acceptodds
Under review as a conference paper at ICLR 2027

Learning Perceptual Rules from a Fragmented Visual World

Abstract

Visual perception organizes incomplete evidence into coherent contours and shapes. Gestalt principles describe regularities in this organization, but how might such expectations be acquired from visual experience? We investigate whether masked image reconstruction supports learned structural priors that guide predictions beyond local interpolation. Using controlled stimuli, we find that masked autoencoders (MAEs) trained on natural images exhibit contour continuation, symmetry-dependent reconstruction, and boundary completion in Kanizsa configurations. Changing distant elements alters the predicted content even when the visible neighborhood of the missing region remains fixed, showing that completion depends on broader spatial relationships. We then test whether visual experience changes the balance between local and global completion by continually pretraining the same initial model on natural images or matched texforms that disrupt global organization while retaining selected texture and coarse-form information. Texform adaptation weakens boundary completion and shifts reconstruction toward local continuation when local contour geometry conflicts with global symmetry, including a reversal of the average global preference for cross stimuli. These findings support an account of perceptual organization in which regularities learned from visual experience constrain predictions about missing content. The organizing tendencies expressed in reconstruction are sensitive both to the surrounding configuration and to the structural statistics encountered during learning.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Masked Autoencoders, Gestalt Principles, Perceptual Grouping, Implicit Rule Learning, Structural Representations, Visual Pattern Completion, Figure–Ground Organization, Self-Supervised Learning, Attention Mechanisms, Representation Analysis

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_qcLqOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates Gestalt-like perceptual completion in Masked Autoencoders (MAEs). The authors use controlled visual stimuli to study contour continuation, symmetry, and boundary completion. They also compare MAEs continually pretrained on natural images and texform images. The results show that training data affects the preference for local and global completion. Finally, the authors analyze the contribution of decoder attention heads to these behaviors.

Critical strengths

S1. Well-designed controlled experiments. The authors use controlled stimuli to study perceptual completion. Changing distant visual elements while keeping local evidence fixed is a useful experimental design.

S2. Interesting observations on training data. The comparison between natural-image and texform adaptation shows clear changes in completion behavior. The experiments provide quantitative results for boundary reconstruction and local-global preferences.

Critical weaknesses

W1. Limited novelty compared with existing work. Several studies have already investigated Gestalt-like behavior in neural networks. Li et al. (2025) studied Gestalt organization in self-supervised vision models, including MAEs. Kim et al. (2021) examined how natural image statistics affect Gestalt closure. Recent studies have also analyzed attention heads related to perceptual organization.The paper does not clearly explain its new findings beyond these studies. The novelty of the proposed analysis remains unclear.

W2. Insufficient evidence for the proposed mechanism. Texform adaptation changes more than global image structure. It may also affect general reconstruction quality. Therefore, the observed changes may not be caused only by weaker global perceptual priors.

Questions for authors

Q1. Novelty compared with prior work.

Several studies have already investigated Gestalt-like perception in neural networks. Could the authors clearly explain the main differences and novel contributions compared with the following works?

  1. From Local Cues to Global Percepts: Emergent Gestalt Organization in Self-Supervised Vision Models (2025).
  2. Neural Networks Trained on Natural Scenes Exhibit Gestalt Closure (2021).
  3. Does Resistance to Style-Transfer Equal Global Shape Bias? Measuring Network Sensitivity to Global Shape Configuration (2023).
  4. Locating and Editing Figure-Ground Organization in Vision Transformers (2026).
  5. I Walk the Line: Examining the Role of Gestalt Continuity in Object Binding for Vision Transformers (2026).

In particular, what new insights do the texform adaptation and attention-head analyses provide beyond these existing studies?

Q2. Training data. How can the authors confirm that the changes after texform adaptation are caused by disrupted global structure rather than reduced reconstruction quality?

Q3. Do the identified attention heads specifically support Gestalt-like completion, or are they also important for general image reconstruction? Could the authors compare their effects on perceptual completion and ordinary reconstruction tasks?

I'm willing to raise my score to 3 points after you provide a convincing answer to my question.

2 comments

Thank you for recognizing the controlled stimulus design and the training-data results. We agree that our contribution should be distinguished more clearly from earlier work on learned Gestalt-like organization.

1. What advances beyond prior work?

Our contribution concerns which missing structure an MAE generates when local and global principles disagree, and how that preference changes with visual experience. We combine these cue-conflict tests with interventions on distant context under fixed local evidence and causal tests of decoder contributions.

  • Li et al. (2025) and Kim et al. (2021): These studies establish Gestalt-like responses through internal activations, representational similarity, and global-structure sensitivity. We examine the missing structures produced by the MAE’s reconstruction process. Manipulating distant elements while preserving nearby evidence tests their influence on the predicted pixels; cue-conflict stimuli then quantify whether reconstruction more closely follows local continuation or global symmetry.

  • Wen et al. (2023) investigate sensitivity to global image configuration and introduce training that explicitly distinguishes intact from disrupted structures. We extend this question to perceptual completion: when local continuation and global symmetry predict different missing shapes, which principle guides reconstruction? Under the same masked-reconstruction objective, natural-image versus texform adaptation changes this preference, reversing the average global preference for cross stimuli. Our contribution therefore concerns how training experience changes the competing structural constraints expressed in generated completions.

  • Arnold and Gröbner (2026) and Tartaglini and Lepori (2026) identify heads supporting BEiT’s convex-completion preference and continuity-based object binding in vision transformers, including MAEs, respectively. We extend this analysis to MAE decoder computation, using bidirectional activation replacement followed by ablation on held-out contour stimuli. We further compare individual-head effects across contour continuation, Kanizsa completion, and local–global preference. Heads ranked highest for contour continuation overlap little with those most influential for Kanizsa completion or local–global preference, suggesting that different Gestalt-like completion behaviors rely on partially distinct sets of decoder heads.

2. Could texform adaptation simply reduce reconstruction quality? Both adaptation conditions start from the same checkpoint and use matched sets of 100,000 natural images or their texform counterparts for 500 epochs. Evaluation uses identical inputs and masks.

We have added non-conflict reconstruction examples in Appendix §5.10, Figure 16. In these examples, the natural-image- and texform-adapted models produce qualitatively similar completions, and the texform-adapted model successfully reconstructs missing contours following good continuation. These examples show that local contour completion remains functional after texform adaptation.

When local and global cues conflict, however, adaptation systematically changes the preferred completion. Table 2 shows that texform adaptation improves agreement with the local reference while reducing agreement with the global reference:

AdaptationLocal-reference MSE ↓Global-reference MSE ↓
Natural images0.26900.0731
Texforms0.00740.2468

Together, successful non-conflict completion and a systematic shift toward local alternatives under cue conflict argue against a uniform loss of contour-reconstruction ability. They support a change in the relative influence of local continuation and global shape constraints. The added examples are qualitative and do not establish equal reconstruction quality across natural images. Nevertheless, they strengthen the interpretation that texform adaptation changes completion preferences, consistent with coherent training-image structure helping sustain globally organized completion.

3. Are the identified heads specifically involved in Gestalt completion?

We thank the reviewer for suggesting a comparison with ordinary reconstruction. We have added this control in Appendix §5.9, Figure 15b. We individually ablated all 128 decoder heads on 400 held-out ImageNet validation images. Each image provided two nonoverlapping 33-patch gaps: one crossing an automatically detected straight boundary and one at a random location. The control shows that heads identified through Kanizsa completion also contribute to ordinary masked reconstruction, without clear evidence of a larger effect for boundary-crossing gaps. However, the overall head-ablation profiles differ between natural-image reconstruction and the artificial perceptual tasks, suggesting that the relative contributions of decoder heads depend on the completion setting.

Our perceptual-task interventions provide complementary evidence. For contour completion, heads are ranked through bidirectional activation replacement on 24 discovery instances, with the ranking fixed before ablation on 40 held-out instances. Ablating the top 16 heads increases target-patch MSE from 0.05596 to 0.07734. Cross-task comparisons further show that the most influential heads differ across contour continuation, Kanizsa completion, and local–global preference.

Together, these results support partially distinct head contributions across completion behaviors, alongside involvement in ordinary reconstruction. We therefore describe these heads as contributing to Gestalt-like completion, without claiming exclusive specialization. The diverse profile between the natural-image reconstruction and the gestalt-like perceptual completion also motivates further investigation into how computations learned through natural-image reconstruction generalize to Gestalt-like completion in artificial displays.

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

The paper asks if a masked autoencoder (MAE) acquires Gestalt-like completion rules (good continuation, symmetry, closure) from natural images, and whether those rules depend on the global structure of its training data. It has three parts. First, pretrained MAEs are given synthetic stimuli and their reconstructions of the masked region is inspected. They observe that the model can continue collinear line fragments, place a missing dot where its mirror-image counterpart predicts, and change their prediction when distant contour segments are rotated while the segments next to the mask stay fixed. Second, one MAE-B checkpoint is continually pretrained for 500 epochs on 100,000 natural images or on their texforms (images that keep local texture statistics but scramble global shape). On curved Kanizsa squares, the texform model’s boundary F1 and edge strength both drop. On stimuli where local continuation and global symmetry predict different fills, the texform model’s output moves to the local completion whereas the original models (trained on regular natural images) favor the global one. Third, the 128 decoder attention heads are ranked by activation replacement, and zeroing the top 8 to 32 heads raises contour-completion error toward the level of a disrupted stimulus. The authors conclude that MAEs learn completion rules from natural images, that these rules use the whole configuration rather than just the pixels next to the mask, and that removing global structure from training shifts completion toward local continuation.

Critical strengths
  1. The cue-conflict test is nice. In the cross and lobed-circle stimuli, both the local and the global completion fit the visible pixels exactly, so the result shows which completion the model prefers, not just whether it draws a boundary.
  2. The training experiment is also well controlled. Both models start from the same checkpoint, see the same images either in natural or in texform form, train for the same number of epochs, and are tested on identical inputs. This way the difference observed between the models can be attributed to the visual experience that they have.
  3. The writing of the paper is very clear.
Critical weaknesses
  1. The “preference for local completion” may just be a failure to complete anything. In the cross and lobed-circle stimuli, the local completion is the one where the masked region stays mostly background, while the global completion adds the arm or lobe. So a model that fills the mask with nothing automatically scores as being local. Some of these signatures is what the texform model is also showing, with its reconstruction being a gray haze. The paper therefore is not cleanly separating “prefers local continuation” from “can no longer produce structure”. Not sure if I am missing something here but even if this is the case, it would be helpful to discuss this point in the paper a bit.

  2. There is no comparison with human behavioral data and the paper relies on qualitative signatures in Sections 3.1–3.3 to emphasize “Gestalt-like”. Also closely related work is not discussed: "Human-like Object Grouping in Self-supervised Vision Transformers" by Adeli et al & "A feedforward mechanism for human-like contour integration" by Doshi et al.

  3. The head analysis for the contour completion lacks a control baseline. Do we get similar reconstruction errors when we randomly ablate the same number of heads?

Questions for authors
  1. A direct test of the local-versus-global mechanism is possible with the approach used in “Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models” which restricts patch-to-patch attention to a spatial radius and measures when configural processing breaks down. This can be also applied here: at what attention radius does the global completion collapse into local continuation, and do the heads that are isolated in Section 3.7 capture context by attending long range?

  2. Discussion in the paper can integrate the broader literature a bit more

  3. Random baseline for contour completion in the head analysis can be added in the appendix

  4. For the cross and lobed-circle stimuli, could you add a third reference in which the masked region is filled with background only, and report the error against it? This will help clarify the point I raised earlier about the model doing true local completion

2 comments

We thank the reviewer for recognizing the value of the cue-conflict design and matched training comparison. The comments help sharpen the distinction between successful local completion, loss of global completion, and nonspecific reconstruction impairment.

1. Does texform adaptation favor local completion, or simply produce background?

We thank the reviewer for raising the background-filling alternative. We added a background-only reference that preserves the visible stimulus and fills the mask uniformly with its background color. Across 144 variants, the texform-adapted model produced the following masked-region MSE:

Stimulus familyLocal-reference MSE ↓Background-only-reference MSE ↓
Crosses0.00740.1886
Bumps0.02180.4169

These results support reconstruction of the local shape rather than uniform background filling. We will incorporate this comparison into the final manuscript.

We have also added non-conflict examples in Appendix §5.10, Figure 16. In these examples, the texform-adapted model successfully reconstructs missing contours following good continuation, with qualitatively similar completions to the natural-image-adapted model. These observations demonstrate retained contour-completion ability after texform adaptation. They argue against a uniform failure to generate structure.

2. Human behavioral comparisons and broader literature

We agree that our experiments do not establish quantitative agreement with human behavior. We use “Gestalt-like” to describe specific completion signatures—good continuation, symmetry, and boundary completion—rather than equivalence to human perceptual judgments or mechanisms.

Our evidence also extends beyond visual resemblance. Changing distant elements while holding nearby evidence fixed tests contextual dependence, and the cue-conflict experiments quantify agreement with competing structural predictions. These experiments characterize model behavior, while a matched human study would be needed to establish whether humans exhibit the same preferences and sensitivity to these manipulations.

We will expand the related work to discuss both suggested papers:

  • Adeli et al. connect object-grouping structure in self-supervised vision-transformer representations to human same/different-object judgments and reaction times. Our experiments complement this behavioral approach by examining the missing structures generated during reconstruction.

  • Doshi et al. demonstrate human-like contour integration in feedforward CNNs fine-tuned for contour detection, highlighting progressively increasing receptive fields and training on gradual curvature. Our MAE experiments examine completion without contour-detection training, using the model’s learned reconstruction computation.

These connections will help distinguish the computational behaviors demonstrated here from the stronger question of human perceptual alignment.

3. Random-ablation baseline for contour completion

We agree that a discovery/evaluation split alone does not establish whether the selected heads affect contour completion more than arbitrary heads. We therefore added a contour-specific random-ablation control in Appendix §5.11 Figure 17. We fixed the decoder-head ranking using bidirectional activation replacement on 24 discovery instances and evaluated all interventions on the same 40 held-out contour instances, with the same normally fine-tuned MAE-B checkpoint and central-target RGB MSE measure. At each ablation count (k=1,4,8,16,24,32), we jointly ablated the ranked heads and compared them with 1,000 unrestricted random head orderings and 1,000 orderings matched to the ranked set's decoder-layer composition. The new appendix figure plots the fixed-ranking effect against both random baselines.

At k=16, ablating the discovery-ranked heads increased target-patch MSE from 0.05596 to 0.07734 (=0.02138). The mean increases were 0.00402 for unrestricted random sets and 0.00988 for layer-matched sets. Only 3/1,000 unrestricted sets and 11/1,000 layer-matched sets caused an increase at least as large as the ranked set (one-sided Monte Carlo tail estimates 0.0040 and 0.0120, respectively). The ranked set also exceeded most random sets at k=4,8,24; the separation was weaker at k=1,32. This supports a greater contribution of the discovery-ranked group to the measured contour reconstruction than expected from ablation size or decoder-layer composition alone.

The earlier 20-random-ordering analysis in Appendix §5.7, Figure 14 addresses Kanizsa completion; the new contour-specific control provides the comparison requested here. We interpret the result as evidence that these heads contribute to contour reconstruction under this intervention, not that they are exclusively dedicated to Gestalt completion. In the separate ordinary-reconstruction control (Appendix §5.9, Figure 15b), individual ablations of all 128 heads on 400 held-out natural images show that heads implicated in Kanizsa completion also contribute to ordinary masked-image reconstruction although the overall head ablation profile differs.

4. Attention radius and long-range contextual processing

The suggested attention-radius intervention would provide a valuable mechanistic extension. Visual Anagrams demonstrates how restricting patch-to-patch attention can reveal the contribution of long-range interactions to configural processing.

Our current experiments establish that distant input structure influences completion under fixed nearby evidence. They do not determine whether that influence is carried by direct long-range attention, propagated through successive shorter-range interactions, or already integrated into encoder representations. Likewise, identifying a head through activation replacement does not establish its attention range.

An MAE-specific radius analysis should therefore distinguish encoder and decoder restrictions: limiting decoder attention alone can leave globally integrated encoder features available. Class-token communication and propagation across layers also require consideration when interpreting the effective spatial restriction.

We have not yet measured a critical attention radius or established that the identified heads specifically transmit long-range information. We will incorporate this study into the discussion and clearly distinguish our demonstrated dependence on distant context from the unresolved computational route by which that context affects reconstruction.

Reviewer_M6ZGOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether masked autoencoders (MAEs) exhibit Gestalt-like perceptual organization. The authors use controlled visual stimuli to examine how MAEs reconstruct missing image regions and whether their predictions reflect local geometric cues or global structural information. The paper also investigates the role of training inputs by comparing MAEs further trained on natural images vs. texforms, which preserve certain local image statistics while disrupting global object structure. The results suggest that training experience influences the models’ completion preferences.

Critical strengths
  • Interesting architectural focus: Investigating Gestalt principles in MAEs is interesting, particularly because masked reconstruction provides a direct way to visualize how learned image statistics influence perceptual completion. -Attention-head interventions: The attention-head interventions are a useful attempt to identify components involved in the observed completion behaviors.
Critical weaknesses

My main concern is the paper’s positioning relative to existing literature. A substantial body of prior research has proposed that Gestalt principles reflect statistical regularities acquired through visual experience. For example, Field et al. (1993), Geisler et al. (2001), and Sigman et al. (2001) established important connections between natural-image statistics and perceptual grouping, particularly good continuation and contour integration. More recent studies have demonstrated that Gestalt-like perceptual organization can emerge in neural networks trained on natural images (Kim et al., 2021; Toosi & Miller, 2024, 2025).

In particular, Toosi and Miller (2024, 2025, 2026) have shown that neural networks trained for object recognition can exhibit illusory contours, perceptual grouping, and shape completion through the integration of implicitly learned priors. Their work also investigates how differences in training experience influence the perceptual interpretations generated by neural networks. For example, they demonstrate training-dependent responses to ambiguous visual stimuli, including different interpretations of Rubin’s vase. Their more recent work also examines Gestalt organization (many more examples) and Kanizsa stimuli.

These studies are closely related to the central motivation and findings of the present paper, but some are not discussed in the related work. I would encourage the authors to more clearly distinguish their specific contributions from these earlier findings. In my view, the paper’s strongest contribution lies in its MAE-based experimental framework and its analysis of how natural-image versus texform training affects the balance between local and global completion, rather than in establishing a novel connection between Gestalt principles and learned perceptual priors.

References

  1. Field, D. J., Hayes, A., & Hess, R. F. (1993). Contour integration by the human visual system: Evidence for a local “association field.” Vision Research, 33(2), 173–193.
  2. Geisler, W. S., Perry, J. S., Super, B. J., & Gallogly, D. P. (2001). Edge co-occurrence in natural images predicts contour grouping performance. Vision Research, 41(6), 711–724.
  3. Sigman, M., Cecchi, G. A., Gilbert, C. D., & Magnasco, M. O. (2001). On a common circle: Natural scenes and Gestalt rules. Proceedings of the National Academy of Sciences, 98(4), 1935–1940.
  4. Kim, B., Reif, E., Wattenberg, M., & Bengio, S. (2021). Neural networks trained on natural scenes exhibit Gestalt closure. Computational Brain & Behavior.
  5. Toosi, T., & Miller, K. D. (2024). Generative perceptual inference in deep neural network models of object recognition induces illusory contours and shapes. Conference on Cognitive Computational Neuroscience (CCN).
  6. Toosi, T., & Miller, K. D. (2025). Unifying Gestalt principles through inference-time prior integration. NeurIPS Workshop on Interpreting Cognition in Deep Learning Models.
  7. Toosi, T., & Miller, K. D. (2025). Generative inference unifies feedback processing for learning and perception in natural and artificial vision. bioRxiv.
  8. Toosi, T., & Miller, K. D. (2026). Feedback-mediated prior integration unifies Gestalt perceptual organization. Vision Sciences Society (VSS), conference abstract.
Questions for authors

-How does this work advance beyond previous studies demonstrating that Gestalt-like perceptual organization can emerge from learned visual priors?

-How do the natural-image versus texform experiments differ conceptually from previous demonstrations that training experience influences perceptual completion? For example, Toosi & Miller (2025) investigated how networks trained on different visual datasets, including faces, places, and ImageNet, produce different perceptual completions when presented with ambiguous stimuli.

-How can the authors establish that the observed differences between natural-image- and texform-adapted models are specifically attributable to global image structure? Although texforms preserve certain local statistics, they also differ from natural images in other ways. Could changes in reconstruction quality, image distribution, or sensitivity to particular visual features explain some of the observed differences?

2 comments

We thank the reviewer for recognizing the value of the MAE framework and attention-head interventions, and for highlighting closely related literature. We agree that connections between learned visual regularities and Gestalt organization, including experience-dependent completion in neural networks, are established. Our contribution concerns how these learned constraints guide MAE reconstruction, how they compete, and how their influence changes with visual experience.

1. How does this work advance beyond previous demonstrations of learned perceptual priors?

MAEs provide a useful setting because completion is both their self-supervised training task and their observable output. We use this setting to address four specific questions:

  1. Do distant relationships influence completion beyond nearby evidence? Moving a distant symmetric counterpart changes the reconstructed element (§3.2), and rotating distant contour segments changes completion despite unchanged neighboring fragments (§3.3). These interventions directly establish dependence on relationships beyond the gap’s local neighborhood.
  2. Which principle guides completion when plausible constraints disagree? Our cross and lobed-circle stimuli make local continuation and global symmetry predict different missing structures compatible with identical visible pixels (§3.6). Reconstruction error measures which alternative the model more closely follows. The original MAE favors global symmetry on average in these tests, particularly for cross stimuli.
  3. How does training experience change this preference? Matched natural-image and texform continued training shifts the balance between these alternatives. Texform adaptation moves both shape families toward local continuation and reverses the average global preference for cross stimuli (§3.6). This reveals a change in the preferred completion geometry, beyond weaker boundary reconstruction alone.
  4. Which decoder components contribute to these predictions? Activation replacement followed by ablation on held-out stimuli links contour completion to decoder computation (§3.7). Comparisons across tasks show that component importance depends on the completion task and evaluation metric (§3.8), without establishing dedicated heads for individual Gestalt rules. These experiments complement Toosi and Miller’s generative-inference framework, which elicits perceptual interpretations through iterative feedback in recognition networks. Our study examines missing structures generated directly through standard MAE encoder–decoder computation, using self-supervised reconstruction without recognition labels or additional inference-time optimization. This provides a complementary setting for studying how learned spatial relationships guide completion.

We will incorporate the relevant references and revise the Introduction and Conclusion to make this positioning and these specific contributions explicit.

2. How does natural-image versus texform adaptation differ from comparing training domains?

We agree that comparisons of face-, scene-, and object-trained networks already demonstrate that visual experience influences perceptual interpretation. Our experiment asks a more targeted question: how does transforming corresponding training images to disrupt object structure change the balance between local continuation and global symmetry?

Both adaptation conditions start from the same pretrained MAE checkpoint and use 100,000 corresponding natural images or texforms for 500 epochs. Architecture and reconstruction objective are held fixed, and evaluation uses identical inputs and masks. Texforms disrupt recognizable object structure while retaining selected texture and coarse-form information. Using transformations of the same source-image pool reduces differences in dataset composition that accompany comparisons across semantic domains.

The resulting change concerns which missing structure the model predicts: texform adaptation reverses the average global preference for cross stimuli and shifts lobed-circle completions toward local continuation (§3.6, Figure 7). Thus, our contribution is a matched comparison linking a transformation of training-image structure to a directional change in completion preference, rather than another demonstration that training experience matters.

3. Could reconstruction quality or other distributional changes explain the texform results?

To address this concern, we have added qualitative comparisons in non-conflict settings (Appendix 5.10, Figure 16). In the illustrated examples, natural-image- and texform-adapted MAEs produce visually similar, coherent completions consistent with good continuation. Thus, the texform-adapted model retains the ability to reconstruct missing contours when the available cues support a consistent completion.

The contrast becomes apparent when local continuation and global symmetry favor different structures. Both alternatives are compatible with identical visible pixels, so the experiment measures which completion is favored, rather than accuracy against a uniquely specified target. Averaged across 72 cross variants:

AdaptationLocal-reference MSE ↓Global-reference MSE ↓
Natural images0.26900.0731
Texforms0.00740.2468

After texform adaptation, the reconstruction closely matches the locally continued alternative. The lobed-circle stimuli show the same direction of preference change.

Together, the non-conflict examples and cue-conflict results support a change in completion preference rather than an across-the-board inability to reconstruct contours. Locally coherent completion remains possible, while the preference for global symmetry weakens when it competes with local continuation.

These observations strengthen the interpretation that coherent relationships in natural training images help sustain global completion preferences. Because texforms also alter other image properties, the experiment does not uniquely identify the responsible statistics; our conclusion concerns a training-dependent shift in the balance between completion principles.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.