acceptodds
Under review as a conference paper at ICLR 2027

Visualizing Dense Neural Representations

Abstract

Modern dense-feature extractors such as DINOv3 or Perception Encoder achieve strong performance on a variety of spatially fine-grained tasks. For a qualitative understanding, the features are usually mapped onto the three first principal components which are treated as RGB values. Surprisingly, the noisiness of these visualizations varies substantially for models with similar downstream performance, suggesting a much larger gap in feature quality than indicated by quantitative evaluation. We propose to fix this issue by using an alternative visualization method based on the Laplace operator that inherits all advantages of PCA, while taking the spatial structure of the data into account. We first show that models with noisy PCA feature maps indeed contain linear components with substantially less noise, qualitatively explaining the relatively modest performance differences on dense tasks. Interestingly, the method reveals an opposite problem for the DINOv3 model: PCA visualizations overestimate the spatial smoothness of the model's representations by encoding a strong spatial bias. The proposed Laplacian PCA effectively removes this bias, revealing additional semantically meaningful structure. Quantitatively, we demonstrate that for all tested models linear components identified by Laplacian PCA align better with dimensions relevant for semantic segmentation and monocular depth estimation.

Primary area:
interpretability and explainable AI
Keywords:
visualization, dense feature models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jj3WOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose a method for qualitative evaluation of dense visual representations. They suggest that the standard PCA-to-RGB visualization can be misleading. Their method, based on Laplacian principal component analysis, results in more spatial consistency without additional complexity by adding a spatial smoothness term to PCA. They compare their method against plain PCA on semantic segmentation and depth estimation tasks by measuring how much performance survives when the features are projected.

Critical strengths

These visualizations are commonly used, so the observation that PCA to RGB can result in clean linear components for noisy looking models and vise versa for DINO, is valuable.

Critical weaknesses
  1. I find the novelty of the method limited. The penalized form of the smoothness term is new; however, the idea of a PCA objective plus a spatial smoothness term has been introduced in methods like LPP.
  2. On the experimental side, the only baseline is PCA, and the authors mention that they pick α\alpha on test data. PCA is a special case of the proposed method (α=0\alpha = 0), the gains with fixed α\alpha are not consistent, and the absence of error bars also makes me suspicious.
Questions for authors

Can you add a couple more baselines?

2 comments

Thank you for your valuable feedback. We also appreciate the encouraging scores on the sub-categories. We have improved the manuscript in terms of baseline comparisons and experimental rigor, and formulated the novelty of the method more clearly. We believe that incorporating your feedback has strengthened the paper quite a bit and would appreciate it if you could take the time to re-evaluate. Specifically:

I find the novelty of the method limited. The penalized form of the smoothness term is new; however, the idea of a PCA objective plus a spatial smoothness term has been introduced in methods like LPP.

We now establish the novelty more clearly in the text by incorporating a new table (Tab. 1), which more explicitly highlights differences to previous methods. We reference the table in the related works section. For example, while LPP also relies on the graph Laplacian to reduce dimensionality, it has no notion of spatial contiguity as we explore in this work. The table also states more clearly why the method proposed by Jiang et al. (2013), which is most similar to the one discussed here, is unsuitable for our application as it does not provide a parametric map from features to visualization. This is also discussed in the new 'Baselines' paragraph (4.2; see below.). We also now state the novelty more explicitly with respect to the negative regularization that is necessary to analyze the DINOv3 features; the corresponding paragraph now cleanly mentions that this discovery is not possible with previous methods.

On the experimental side, the only baseline is PCA

We previously did not include the prior method by Bolya et at. (2025) because of the conceptual arguments against it provided in the text. However, for completeness, we now include it as an additional quantitative baseline. We incorporate a dedicated paragraph on baselines in section 4.2, explicitly describing this choice and why other candidates (such as Jiang et al. (2013)) are not suitable for comparison (i.e. the lack of a parametric map). We agree that this was not communicated clearly enough before. Comparisons to the prior method by Bolya et at. (2025) show that both outperform PCA, while Laplacian PCA demonstrates clear advantages over both baselines.

the authors mention that they pick on test data.

You are likely referring to Table 2, which also showed performance for the best value of α\alpha for each model. We agree that this was confusing and have removed that column. These results are now presented in Table 3, together with the new baseline method. All comparisons are now without tweaking any hyperparameters.

the gains with fixed α\alpha are not consistent ;the absence of error bars also makes me suspicious.

We used the rebuttal time to compute error bars for all relevant quantitative comparisons to baselines. These are now included in the tables, as well as the line plots. Since we only evaluate linear probes, error margins are quite small, and improvements over PCA, as well as the method proposed by Bolya et al., are consistent for the vast majority of comparisons. In case you have further concerns regarding this point, we would appreciate a brief update.

Thank you again for your helpful suggestions.

Thank you for the additional experiments and the new comparison table. They address part of my concern. My remaining concern is novelty. The clarifications improve the positioning, but the core objective (PCA with a graph-Laplacian smoothness term) remains close to LPP and Jiang et al. (2013). the negative regularization is the most novel aspect, but I'm not convinced it is sufficient on its own. Given the improved experiments, I therefore raise my score from 1 to 2.

Reviewer_q3jfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates PCA visualisations of vision model single image representations and proposes Laplacian PCA as an alternative. PCA finds directions of high variance within an image, which could reflect localised objects, texture, global information, or mixtures of these. This work proposes regularisation such that directions whose activations vary smoothly across neighbouring image patches are favoured. It evaluates the selected components through downstream probes. For DINOv3, they find that negative regularisation is beneficial, which suggests that the leading PCA components are already too spatially smooth for this model.

Critical strengths

This work quantifies observations that multiple previous studies (Simeoni et al., 2025; Bolya et al., 2025) have made, thereby clarifying characteristics of PCA visualisation. I see its main contribution in characterising the properties a per-image, low-dimensional feature visualisation method should have in Sec. 3.1: Property 1: the projection should be linear; Property 2: applying the projection should not mix tokens.

Critical weaknesses

Weakness 1 – Limited novelty and demonstration of utility:

The problem of noisy PCA visualisations has already been pointed out by Bolya et al., and a related formulation of Laplacian-regularised PCA was proposed by Jiang et al. (2013). The paper acknowledges these connections, but the methodological contribution appears limited.

Also, Bolya et al. discuss a possible cause of misleading PCA visualisations in their appendix: global tokens can dominate the leading components and in this way obscure spatial information. The authors should build on that and provide a more detailed discussion and deeper investigation of the causes in this work, given that they focus on the problem. Specifically, the paper quantifies spatial irregularity but provides limited analysis of what produces it or when it should be regarded as noise. High Dirichlet energy may reflect artifacts, but can also arise from meaningful fine-scale structure or spatially concentrated global information. Since the proposed method changes which feature directions are visualised, it would be helpful to characterise the suppressed directions and clarify what notion of visualisation faithfulness is intended. Improved alignment with segmentation and depth probes supports task-specific usefulness, but does not alone establish that the resulting visualisation is a more faithful summary of the representation overall.

Weakness 2 – Limited comparison to previous methods:

The authors suggest two properties a method should fulfil that put their method in contrast with two previous works, Bolya and Jiang. I see a missed opportunity to test the importance of these properties through direct comparisons. Bolya’s smoothing is linear but mixes tokens, while Jiang’s embedding is not generally guaranteed to be expressible as a tokenwise linear projection of the original features. Comparing against both methods (using the same spatial graph for Jiang) would help to establish the importance of the proposed constraints.

Questions for authors

Questions

  • What kinds of information are suppressed by positive Laplacian regularisation, and what evidence distinguishes unwanted noise from meaningful spatial variation?
  • How much do the proposed constraints matter empirically compared with Bolya’s visualisation and Jiang’s formulation?

Suggestions:

  • “To make an educated guess as to how such components might be identified, note that scene semantics are usually piecewise constant in space [….]” - rephrase this, as “piecewise constant” is too strong here, suggesting components stay exactly constant within regions and have abrupt changes at their boundaries. I think “spatially coherent” would be better.
  • L; 162-164: Clarify that you only refer to standard PCA at α=0\alpha=0 in Fig. 2 for this point. α\alpha has not been defined here.
  • More interpretation of their Laplacian PCA in eq. 5 would be useful.
  • In Table 1, ordinary PCA is the appropriate baseline for which three-dimensional visualization better preserves the probe’s performance. However, it would be useful to also provide full-feature performance to show how much performance either visualisation loses.
2 comments

Thank you for the interesting comments on this manuscript. As per your suggestions, we have improved baseline comparisons (W2), more clearly formulated the work's novelty (W1), and now provide a detailed of analysis of directions of variance that are removed from the visualizations through the Laplacian penalty (W1). We believe that the incorporation of your feedback has strengthened the paper and would appreciate it if you could take another look.

W1

The problem of noisy PCA visualizations has already been pointed out by Bolya et al., and a related formulation of Laplacian-regularized PCA was proposed by Jiang et al. (2013). The paper acknowledges these connections, but the methodological contribution appears limited.

We now state the novelty of the method more clearly through multiple adjustments: We incorporate a table (Table 1) to more concisely summarize differences with respect to other approaches. A new 'baselines' paragraph in 4.2 properly establishes that the formulation of Jiang et al. (2013) is not suitable for our application as it does not provide a parametric map from feature space to the visualization. The paragraph on negative regularization of DINOv3 in 4.2 now states more clearly that this discovery and analysis is not possible with the prior approach by Bolya et al., as their formulation can only produce smoother feature maps, which are suboptimal for DINOv3.

the paper quantifies spatial irregularity but provides limited analysis of what produces it or when it should be regarded as noise. High Dirichlet energy may reflect artifacts, but can also arise from meaningful fine-scale structure or spatially concentrated global information.

This is a very good point, and we believe that addressing it has lead to a substantial improvement of the manuscript. Previously, we probed the potential removal of global outlier tokens through Laplacian PCA in Fig. 5b. (original submission). We acknowledge your point that this was a relatively rough evaluation. We ran two additional experiments and present them in an entirely new section (4.3) dedicated to probing what Laplacian regularization removes from the visualizations.

First, we evaluated alignment with semantic segmentation on ADE20k classes stratified by their spatial granularity. We observe that Laplacian PCs align better with dimensions relevant for semantic segmentation across all levels of granularity, yielding strong evidence against the hypothesis that regularization destroys meaningful fine-grained semantic information (Fig. 7, new version).

We also extracted directions of variance that are unique to either the PCA or the Laplacian PCA representations to study what exactly is added/removed by regularizing (Fig. 8, new version). The most striking effect is that Laplacian PCA removes standard PCA's very strong ability to decode a token's spatial position in DINOv3, directly corresponding to the qualitative observation that a strong spatial bias is removed. Unique Laplacian PCA components segment fine-grained classes better than unique PC components, aligned with the findings from the previous paragraph. Finally, we do not find consistent evidence for unique standard PCA components to be strongly aligned with global image classification, going against the mentioned hypothesis that noisy feature maps arise from 'global tokens'. In summary, we do not find any specific 'use' for the directions of variance that Laplacian PCA removes, supporting the motivation to remove them.

W2

The authors suggest two properties a method should fulfil that put their method in contrast with two previous works, Bolya and Jiang. I see a missed opportunity to test the importance of these properties through direct comparisons.

We now incorporate a 'baselines' section in the paper in which we address this comment. We include the method by Bolya et al. in the quantitative comparisons, showing that Laplacian PCA outperforms the prior method in terms of alignment with task-relevant feature directions. This serves as quantiative evidence of the arguments provided in the original submission, and we do agree that it strengthens the paper. As mentioned before, we also emphasize more clearly that prior methods cannot produce the analysis on DINOv3 to set the work at hand apart better. Based on your comment, we acknowledge that we also did not clarify strongly enough that the method by Jiang et al. cannot be evaluated in terms of semantic alignment, as it does not provide a parametric map from feature space to the visualization. We now properly state this in the 'baselines' section, as well as in the new table for method comparison (Table 1). Importantly, we clarify that this is an inherent, strong disadvantage of that formulation, rather than a peculiarity of the specific evaluation protocol.

Finally, thank you also for the helpful additional suggestions, and the careful reading it required to provide them.

Thank you for the detailed response and additional experiments. The baseline comparison and analyses of the information retained and suppressed by Laplacian PCA have addressed my questions and strengthened the empirical contribution. While I still think that the methodological novelty is modest, I will raise my rating to weak acceptance.

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

This paper introduces the Laplacian PCA technique to iterate the use of PCA analysis as a tool for visualizing the representations of modern dense-feature extractors. This approach is motivated by the paper’s argument that visualizations of such representations should aim to be contiguous. The authors then move to quantify noisiness in modern visualizations, with the finding that modern models show a range of quantifiable noisiness results. Laplacian PCA is thus presented as a principled and simple alternative to reduce noise and create contiguous representations. Laplacian PCA is finally demonstrated to be superior to standard PCA in producing higher-quality linear components and feature maps that align better with downstream semantic tasks.

Critical strengths

In my opinion, the key strength of this paper is its careful work to ground and motivate the use of Laplace operators within the context of existing visualization techniques that use PCA. The reasoning and flow in sections 3.1, 3.2, and 3.3 all serve to successfully present the case for why an evolution of PCA is necessary and why the Laplacian PCA is a well-motivated option, though I have trouble following some mathematical steps. These steps make a strong argument for acceptance on the grounds of the paper’s value for current representation visualizations, and neural interpretations more broadly.

Critical weaknesses

My primary concern with Laplacian PCA is its framing as a tool for visualization that iterates and improves on standard PCA. The broad experimental results seem to validate its ability to improve on standard PCA’s feature maps, though this story is complicated by DINOv3. However, it is not clear to me that the significance of Laplacian PCA as a tool is justified by improvements on an existing tool. This paper would be much stronger if there was at least one case where Laplacian PCA uncovered something that standard PCA entirely missed, or which led to different conclusions, rather than strengthening a story that standard PCA is already telling.

Questions for authors

DINOv3 is noted as the outlier in your experimental results. The findings specifically note that DINOv3 might exist on an inverse U-shape as the other models do. In terms of the Laplacian PCA, what would this actually mean if α\alpha needs to be less than 0 to reach the earlier part of the U-shape?

Are there any cases where standard PCA systematically underestimating dense-feature quality, as the paper notes is a possibility, could have produced a conclusion that Laplacian PCA might overturn?

2 comments

Thank you for your thoughtful review and the interesting feedback; we also appreciate your support of the presented work.

You raised the question of what it actually means that DINOv3 is regularized with negative strength for optimal semantic alignment. We believe that this a strong point of the provided analysis, so we appreciate the feedback that it was not communicated clearly enough. We have made changes to the paragraph from l.373 onwards to explain that negative regularization attenuates the smoothest components of the visualization. There, we now also properly clarify why this findings is not possible with previously proposed methods. Further, we provide an additional analysis in 4.3 (Fig. 8), which demonstrates quantitatively why DINOv3 benefits from negative regularization: The directions of variance that are removed from PCA by the Laplacian penalty are highly aligned with those relevant for decoding a token's spatial position on the grid of image patches. Penalizing these directions frees the visualization up to encode more semantically meaningful structure. This finding aligns nicely with the qualitative plots shown in Fig.6, where this spatial bias of PCA is clearly visible as well.

You pointed out your concern that the presented approach primarily iterates and improves on standard PCA. We agree with the assessment, though would offer a different perspective on it: As we argue in the text, PCA already has multiple attractive properties. The proposed method is a simple, cheap way of extracting additional insight from this approach, which is already well-established in the literature. As you stated, the experiments demonstrate the method's ability to extract components that are more relevant for the usual downstream tasks. Since it inherits PCA's interpretability and computation/conceptual simplicity, we also see little reason not to incorporate Laplacian PCA into future model evaluations. To better drive this point in the paper, we now include an additional table (Tab. 1) that more explicitly establishes the method's benefits in comparison to PCA and other prior methods.

We also argue that Laplacian PCA can yield insights that PCA clearly fails to reveal: As an example, take the granularity of semantic structure in the second block of images in Fig. 3: PCA indicates that CLIP and the SAM3 encoder do not reflect the scene's more fine-grained semantic structure. However, Laplacian PCA demonstrates that both models clearly encode the semantic correspondences of animal body parts in the image. We also tried to improve the presentation of this argument in relation to the negative regularization of DINOv3, which clearly demonstrates different 'failure modes' of PCA (too smooth vs too noisy), which would not be discoverable without the Laplacian penalty. As described above, we refined these arguments in the text.

Thank you again for your comments on this work, which we believe have further strengthened the paper.

Thank you for your detailed response and clarifications. The DINOv3 regularization question has been resolved to my satisfaction. I also appreciate the further elaboration on 'failure modes' of PCA. However, my concerns about novelty remain, as it's not clear to me what the broader significance of the Laplacian PCA-specific insights are. As a result, I will maintain my rating at a weak acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.