A Better Concept Map Can Reconstruct Less

TL;DR for operators

When inspecting a vision model, the sparse autoencoder that best reconstructs its internal activations may not produce the clearest explanation of what image regions those activations represent.

Spatial-SAE addresses a specific modeling limitation: conventional sparse autoencoders do not explicitly coordinate features across neighboring image patches. Its two-stream representation improves concept recovery and spatial alignment in the studied settings, but generally sacrifices reconstruction fidelity.

For engineers selecting interpretability tools, reconstruction accuracy should be evaluated alongside semantic concept alignment and region coherence. The method is promising when spatial relationships are meaningful and annotated examples are available. It is not yet evidence that more coherent features reveal causal model mechanisms or improve downstream decisions.

The better concept map has the worse reconstruction

Consider a team examining which visual concepts a deployed vision model has learned. Two feature-extraction models are available. One reproduces the model’s internal activations with relatively little error. The other reconstructs those activations less accurately, but its feature maps correspond more closely to annotated image regions.

Which representation should the team trust when investigating the model?

In The Independence Prior of SAEs Fragments Visual Concepts, Tommaso Mencattini and colleagues investigate precisely this disagreement.¹

Their experiments on frozen DINOv2 activations from ADE20K report the following comparison for TopK sparse autoencoders at an expected 16 active features per patch:

Source: Section 5.2, Figure 6; Appendix D, Table 3. The reported Normalized F1 and reconstruction results are means over five seeds. Region coherence uses the paper’s own scale.

Normalized concept F1 compares the best individual SAE feature for each eligible semantic class against a linear probe trained to identify that class. It measures how much linearly decodable semantic information is captured by a single feature, rather than ordinary segmentation accuracy. Region coherence measures whether patches within the same annotated region have more similar codes than equally distant patches belonging to different regions.

These measures answer questions that reconstruction R² cannot. An autoencoder can preserve substantial activation information without organizing that information into features that align with recognizable regions.

The observed 18-point reconstruction loss is nevertheless significant. Spatial-SAE represents a different allocation of representational fidelity, not an unconditional improvement.

Conventional SAEs overlook relationships between neighboring patches

A vision transformer represents an image through a grid of patch activations. Neighboring patches often contain related information because they belong to the same object, surface, or semantic region.

A conventional sparse autoencoder learns a dictionary of features and reconstructs each activation using a sparse combination of those features. Its training objective typically adds reconstruction and sparsity costs across patches.

This creates a subtle limitation.

The objective does not explicitly reward neighboring patches for using mutually consistent explanations. Its sparsity penalty corresponds to a factorized prior over patch-level latent codes. The authors prove that the conventional objective is invariant to simultaneously permuting activation and code rows.

That result concerns the model’s inductive bias. It does not imply that the codes learned from real images are necessarily statistically independent.

The distinction matters because the same visual concept may be represented by different dictionary features across adjacent patches. The autoencoder may still reconstruct those activations well, even when a concept appears fragmented in the resulting feature maps.

The researchers also examine frozen DINOv2 representations directly. Genuine spatial neighborhoods contain predictive information beyond neighborhoods shuffled within the same image. Spatial organization therefore persists in the representations being analyzed, even though conventional SAE objectives do not explicitly model it.

This motivates the paper’s central hypothesis: sparse visual representations should account for relationships between neighboring patches, not only the sparse composition of individual activations.

Spatial-SAE separates persistent concepts from local innovations

The authors introduce the Markov-Field Linear Representation Hypothesis (MF-LRH), extending conventional sparse representation assumptions with spatial interactions.

Its central distinction is between two kinds of information:

• A smooth field captures features that persist across neighboring patches, encouraging spatial agreement.
• An innovation field captures localized differences, retaining variation that should not be forced into a smooth representation.

Image patches form nodes in a graph, with edges connecting nearby patches. Spatial frequencies describe how quickly code values vary across that graph: low frequencies indicate gradual spatial variation, while higher frequencies indicate more localized contrasts.

Spatial-SAE uses complementary graph-spectral penalties to encourage low-frequency structure in the smooth field and relatively higher-frequency structure in the innovation field. A shifted graph Laplacian keeps the innovation penalty mathematically well behaved.

The resulting objective follows from maximum a posteriori inference under the proposed spatial prior and Gaussian observation noise. Its implementation learns separate sparse representations for shared neighboring content and local residual information.

This is a change to feature inference and training, not simply post-processing that makes existing feature maps look smoother.

The distinction is experimentally consequential. At regularization strength 0.001, a two-stream Spatial-SAE improves ADE20K Normalized F1 by 6.12 ± 1.86 percentage points over a single-stream model with spatial regularization, across 14 matched TopK comparisons (Appendix D, Table 4).

The two-stream model wins all 14 comparisons on concept F1 at that setting. It does not simultaneously improve every metric: its mean region-coherence difference relative to the single-stream control is negative.

The result suggests that separating persistent features from localized variation contributes something beyond generic spatial smoothing.

The recovery experiments identify the benefit and its cost

The paper first tests concept identification on synthetic activations generated from known dictionaries, smooth spatial fields, localized innovations, and noise.

These experiments provide direct recovery evidence because the underlying concepts are known.

Across 300 matched comparisons per sparsifier, Spatial-SAE achieves approximately a 96% average win rate across dictionary-matching correlation, polysemanticity, and support F1, averaged over four sparsifier families (Section 5.1, Table 1).

For TopK and BatchTopK, it wins every matched comparison on those three identification measures. Yet both families record a 0% win rate on reconstruction R². JumpReLU provides an exception, improving both concept identification and reconstruction throughout its matched synthetic comparisons.

The 96% figure is a win rate, not an absolute identification accuracy. It measures how often Spatial-SAE outperforms its matched baseline under the tested synthetic configurations.

Natural-image experiments provide a complementary, less direct test. On ADE20K, all four sparsifier families improve Normalized F1 at the selected sparsity budget, while reconstruction fidelity declines.

Spectral analysis helps locate that cost. The baseline’s reconstruction advantage over Spatial-SAE grows from 0.11 at low graph frequencies to 0.37 at high frequencies. The spatial method therefore loses disproportionately more reconstruction fidelity in components that vary rapidly across patches.

This supports an interpretation of the tradeoff: encouraging persistent spatial concepts changes which activation details the learned representation preserves.

It does not establish that the discarded high-frequency information is unimportant for the original model’s predictions.

Stronger spatial coherence can conceal worse concepts

The regularization ablations reveal a separate failure mode.

In the static TopK configuration on ADE20K, increasing spatial regularization from 0 to 0.001 produces a median 6.53-point gain in Normalized F1, alongside a 5.52-point reduction in reconstruction R² (Appendix D, Table 5).

Moderate regularization can therefore improve semantic alignment while imposing a measurable reconstruction cost.

But increasing the regularization strength from 0.01 to 0.1 produces a different outcome:

• Median Normalized F1 declines by 43.14 points.
• Median reconstruction R² declines by 27.94 points.
• Median region coherence increases by 16.04 points.

The feature codes become more spatially coherent by the reported measure even as their semantic alignment deteriorates sharply.

This is an important boundary for model selection. Spatial coherence is a diagnostic, not independent proof of interpretability. A representation may become increasingly uniform without preserving the distinctions that matter for understanding an image.

Transfer strengthens the result without establishing generality

The authors also evaluate COCO-Stuff, using architectural and hyperparameter choices selected on ADE20K without dataset-specific retuning. The transferred configurations are trained from scratch on the new dataset.

Across four sparsifiers, Normalized F1 improves by 1.4–7.1 percentage points and region coherence by 6.5–9.0 points.

Blinded automated assessments also report improvements in feature-map quality and monosemanticity. These ratings come from a single language-model evaluator, Claude Haiku 4.5. Their reported class-bootstrap intervals do not measure agreement among independent raters.

The COCO-Stuff results are evidence of cross-dataset robustness. They are not evidence of cross-backbone generalization: both natural-image evaluations use frozen DINOv2 ViT-B/14 representations.

Nor does segmentation alignment establish that the discovered features are the unique causal concepts governing the vision model’s behavior.

What model-inspection teams should change

For an engineering or model-governance team choosing an SAE to investigate visual representations, Cognaptus draws a conditional implication from these findings.

When the inspection objective is to identify which regions correspond to learned features, model selection should jointly examine semantic alignment, distance-matched region coherence, and reconstruction fidelity. Relying on reconstruction alone may select a dictionary that preserves activation details while fragmenting region-level concepts.

Spatial-SAE is a candidate when the representation has meaningful neighborhood relationships and the team can validate discovered features against annotated examples.

That evaluation should also test sensitivity to regularization strength, inspect the consequences of lost high-frequency reconstruction, and account for additional training computation. The context-attention variant, for example, introduces greater training cost than the conventional baselines.

For inspection pipelines, clearer feature maps may make diagnostic review more informative. Whether this improves error diagnosis, corrective interventions, or model-governance decisions remains an empirical question requiring downstream validation.

The study’s strongest direct recovery evidence comes from controlled synthetic generators. Its real-image findings depend on segmentation-derived and automated interpretability proxies, using one frozen vision backbone. No downstream intervention study establishes causal utility.

These boundaries make wholesale replacement of conventional SAEs premature, particularly where fine-grained reconstruction is essential.

Interpretability needs a metric matched to the question

The paper identifies a specific reason why accurate reconstruction can coexist with fragmented explanations: an autoencoder can faithfully represent individual patch activations without explicitly accounting for their spatial relationships.

Spatial-SAE demonstrates that changing this assumption can improve concept recovery and region-level alignment. It also shows that those gains have reconstruction costs and that excessive regularization can make apparent coherence misleading.

For teams inspecting vision systems, the choice of interpretability model should begin with the question the representation must answer. Reconstruction, semantic alignment, and spatial coherence measure related but different properties. None alone establishes that the learned features explain the model’s causal behavior.

Cognaptus: Automate the Present, Incubate the Future.