Pith. sign in

REVIEW 2 major objections 4 minor 44 references

Support Operation Factorization: Compositional Readout of Frozen Vision Encoders under Controlled Interventions

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a factored readout of frozen patch-token innovation fields recovers held-out support–operation bindings on Shapes3D-Extended and COCO under injective scoring, while the same protocol on rebuilt MuJoCo exposes a…

desk verdict Injective leave-one-cell-out probing is a genuinely useful addition, but the flagship carrier comparison in Table 3 lacks the per-seed statistics to back its headline number. read the letter →

arxiv 2608.06174 v1 pith:EBTRBEDD submitted 2026-08-06 cs.CV

Zhongyao Wang , Wanli Ouyang , Taoyong Cui , Pheng Ann Heng This is my paper · ORCID
classification cs.CV
keywords compositionalreadoutfrozenvisionencoderssupport-operationfactorizationinnovationfieldinjectiveevaluationoperationlaunderingleave-one-cell-outprotocolgeneralization
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

This paper tries to establish that a frozen vision transformer's patch-token innovation field—the difference between an intervened image's features and its source image's features—can support a compositional readout that binds a spatial support (where the edit happened) to an operation (what the edit was). The claim is that a factored readout, SO-OPF, recovers held-out support–operation bindings on two controlled substrates: 0.874 injective accuracy with known factorial assignment and 0.769 when the assignment must be learned from flat labels on Shapes3D-Extended, and 0.799 and 0.762 on globally image-disjoint COCO. The paper also argues that an injectively aligned, Hungarian-based scoring protocol is necessary to expose operation laundering, in which two operations are credited through the same slot; under matched axis-aware supervision on Shapes3D, the factored carrier raises learned-assignment accuracy from 0.653 for a dense carrier to 0.841 and eliminates the dense carrier's laundering gap. On a rebuilt MuJoCo substrate the same protocol fails, with learned-assignment accuracy reaching only 0.569 with DINOv3 and 0.484 with SigLIP2, so the paper bounds its own claim: compositional recovery is not universal across renderers or material-space operations. A sympathetic reader should care because standard factor probes score axes separately and can silently reward slot sharing, whereas this work offers a recovery protocol, an explicit failure boundary, and a metric that distinguishes true composition from slot reuse.

What carries the argument

The central object is SO-OPF, a factored readout that maps the innovation field $\Delta z = \Phi(x_{\mathrm{int}}) - \Phi(x_{\mathrm{src}})$ to a support×operation energy grid through $E_{s,o} = A_s \, \pi_{s,o}$. A learned router assigns patch tokens to support slots, $A_s$ pools per-support squared-norm salience, and $\pi_{s,o}$ is a competitive operation posterior computed from scale-normalized innovation token features by a small MLP. The evaluation machinery is a leave-one-cell-out protocol over an $|S|\times|O|$ factorial grid, with a Hungarian injective assignment that forces each ground-truth axis value to claim a distinct energy slot, thereby detecting operation laundering that many-to-one alignment would conceal.

What would settle it

Render the same MuJoCo material-space hue and inversion edits as pixel-space overlays with matched apparent magnitude, keep every other protocol detail fixed, and measure injective accuracy for those cells under known factorial assignment with DINOv3: if the hue and inversion cells climb from their observed near-chance range (0.09 to 0.28) to above 0.7 while their material-space twins stay low, the failure boundary lies in the subtraction signal for material edits rather than in the readout or the encoder; if they stay low, the bottleneck sits elsewhere.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that whether a frozen encoder composes held-out support×operation bindings and whether the factorial grid can be recovered from flat cell labels are separate questions, and SO-OPF answers the first affirmatively on two substrates without establishing the second generally. With frozen DINOv3 features, known factorial assignment reaches 0.874 injective accuracy on Shapes3D-Extended and 0.799 on globally image-disjoint COCO, while learning the assignment from flat labels reaches 0.769 and 0.762, respectively. The COCO separation replicates with SigLIP2 (factorial 0.826 versus learned 0.751), and learned assignment recovers an equivalent grid in about 90–96% of fits on the successful substrates. On rebuilt MuJoCo, pattern cells score near perfect while material-space hue and inversion cells score near chance, and learned-assignment accuracy drops from 0.790 under many-to-one alignment to 0.569 under injective scoring with DINOv3, from 0.629 to 0.484 with SigLIP2. The paper therefore claims that factored readout plus injective evaluation recovers held-out bindings on two substrates while exposing, rather than hiding, a renderer- and operation-specific failure boundary, and it explicitly does not claim universal recovery from flat labels.

Load-bearing premise

The readout sees only the difference between the intervened image's features and the source image's features, so the whole result rests on the assumption that this subtraction retains enough spatially localized, operation-discriminating information in the frozen patch tokens; if subtracting an edit strips its identity for a given operation, no readout can recover that cell.

Editorial extensions

If this is right

  • Known factorial assignment alone gives near-clean composition on Shapes3D (0.874) and COCO (0.799), so frozen encoder features are sufficient carriers for spatially localized edits on those domains.
  • When assignment must be learned from flat labels, expect a measurable cost: learned injective accuracy trails factorial by about 0.1 on Shapes3D and about 0.04 to 0.08 on COCO, with imperfect grid recovery in 90 to 96 percent of fits.
  • Many-to-one accuracy should be treated as suspect: on MuJoCo the learned assignment scores 0.790 under many-to-one alignment but drops to 0.569 under injective scoring, revealing slot collapse that the aggregate number conceals.
  • Any claim about compositional readout must name its substrate: material-space operations on a rebuilt MuJoCo renderer fail even with known factorial assignment, while pixel-space overlay cells remain near perfect.
  • The SigLIP2 replication on COCO and MuJoCo indicates that the carrier/assignment separation and the renderer boundary are not artifacts of a single frozen encoder.

Reading between the lines

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

  • I infer that the same leave-one-cell-out plus injective scoring protocol could act as a general diagnostic for factor probes: any probe reporting aggregate accuracy without an injective alignment may contain unreported slot sharing on some substrates.
  • A natural next test would feed the readout a whitened or per-token normalized innovation field on the MuJoCo material-space cells; if hue and inversion cells recover, the bottleneck lies in signal scale rather than in the encoder or the readout.
  • Extending the protocol from predefined supports to discovered spatial decompositions, and from low-level appearance edits to semantic operations such as object removal, would test whether compositional binding is a property of the encoder or of the intervention class; the paper lists these as future directions.
  • If renderer-diverse pretraining repairs the MuJoCo boundary, the paper's negative result would become a training-data recommendation rather than a fixed structural limit; that would sharpen the claim's scope without overturning it.
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

2 major / 4 minor

Summary. The paper proposes a compositional readout for frozen vision encoders based on a support×operation factorization. It defines the innovation field Δz = Φ(x_int) − Φ(x_src), a leave-one-cell-out protocol over support×operation grids, and an injective Hungarian-scored metric intended to penalize operation laundering. The SO-OPF readout factors cell energy into support salience and a competitive operation posterior, and is evaluated under three assignment semantics: known factorial Q, learned Q from flat cell labels, and a compressed identity control. Experiments on Shapes3D, a rebuilt MuJoCo substrate, and globally image-disjoint COCO with frozen DINOv3 and SigLIP2 report that factorial assignment reaches 0.874 and 0.799 injective accuracy on Shapes3D and COCO, learned assignment reaches 0.769 and 0.762, and that under matched axis-aware supervision the factored carrier improves learned-Q accuracy from 0.653 to 0.841 over a dense carrier. The MuJoCo substrate exhibits a failure boundary with substantial slot collapse. The paper explicitly disclaims identifiability and universal recovery from flat labels.

Significance. If the results hold, the main contributions are a metric that separates carrier capacity from assignment recovery and exposes operation laundering, a factored readout that outperforms a dense carrier, and an honestly documented failure boundary on MuJoCo. Strengths include the clean separation of factorial, learned, and compressed assignments; paired t-tests for assignment-semantics comparisons in Table 1; a second-encoder replication on COCO and MuJoCo; collapse-threshold robustness checks; and detailed reproducibility information. The main caveat is that the flagship carrier comparison (C2 and the abstract's '0.653 to 0.841' claim) is reported without per-seed variability or a significance test, so the quantitative strength of the central contribution is not yet established.

major comments (2)
  1. [Experiments, Table 3] The abstract's central carrier claim—'the factored carrier improves learned-assignment accuracy from 0.653 to 0.841 over a dense carrier and eliminates its laundering gap'—rests entirely on the enhanced-protocol rows of Table 3, whose caption states 'Variability is omitted from this compact carrier comparison.' No per-seed values, standard deviations, or significance tests are reported for the factored-versus-dense difference, and the paired tests in Table 1 cover assignment semantics, not carrier architecture. This single-point gap is the entire basis for contribution C2 and for the corresponding sentence in the abstract. Please report the ten per-seed learned-enhanced injective accuracies for both carriers, the across-seed SD, and a paired test (or bootstrap interval) for the difference; if the 0.841 vs 0.653 gap is not statistically robust, the quantitative claim and the 'eliminates its laundering gap' statement must be revised accordingly. This is checkable from the existing per-seed fits and does not require new data.
  2. [Experiments, 'Known Structure and Learned Structure'] The text reports Shapes3D learned-assignment raw and injective scores of 0.780 and 0.769 and a paired factorial-minus-learned raw difference of 0.095 (t = 4.01, p = 0.0031), but Table 1 contains only injective values and no raw-score means or SDs. As a result, the reader cannot verify whether the reported raw difference and its p-value are consistent with the tabulated numbers. Please include raw-accuracy means and SDs for all assignment arms (or a note that the raw scores are in the supplementary material) so the significance statement is reproducible.
minor comments (4)
  1. [Table 2 caption] The caption says variability is omitted and reported for the assignment comparison in Table 1, but this applies only to the SO-OPF rows; the bilinear, MLP, attention, oracle, random-init, and pixel baselines have no reported variability anywhere in the paper, so the text's ordering of these baselines should include SDs or be explicitly flagged as preliminary.
  2. [Eq. (4)] The support router in Eq. (4) uses only the source token map z_src and positional embeddings, while the readout's input is defined as Δz in Eq. (1); please clarify whether the router intentionally ignores the intervened image and how this choice affects localization for edits that alter the support region.
  3. [Figures 2 and 3] Figure 2's caption refers to 'the remaining panels' and Figure 3's caption to 'a green row beneath the axis'; these references are not self-explanatory in the current rendering, and the figures would benefit from explicit panel labels or a legend.
  4. [Appendix C] The main text reports only the collapse threshold of 0.5, while Appendix C shows stability across thresholds 0.4–0.6; consider stating the threshold-stability result in the main text, since it is a useful robustness check for the MuJoCo boundary.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the held-out-cell protocol, optimization-only Hungarian alignment, flat-label learned assignment, and explicit disclaimers keep the derivation self-contained.

full rationale

The paper's derivation chain is self-contained and its empirical claims are not equivalent to their inputs by construction. The innovation field Δz = Φ(x_int) − Φ(x_src) (Eq. 1) is a fixed input; the readout is trained on observed cells and tested only on the held-out cell, with the Hungarian injective alignment (Eq. 3) estimated from optimization-pair energies and never from held-out pairs. The learned-Q arm maps flat observed-cell labels to grid locations (Eq. 9) and is evaluated by injective cell accuracy, while compressed identity acts as a control showing label order alone does not supply the grid. The factored-vs-dense carrier comparison (Table 3) is an architectural comparison under matched axis-aware supervision, not a fitted parameter renamed as a prediction; the dense carrier (Eq. 12) is a distinct, learnable baseline, and the improvement is an empirical outcome rather than a tautology. The paper explicitly disclaims identifiability and flat-label discovery in the Limitations: 'No experiment here constitutes unsupervised representation learning or proves equivalence to known structure,' and the MuJoCo boundary is reported as a failure rather than hidden. The only concerns raised by a skeptical reading—the absence of per-seed error bars in Table 3 and the assumption that Δz preserves operation-discriminative information—are correctness or statistical-evidence risks, not circularity. No load-bearing self-citation or imported uniqueness claim appears in the derivation.

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

The central claims rest on the frozen encoder's innovation field being informative, on the factored carrier assumption, and on a set of hand-chosen hyperparameters (temperature schedule, loss scale, weights, collapse threshold). None of these are fit to held-out test data. The assignment Q is learned in the learned arm, which is an honest part of the protocol. No new physical entities are introduced; 'operation laundering' is a diagnostic concept, not an entity. The data generation ranges are experimental design choices rather than free parameters in the derivation.

free parameters (6)
  • Energy logit scale = 10
    Introduced in Eq. (9) to prevent energy-share logits from staying near-uniform; chosen by hand, not fit to test data.
  • Temperature annealing schedule (τ) = 2.0 to 0.5
    In Eq. (7), τ anneals from 2.0 to 0.5 to sharpen the operation posterior; chosen by hand.
  • Support-use entropy weight λ_mask = 1
    In Eq. (10), unit weight on negative entropy of mean support usage; chosen by hand.
  • Enhanced loss weights and margin = 1, 0.1, 0.5, 1, ρ=0.6
    In Eq. (11): L_loc weight 1, L_bal 0.1, L_axis 0.5, L_opinj 1, localization margin ρ=0.6; hand-chosen for the axis-aware protocol.
  • Collapse threshold = 0.5
    Predeclared threshold for collapse rate; robustness checked at 0.4 and 0.6 in the supplementary material.
  • Optimization schedule = lr=5e-3, 2000 steps, 5 restarts, 64/16/20 split
    From 'Features and optimization' and Appendix A; standard choices, not fit to test data.
assumptions (6)
  • standard math Hungarian algorithm returns a valid bijective alignment maximizing marginal energy assignment (Kuhn 1955).
    Invoked in Eq. (3) and Algorithm 1 for injective evaluation.
  • domain assumption The innovation field Δz = Φ(x_int) − Φ(x_src) is a sufficient signal for both support location and operation identity.
    Eq. (1); the sole input to the readout. The MuJoCo results show this can fail for material-space operations.
  • domain assumption Supports are predefined spatial regions and operations are predefined appearance edits; grid dimensions and flat cell labels are known in every experiment.
    Sections 'Supports and operations' and 'Assignment Semantics'; explicitly stated as known.
  • domain assumption The factored energy form E_{s,o} = A_s π_{s,o} is an adequate carrier separating support salience from operation identity.
    Eqs. (5)-(8); the central architectural assumption, tested against a dense carrier.
  • domain assumption Frozen DINOv3 and SigLIP2 patch tokens preserve the information needed for compositional readout.
    'Frozen encoder and token maps'; the encoders are never fine-tuned.
  • domain assumption Paired-sample t-tests over ten seeds are valid for comparing assignment modes.
    'Reporting' section; assumes seed-level means are approximately normal and independent across modes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Support Operation Factorization: Compositional Readout of Frozen Vision Encoders under Controlled Interventions." pith.science (2026). https://pith.science/paper/EBTRBEDD

@misc{pith2026260806174,
  author       = {Pith},
  title        = {Pith review of: Support Operation Factorization: Compositional Readout of Frozen Vision Encoders under Controlled Interventions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EBTRBEDD}},
  note         = {Machine review of arXiv:2608.06174}
}
read the original abstract

Compositional analysis of frozen vision encoders should determine both what changed and where it changed. Standard factor probes score these axes separately, however, and can reward multiple operations that reuse the same predicted slot. We call this failure operation laundering. We introduce an injectively aligned leave-one-cell-out protocol over support x operation grids and SO-OPF, a readout that factors cell energy into support salience and a competitive operation posterior. This formulation separates two questions that aggregate scores conflate: whether the carrier composes held-out bindings when the grid is known, and whether that grid can be recovered from flat cell labels. With frozen DINOv3 features, known factorial assignment reaches 0.874 injective accuracy on Shapes3D-Extended and 0.799 on globally image-disjoint COCO; learning the assignment from flat labels reaches 0.769 and 0.762, respectively. Under matched-axis-aware supervision on Shapes3D, the factored carrier improves learned-assignment accuracy from 0.653 to 0.841 over a dense carrier and eliminates its laundering gap. SigLIP2 replicates the COCO separation. A rebuilt MuJoCo substrate exposes a boundary: learned-assignment accuracy is 0.569 with DINOv3 and 0.484 with SigLIP2, with substantial slot collapse. Thus factored readout and injective evaluation recover held-out bindings on two substrates while exposing, rather than hiding, a renderer-specific failure boundary; they do not establish universal recovery from flat labels.

Figures

Figures reproduced from arXiv: 2608.06174 by the authors.

Figure 1
Figure 1. Evaluation, carrier, and assignment are separate design choices. a, Leave-one-cell-out evaluation withholds one binding while retaining its support and operation in other cells. b, SO-OPF factors support salience from a competitive operation posterior. c, The structural input each assignment arm receives. Only the learned arm tests structure recovery without the axis taxonomy; compressed identity is a control that m… view at source ↗
Figure 2
Figure 2. MuJoCo support-router diagnostic. Exact support masks are shown above three hue-intervention examples; the remaining panels show the learned support slots. Green frames mark the intervened support. After matching slots to supports, mean mask IoU is 0.821 and assignment agreement is 0.903 over 630 held-out examples. This panel diagnoses support localization; it is not a learned-Q factorial-assignment result [PITH_FU… view at source ↗
Figure 3
Figure 3. DINOv3 assignment comparison. a, Injective cell accuracy for factorial, learned, and compressed-identity assignments. Each seed is first averaged over nine held-out cells; small dots are the ten seed means, and markers with bars are their mean and population SD. Compressed identity is a Shapes3D-only label-order control. The dashed rule marks chance (1/9). b, Collapse rate over 90 seed–cell fits; the green row benea… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 28 canonical work pages

  1. [1]

    Higgins, Irina and Matthey, Loic and Pal, Arka and Burgess, Christopher and Glorot, Xavier and Botvinick, Matthew and Mohamed, Shakir and Lerchner, Alexander , booktitle=. -

    work page

  2. [2]

    International Conference on Machine Learning , pages=

    Disentangling by Factorising , author=. International Conference on Machine Learning , pages=

    work page

  3. [3]

    International Conference on Machine Learning , pages=

    Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations , author=. International Conference on Machine Learning , pages=

  4. [4]

    International Conference on Learning Representations , year=

    The Role of Disentanglement in Generalisation , author=. International Conference on Learning Representations , year=

    work page

  5. [5]

    International Conference on Learning Representations , year=

    Visual Representation Learning Does Not Generalize Strongly Within the Same Domain , author=. International Conference on Learning Representations , year=

    work page

  6. [6]

    International Conference on Machine Learning , pages=

    On Disentangled Representations Learned from Correlated Data , author=. International Conference on Machine Learning , pages=

    work page

  7. [7]

    OpenReview preprint , year=

    A Path Towards Autonomous Machine Intelligence , author=. OpenReview preprint , year=

    work page

  8. [8]

    Conference on Computer Vision and Pattern Recognition , pages=

    Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture , author=. Conference on Computer Vision and Pattern Recognition , pages=

    work page

Show all 44 references
  1. [9]

    arXiv preprint arXiv:2404.08471 , year=

    Revisiting Feature Prediction for Learning Visual Representations from Video , author=. arXiv preprint arXiv:2404.08471 , year=

    arXiv

  2. [10]

    Transactions on Machine Learning Research , year=

    Oquab, Maxime and Darcet, Timoth. Transactions on Machine Learning Research , year=

  3. [11]

    arXiv preprint arXiv:2508.10104 , year=

    Sim. arXiv preprint arXiv:2508.10104 , year=

    arXiv

  4. [12]

    arXiv preprint arXiv:2403.00504 , year=

    Learning and Leveraging World Models in Visual Representation Learning , author=. arXiv preprint arXiv:2403.00504 , year=

    arXiv

  5. [13]

    Advances in Neural Information Processing Systems , pages=

    Object-Centric Learning with Slot Attention , author=. Advances in Neural Information Processing Systems , pages=

  6. [14]

    International Conference on Learning Representations , year=

    Conditional Object-Centric Learning from Video , author=. International Conference on Learning Representations , year=

  7. [15]

    Advances in Neural Information Processing Systems , year=

    Simple Unsupervised Object-Centric Learning for Complex and Naturalistic Videos , author=. Advances in Neural Information Processing Systems , year=

  8. [16]

    International Conference on Machine Learning , pages=

    Generalization without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks , author=. International Conference on Machine Learning , pages=

  9. [17]

    International Conference on Learning Representations , year=

    Measuring Compositional Generalization: A Comprehensive Method on Realistic Data , author=. International Conference on Learning Representations , year=

  10. [18]

    Advances in Neural Information Processing Systems , volume=

    Compositional Generalization from First Principles , author=. Advances in Neural Information Processing Systems , volume=

  11. [19]

    International Conference on Learning Representations , year=

    Recurrent Independent Mechanisms , author=. International Conference on Learning Representations , year=

  12. [20]

    Proceedings of the IEEE , volume=

    Toward Causal Representation Learning , author=. Proceedings of the IEEE , volume=

  13. [21]

    Microsoft

    Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll. Microsoft. European Conference on Computer Vision , pages=

  14. [22]

    Understanding Disentangling in -

    Burgess, Christopher P and Higgins, Irina and Pal, Arka and Matthey, Loic and Watters, Nick and Desjardins, Guillaume and Lerchner, Alexander , journal=. Understanding Disentangling in -

  15. [23]

    Burgess, Chris and Kim, Hyunjik , howpublished=

  16. [24]

    International Conference on Computer Vision , pages=

    Emerging Properties in Self-Supervised Vision Transformers , author=. International Conference on Computer Vision , pages=

  17. [25]

    Bardes, Adrien and Ponce, Jean and LeCun, Yann , booktitle=

  18. [26]

    Kuhn, Harold W , journal=. The

  19. [27]

    Todorov, Emanuel and Erez, Tom and Tassa, Yuval , booktitle=

  20. [28]

    Nonlinear

    Hyv. Nonlinear. International Conference on Artificial Intelligence and Statistics , pages=

  21. [29]

    Advances in Neural Information Processing Systems , year=

    Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style , author=. Advances in Neural Information Processing Systems , year=

  22. [30]

    International Conference on Learning Representations (Workshop) , year=

    Understanding Intermediate Layers Using Linear Classifier Probes , author=. International Conference on Learning Representations (Workshop) , year=

  23. [31]

    Computational Linguistics , volume=

    Probing Classifiers: Promises, Shortcomings, and Advances , author=. Computational Linguistics , volume=

  24. [32]

    arXiv preprint arXiv:2502.14786 , year =

    Michael Tschannen and Alexey Gritsenko and Xiao Wang and Muhammad Ferjad Naeem and Ibrahim Alabdulmohsin and Nikhil Parthasarathy and Talfan Evans and Lucas Beyer and Ye Xia and Basil Mustafa and Olivier H. arXiv preprint arXiv:2502.14786 , year =

    arXiv

  25. [33]

    Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (

    Kim, Been and Wattenberg, Martin and Gilmer, Justin and Cai, Carrie and Wexler, James and Viegas, Fernanda and Sayres, Rory , booktitle =. Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (

  26. [34]

    IEEE Conference on Computer Vision and Pattern Recognition , pages =

    Network Dissection: Quantifying Interpretability of Deep Visual Representations , author =. IEEE Conference on Computer Vision and Pattern Recognition , pages =

  27. [35]

    and Matthey, Loic and Watters, Nicholas and Kabra, Rishabh and Higgins, Irina and Botvinick, Matt and Lerchner, Alexander , journal =

    Burgess, Christopher P. and Matthey, Loic and Watters, Nicholas and Kabra, Rishabh and Higgins, Irina and Botvinick, Matt and Lerchner, Alexander , journal =

  28. [36]

    International Conference on Machine Learning , pages =

    Multi-Object Representation Learning with Iterative Variational Inference , author =. International Conference on Machine Learning , pages =

  29. [37]

    arXiv preprint arXiv:1812.02230 , year =

    Towards a Definition of Disentangled Representations , author =. arXiv preprint arXiv:1812.02230 , year =

    arXiv

  30. [38]

    Neural Computation , volume =

    Separating Style and Content with Bilinear Models , author =. Neural Computation , volume =

  31. [39]

    Conference on Empirical Methods in Natural Language Processing and the International Joint Conference on Natural Language Processing , pages=

    Designing and Interpreting Probes with Control Tasks , author=. Conference on Empirical Methods in Natural Language Processing and the International Joint Conference on Natural Language Processing , pages=

  32. [40]

    International Conference on Machine Learning , year=

    Learning Transferable Visual Models From Natural Language Supervision , author=. International Conference on Machine Learning , year=

  33. [41]

    Artificial Intelligence , volume=

    Tensor Product Variable Binding and the Representation of Symbolic Structures in Connectionist Systems , author=. Artificial Intelligence , volume=

  34. [42]

    Subspace-Decomposed

    Thil, Lucas and Read, Jesse and Kaddah, Rim and Doquet, Guillaume , journal=. Subspace-Decomposed

  35. [43]

    Kuang, Yilun and Dagade, Yash and Rudner, Tim G. J. and Balestriero, Randall and LeCun, Yann , journal=. Rectified

  36. [44]

    IEEE Conference on Computer Vision and Pattern Recognition , pages=

    From Red Wine to Red Tomato: Composition with Context , author=. IEEE Conference on Computer Vision and Pattern Recognition , pages=

Pith tools

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