory optimizers (MCTS, diffusion planners) that sample actions, since it only requires a per-state score, not a change in the dynamics model.
  • The method explicitly biases against unseen states; a natural extension is to anneal $w$ over deployment so the planner can explore novel regions once the world model has accumulated enough data.
  • Because the VAE is trained on the same data as the world model, reconstruction loss may encode memorization rather than true distributional shift; training the VAE on a held-out split would separate these two failure modes.
  • If the VAE's reconstruction loss is a faithful uncertainty estimate, it could be reinterpreted as a learned covariance for the transition model, connecting novelty detection to calibrated probabilistic dynamics.
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 / 7 minor

Summary. The paper proposes WM-VAE, an augmentation of the DINO-WM world-model planner in which a variational autoencoder is trained on the same latent representations as the world model and used as a novelty detector. During CEM planning, the VAE reconstruction loss of each predicted latent state is added as a per-action cost to the trajectory cost (Eq. 4, Algorithm 1). The authors report improved Chamfer distance on three simulated FleX manipulation tasks (Granular, Rope, Cloth) compared with the DINO-WM baseline, and interpret the gains as evidence that the method improves data efficiency and robustness to imperfect world models.

Significance. If validated, the core idea is appealing and potentially useful: a simple, model-agnostic novelty penalty integrated into a sampling-based planner could improve deployment of learned world models under limited training coverage. The paper identifies a real problem—world-model predictions drifting out of the training distribution—and proposes a lightweight solution. The method is clearly described and the evaluation spans three non-trivial manipulation domains. However, the significance is currently limited by the lack of a direct validation of the novelty proxy, by the absence of statistical analysis, and by the mismatch between the claimed data-efficiency contribution and the experiments actually performed. The paper would be a stronger contribution if it demonstrated that the reconstruction loss is correlated with world-model prediction error, rather than relying on an untested assumption.

major comments (4)
  1. [Section III-B, III-C, Eq. (4)] The central mechanism rests on the claim that the VAE reconstruction loss L_r = MSE(y,z) is a measure of world-model confidence for a predicted latent state. This is not validated. L_r is computed only from the predicted state z and its VAE reconstruction; it is never compared with the actual next observation z_actual = enc(obs_next). The paper provides no evidence that L_r is monotonically related to the transition model's true prediction error ||z_pred - z_actual||, nor that a low L_r implies a trustworthy prediction. A VAE with a regularized latent space can assign low reconstruction error to over-smoothed but inaccurate predictions, and high error to a novel but correct state. Since Eq. (4) directly integrates this unvalidated proxy into the planner, the reported gains in Table I could be caused by generic trajectory smoothing rather than by 'bounding distributional shifts.' A concre
  2. [Section IV-C, Table III] Table III shows that the best reconstruction-loss weight w varies by environment and that some weights worsen performance relative to the baseline (e.g., Granular with w=0.125 gives 0.422 vs. the baseline 0.391). The paper does not specify how w was chosen for the results in Table I. If w was selected per environment using the same test environments reported in Table I, the comparison is not a fair evaluation of a fixed method but rather a per-scene tuned configuration. Moreover, all numbers are means over five scenes with no error bars, standard deviations, or significance tests; the phrase 'significantly improve' in the Abstract and Section I is therefore unsupported. Please report per-scene results, error bars, and a statistical test (or at least a clear statement that the effect sizes are not statistically tested), and provide a principled procedure for selecting w that does not use
  3. [Abstract, Section IV-B, Section IV-C] The paper claims 'data efficiency' as a central outcome, but no experiment supports this claim directly. Section IV-B states that the training dataset is smaller than in prior works, and Table I compares WM-VAE to DINO-WM at a single training-data size per environment. There is no sweep over dataset sizes, no learning curve, and no comparison of the number of trajectories needed to reach a given Chamfer distance. The data-efficiency claim would require at least two or three training set sizes and a comparison of performance versus number of trajectories. As written, the results only show that, for the specific reduced dataset used, the novelty penalty lowers Chamfer distance; they do not show that the method needs less data to achieve the same performance.
  4. [Section IV-C, Table I and Table II] The comparison between WM-VAE and DINO-WM would be more convincing if the evaluation conditions were matched in detail. The paper does not report the number of CEM iterations, number of trajectory samples, elite fraction, number of seeds, or the exact training hyperparameters for each environment. Since the contribution is a change to the planner's cost function, it is important to show that the baseline and the proposed method use the same planner hyperparameters and that any gains are not due to differences in CEM tuning. Please include these details or state that they are identical to DINO-WM's defaults.
minor comments (7)
  1. [Throughout] The name 'WM-V AE' contains an unnecessary space in several places; please use 'WM-VAE' consistently.
  2. [Eq. (3)] The equation 'Lr = M SE(y, z) (3) .' has a misplaced period. Also, MSE is not defined; specify that it is the mean squared error over the latent patch dimensions.
  3. [Section III-B] The VAE architecture is described only as 'convolution layers' and 'transposed convolution layers.' Please provide the layer sizes, latent dimension, training epochs, and any regularization (e.g., KL weight) so that the novelty detector can be reproduced. The paper does not state whether the VAE is trained on flattened patch sequences or on spatially arranged patches.
  4. [Section IV-A] The environment descriptions are minimal. Please report the observation resolution, the action space bounds, the number of particles for Granular and Cloth, and the number of evaluation episodes per scene.
  5. [Section IV-C, Fig. 5] Figure 5 is not referenced in the text. It would help to refer to it explicitly when discussing the effect of the weight w.
  6. [Section III-C] The sentence 'the reconstruction loss is analogous to the regularization component of many loss functions' is vague. If this is intended as an explanation of the mechanism, please make the analogy precise or remove it.
  7. [References] Reference [9] (DINO-WM) is cited as the state of the art, but the paper does not compare with other recent world-model planners or with alternative confidence measures (e.g., ensemble disagreement, epistemic uncertainty). Adding such a comparison would strengthen the contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is an empirical heuristic with an untested proxy, not a derivation from its own inputs.

full rationale

The paper proposes adding a VAE reconstruction loss as a per-action cost in CEM planning with a learned world model. The VAE is trained on the same latent states as the world model, and its reconstruction loss (Eq. 3) is added to the trajectory cost (Eq. 4). This is a heuristic regularizer, not a derivation of a result from its inputs. The claim that high reconstruction loss indicates out-of-distribution predictions is an empirical assumption about learned representations, but it is not a tautology, a fitted parameter renamed as prediction, or a self-citation chain. The per-environment weight w in Table III is a hyperparameter chosen by the authors; selecting the best weight from a sweep does not make the reported Chamfer distances circular, though it may raise evaluation-quality concerns. The paper cites DINO-WM [9] and standard VAE references, but there are no load-bearing self-citations: the authors are not the DINO-WM authors, and the cited prior work is used as baseline and standard technique, not as an unverified premise. The Conclusion explicitly acknowledges the bias against unseen states, which further indicates that the trade-off is an admitted design choice rather than a hidden circularity. Thus, no circular step can be exhibited from the paper's equations or citations.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

No new physical or model entities are introduced. The main load-bearing input is the assumption that a VAE trained on the same latent data as the world model yields a useful reliability signal, plus the per-environment tuned weight w. All other machinery (DINOv2, ViT transition model, CEM) comes from prior work.

free parameters (1)
  • reconstruction loss weight w = 0.125 or 0.25 per environment (Table III: Granular best 0.25, Rope best 0.125, Cloth best 0.25)
    Chosen from a sweep; variance across environments shows the method is sensitive to this value.
assumptions (3)
  • domain assumption VAE reconstruction loss is a valid out-of-distribution novelty score in DINOv2 patch latent space.
    Section III-B states this without empirical validation on correlation with world-model error.
  • domain assumption DINOv2 latent states are sufficient for predicting future states and for goal comparison.
    Adopted from DINO-WM [9]; not revalidated in this paper.
  • domain assumption CEM with element-wise variance converges to good trajectories.
    Assumed from prior work [9]; no convergence analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bounding Distributional Shifts in World Modeling through Novelty Detection." pith.science (2026). https://pith.science/paper/MDBMTWCR

@misc{pith2026250806096,
  author       = {Pith},
  title        = {Pith review of: Bounding Distributional Shifts in World Modeling through Novelty Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MDBMTWCR}},
  note         = {Machine review of arXiv:2508.06096}
}
read the original abstract

Recent work on visual world models shows significant promise in latent state dynamics obtained from pre-trained image backbones. However, most of the current approaches are sensitive to training quality, requiring near-complete coverage of the action and state space during training to prevent divergence during inference. To make a model-based planning algorithm more robust to the quality of the learned world model, we propose in this work to use a variational autoencoder as a novelty detector to ensure that proposed action trajectories during planning do not cause the learned model to deviate from the training data distribution. To evaluate the effectiveness of this approach, a series of experiments in challenging simulated robot environments was carried out, with the proposed method incorporated into a model-predictive control policy loop extending the DINO-WM architecture. The results clearly show that the proposed method improves over state-of-the-art solutions in terms of data efficiency.

Figures

Figures reproduced from arXiv: 2508.06096 by the authors.

Figure 1
Figure 1. Depiction of potential action trajectories a planner [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the world model architecture, along with the novelty detection component [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Environments in which WM-VAE is evaluated. From [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Overview of the CEM planner, showing the iterative [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Graphs showing how Chamfer distance changes as reconstruction loss weight increases for various environments [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Rollout images comparing actual with predicted states in various environments. For each example, the top row [PITH_FULL_IMAGE:figures/full_fig_p006_6.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. Safety, Security, and Cognitive Risks in World Models

    cs.CR 2026-04 unverdicted novelty 6.0 of 10

    World models enable efficient AI planning but create risks from adversarial corruption, goal misgeneralization, and human bias, demonstrated via attacks that amplify errors and reduce rewards on models like RSSM and D...

Reference graph

Works this paper leans on

34 extracted references · 13 canonical work pages · cited by 1 Pith paper

  1. [1]

    World models

    D. Ha and J. Schmidhuber, “World models.” CoRR, vol. abs/1803.10122, 2018. [Online]. Available: http://dblp.uni- trier.de/db/journals/corr/corr1803.html#abs-1803-10122

    arXiv 2018

  2. [2]

    Deep learning, reinforcement learning, and world models,

    Y . Matsuo, Y . LeCun, M. Sahani, D. Precup, D. Silver, M. Sugiyama, E. Uchibe, and J. Morimoto, “Deep learning, reinforcement learning, and world models,” Neural Networks, vol. 152, pp. 267–275, 2022

    work page 2022

  3. [3]

    Estimation of inertial parameters of manipulator loads and links,

    C. G. Atkeson, C. H. An, and J. M. Hollerbach, “Estimation of inertial parameters of manipulator loads and links,” The International Journal of Robotics Research , vol. 5, no. 3, pp. 101–119, 1986. [Online]. Available: https://doi.org/10.1177/027836498600500306

    work page doi:10.1177/027836498600500306 1986

  4. [4]

    Efficient optimization for autonomous robotic manipulation of natural objects,

    A. Boularias, J. Bagnell, and A. Stentz, “Efficient optimization for autonomous robotic manipulation of natural objects,” Proceedings of the AAAI Conference , vol. 28, no. 1, Jun. 2014

    work page 2014

  5. [5]

    Offline reinforcement learning: Tutorial, review, and perspectives on open problems,

    S. Levine, A. Kumar, G. Tucker, and J. Fu, “Offline reinforcement learning: Tutorial, review, and perspectives on open problems,” 2020. [Online]. Available: https://arxiv.org/abs/2005.01643

    arXiv 2020

  6. [6]

    The class imbalance problem in deep learning,

    K. Ghosh, C. Bellinger, R. Corizzo, P. Branco, B. Krawczyk, and N. Japkowicz, “The class imbalance problem in deep learning,” Mach. Learn. , vol. 113, no. 7, p. 4845–4901, Dec. 2022. [Online]. Available: https://doi.org/10.1007/s10994-022-06268-8

    doi:10.1007/s10994-022-06268-8 2022

  7. [7]

    Fast model identification via physics engines for improved policy search,

    S. Zhu, A. Kimmel, K. E. Bekris, and A. Boularias, “Fast model identification via physics engines for improved policy search,” in Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), Stockholm, Sweden , 2018

    work page 2018

  8. [8]

    Automatic vs. manual feature engineering for anomaly detection of drinking-water quality,

    V . Fehst, H. C. La, T.-D. Nghiem, B. E. Mayer, P. Englert, and K.-H. Fiebig, “Automatic vs. manual feature engineering for anomaly detection of drinking-water quality,” in Proceedings of the Genetic and Evolutionary Computation Conference Companion , ser. GECCO ’18. New York, NY , USA: Association for Computing Machinery, 2018, p. 5–6. [Online]. Availabl...

    arXiv 2018

Show all 34 references
  1. [9]

    Dino-wm: World models on pre-trained visual features enable zero-shot planning,

    G. Zhou, H. Pan, Y . LeCun, and L. Pinto, “Dino-wm: World models on pre-trained visual features enable zero-shot planning,” 2024. [Online]. Available: https://arxiv.org/abs/2411.04983

    2024 arXiv

  2. [10]

    Exploring model-based planning with policy networks,

    T. Wang and J. Ba, “Exploring model-based planning with policy networks,” 2019. [Online]. Available: https://arxiv.org/abs/1906.08649

    2019 arXiv

  3. [11]

    On the role of planning in model-based deep reinforcement learning,

    J. B. Hamrick, A. L. Friesen, F. Behbahani, A. Guez, F. Viola, S. Witherspoon, T. Anthony, L. Buesing, P. Veliˇckovi´c, and T. Weber, “On the role of planning in model-based deep reinforcement learning,” 2021. [Online]. Available: https://arxiv.org/abs/2011.04021

    2021 arXiv

  4. [12]

    Model-based visual planning with self-supervised functional distances,

    S. Tian, S. Nair, F. Ebert, S. Dasari, B. Eysenbach, C. Finn, and S. Levine, “Model-based visual planning with self-supervised functional distances,” 2020. [Online]. Available: https://arxiv.org/abs/2012.15373

    2020 arXiv

  5. [13]

    Learning to predict vehicle trajectories with model-based planning,

    H. Song, D. Luan, W. Ding, M. Y . Wang, and Q. Chen, “Learning to predict vehicle trajectories with model-based planning,” in Proceedings of the 5th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, A. Faust, D. Hsu, and G. Neumann, Eds., vol. 164. P...

    2022

  6. [14]

    Optimal cost design for model predictive control,

    A. Jain, L. Chan, D. S. Brown, and A. D. Dragan, “Optimal cost design for model predictive control,” in Proceedings of the 3rd Conference on Learning for Dynamics and Control , ser. Proceedings of Machine Learning Research, A. Jadbabaie, J. Lygeros, G. J. Pappas, P. A. Parrilo...

    2021

  7. [15]

    Recurrent world models facilitate policy evolution,

    D. Ha and J. Schmidhuber, “Recurrent world models facilitate policy evolution,” in Advances in Neural Information Processing Systems , S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds., vol. 31. Curran Associates, Inc., 2018

    2018

  8. [16]

    Combining physics and deep learning to learn continuous-time dynamics models,

    M. Lutter and J. Peters, “Combining physics and deep learning to learn continuous-time dynamics models,” The International Journal of Robotics Research , vol. 42, no. 3, pp. 83–107, 2023. [Online]. Available: https://doi.org/10.1177/02783649231169492

    2023 doi

  9. [17]

    Towards physically interpretable world models: Meaningful weakly supervised representations for visual trajectory prediction,

    Z. Mao and I. Ruchkin, “Towards physically interpretable world models: Meaningful weakly supervised representations for visual trajectory prediction,” 2025. [Online]. Available: https://arxiv.org/abs/2412.12870

    2025 arXiv

  10. [18]

    Worlddreamer: Towards general world models for video generation via predicting masked tokens,

    X. Wang, Z. Zhu, G. Huang, B. Wang, X. Chen, and J. Lu, “Worlddreamer: Towards general world models for video generation via predicting masked tokens,” 2024. [Online]. Available: https://arxiv.org/abs/2401.09985

    2024 arXiv

  11. [19]

    Eva: An embodied world model for future video anticipation,

    X. Chi, H. Zhang, C.-K. Fan, X. Qi, R. Zhang, A. Chen, C. min Chan, W. Xue, W. Luo, S. Zhang, and Y . Guo, “Eva: An embodied world model for future video anticipation,” 2024. [Online]. Available: https://arxiv.org/abs/2410.15461

    2024 arXiv

  12. [20]

    Combating the compounding-error problem with a multi-step model,

    K. Asadi, D. Misra, S. Kim, and M. L. Littman, “Combating the compounding-error problem with a multi-step model,” 2019. [Online]. Available: https://arxiv.org/abs/1905.13320

    2019 arXiv

  13. [21]

    An analysis of frame- skipping in reinforcement learning,

    S. Kalyanakrishnan, S. Aravindan, V . Bagdawat, V . Bhatt, H. Goka, A. Gupta, K. Krishna, and V . Piratla, “An analysis of frame- skipping in reinforcement learning,” 2021. [Online]. Available: https://arxiv.org/abs/2102.03718

    2021 arXiv

  14. [22]

    Diffusion world model: Future modeling beyond step-by-step rollout for offline reinforcement learning,

    Z. Ding, A. Zhang, Y . Tian, and Q. Zheng, “Diffusion world model: Future modeling beyond step-by-step rollout for offline reinforcement learning,” 2024. [Online]. Available: https://arxiv.org/abs/2402.03570

    2024 arXiv

  15. [23]

    Variational autoencoder based anomaly detection using reconstruction probability,

    J. An and S. Cho, “Variational autoencoder based anomaly detection using reconstruction probability,” Special lecture on IE , vol. 2, no. 1, pp. 1–18, 2015

    2015

  16. [24]

    Variational autoencoder for anomaly detection: A comparative study,

    H. H. Nguyen, C. N. Nguyen, X. T. Dao, Q. T. Duong, D. P. T. Kim, and M.-T. Pham, “Variational autoencoder for anomaly detection: A comparative study,” 2024. [Online]. Available: https://arxiv.org/abs/2408.13561

    2024 arXiv

  17. [25]

    Anomaly-based intrusion detection from network flow features using variational autoencoder,

    S. Zavrak and M. ˙Iskefiyeli, “Anomaly-based intrusion detection from network flow features using variational autoencoder,” IEEE Access , vol. 8, pp. 108 346–108 358, 2020

    2020

  18. [26]

    Learning to discover anomalous spatiotemporal trajectory via open-world state space model,

    Q. Gao, C. Liu, L. Huang, G. Trajcevski, Q. Guo, and F. Zhou, “Learning to discover anomalous spatiotemporal trajectory via open-world state space model,” Knowledge- Based Systems , vol. 310, p. 112918, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/p...

    2025

  19. [27]

    Real-time anomaly detection and reactive planning with large language models,

    R. Sinha, A. Elhafsi, C. Agia, M. Foutter, E. Schmerling, and M. Pavone, “Real-time anomaly detection and reactive planning with large language models,” 2024. [Online]. Available: https://arxiv.org/abs/2407.08735

    2024 arXiv

  20. [28]

    Enhancing reconstruction-based out-of-distribution detection in brain mri with model and metric ensembles,

    E. M. C. Huijben, S. Amirrajab, and J. P. W. Pluim, “Enhancing reconstruction-based out-of-distribution detection in brain mri with model and metric ensembles,” 2024. [Online]. Available: https://arxiv.org/abs/2412.17586

    2024

  21. [29]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, A. ...

    2024 arXiv

  22. [30]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” 2021. [Online]. Available: https://arxi...

    2021 arXiv

  23. [31]

    An introduction to variational autoencoders,

    D. P. Kingma and M. Welling, “An introduction to variational autoencoders,” Foundations and Trends® in Machine Learning , vol. 12, no. 4, p. 307–392, 2019. [Online]. Available: http://dx.doi.org/10.1561/2200000056

    2019 doi

  24. [32]

    Deep convolutional inverse graphics network,

    T. D. Kulkarni, W. F. Whitney, P. Kohli, and J. Tenenbaum, “Deep convolutional inverse graphics network,” in Advances in Neural In- formation Processing Systems , C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, Eds., vol. 28. Curran Associates, Inc., 2015

    2015

  25. [33]

    Deconvo- lutional networks,

    M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus, “Deconvo- lutional networks,” in 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition , 2010, pp. 2528–2535

    2010

  26. [34]

    Neural discrete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,” 2018. [Online]. Available: https://arxiv.org/abs/1711.00937

    2018 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.