Into the Rabbit Hull: From Task-Relevant Concepts in DINO to Minkowski Geometry

AI summary

Overview

Research area: Interpretability of self-supervised vision transformers (specifically DINOv2), combining sparse dictionary learning, representation geometry, and cognitive theory.

Technical level: Intermediate. Familiarity with transformers and basic linear algebra helps, but the paper's core ideas are explained conceptually.

Scope: A three-part empirical and theoretical study of what DINOv2 encodes internally, using a 32,000-concept sparse autoencoder dictionary to probe task usage, geometric structure, and a proposed refinement of the Linear Representation Hypothesis toward convex/Minkowski geometry.

What This Paper Is About

DINOv2 is widely used as a visual backbone, yet it is unclear what its internal representations actually capture and how they are organized. The authors build a large interpretable dictionary of concepts from DINOv2 activations and then ask which concepts different tasks use, how the concepts are geometrically structured, and whether the standard assumption that features are sparse linear directions fully explains what they find.

Key Contributions

1. A 32,000-concept interpretability dictionary for DINOv2-B, trained with a stable sparse autoencoder whose atoms are constrained to the convex hull of real activations, achieving R² > 88% reconstruction.

2. Task-level functional specialization analysis showing that classification, segmentation, and depth estimation recruit distinct, low-dimensional subspaces of the concept dictionary, each with recognizable families of concepts.

3. A detailed geometric and statistical characterization of the dictionary, revealing departures from the idealized Grassmannian quasi-orthogonality assumed by the Linear Representation Hypothesis: heavier-tailed inner products, sharper singular value decay, partly dense positional concepts, and antipodal pairs.

4. The Minkowski Representation Hypothesis (MRH), a refined view in which tokens arise as convex mixtures of archetypal landmarks (e.g., "rabbit," "brown," "fluffy"), grounded in Gärdenfors' conceptual spaces and in the mechanics of multi-head attention.

Main Findings

• "Elsewhere" concepts support classification via learned negation. Across many ImageNet classes, top concepts include units that fire broadly except on the object itself, and that vanish entirely when the object is removed. These implement a conditional structure akin to "the object exists elsewhere, but this token is not it," which can mislead naive attribution heatmaps.

• Segmentation relies on a tight cluster of border concepts. The most important segmentation concepts consistently localize along object contours and silhouette transitions across diverse classes. Their mutual cosine similarity is higher than random concepts, and their eigenvalue spectrum decays faster, indicating a dedicated low-dimensional boundary subspace.

• Depth estimation uses three monocular cue families. Controlled image perturbations (shadow removal, edge-preserving smoothing, high-pass filtering) isolate three clusters: projective geometry cues (vanishing lines, converging structure), shadow-based cues (soft lighting gradients), and local frequency transitions (texture/detail changes). This taxonomy echoes classical visual neuroscience accounts of monocular depth.

• Register tokens carry global, non-spatial scene properties. Hundreds of concepts fire exclusively on register tokens and respond to holistic attributes like motion blur, illumination, caustics, lens distortion, and artistic style. By contrast, only a single concept fires exclusively on the CLS token, suggesting an asymmetric division of labor across token types.

• The dictionary is only partly sparse. Three outlier concepts activate densely across the entire dataset and encode persistent positional information (left, right, bottom). This indicates a hybrid regime where a small dense set coexists with a large sparse, image-contingent set.

• The dictionary departs from an ideal Grassmannian frame. Rather than becoming more mutually incoherent during training, the SAE develops higher local coherence and clustered structure. Its singular value spectrum decays more sharply than random or Grassmannian baselines, indicating anisotropy with some heavily used directions and others barely used.

• Antipodal concept pairs form signed semantic axes. Many concepts have near-opposite counterparts with cos θ ≈ −1, meaning that negative cosine similarity can reflect two poles of a single signed feature rather than two unrelated features. This has practical implications for clustering and retrieval pipelines that rely on cosine similarity.

• Per-image token clouds are locally connected and low-dimensional. Even after removing positional components and despite positional information compressing into roughly a 2D subspace in later layers, tokens within an image occupy a smooth, low-dimensional, locally connected manifold.

• Concept co-activation shows no clean modular structure. The Gram matrix of concept co-activation has a smoothly decaying spectrum without sharp gaps, lying between random and shuffled baselines, indicating a high-dimensional distributed regime rather than clean clusters.

Methodology in Plain English

The authors first train a sparse autoencoder on DINOv2-B token activations. An autoencoder compresses each token into a small set of active "concept" units and then reconstructs the original activation from them. To keep the learned concepts realistic and stable across runs, each dictionary atom is forced to be a weighted average of actual activation vectors drawn from a large pool of centroids (from k-means over 1.4M ImageNet images), rather than a free-floating direction. They use 32,000 concepts, allow only about 8 active per token, and use BatchTopK projection for sparsity.

To study downstream usage, they take linear probes trained for classification, segmentation, and depth, re-express them in the concept basis, and measure how much each concept contributes. They verify concepts causally by masking them and checking the effect on outputs.

To study geometry, they compute pairwise inner products, singular value spectra, Hoyer sparsity scores, and Gram matrices of concept co-occurrence, and compare each against random and Grassmannian baselines. They also examine per-concept activation footprints to distinguish position-specific, CLS-only, and register-only units.

For depth, they apply targeted image manipulations designed to remove specific monocular cues and check which concept clusters respond, then visualize prototype patches.

Finally, they synthesize the observed structure into the Minkowski Representation Hypothesis, connecting it to Gärdenfors' conceptual spaces and to the fact that a multi-head attention layer sums convex combinations of value vectors.

Why This Matters

Impact on research. The paper challenges the assumption that vision transformer representations are cleanly linear-sparse. The Minkowski Representation Hypothesis reframes interpretability around convex regions and archetypal landmarks rather than unbounded directions, which changes how researchers should design probing methods, steering interventions, and similarity-based retrieval. The findings on "Elsewhere" concepts also warn that standard attribution heatmaps can systematically mislead.

Real-world applications:

• Model auditing and safety. Understanding that concepts can encode conditional negations and global scene factors helps auditors detect unanticipated reasoning patterns before models are deployed in high-stakes settings.

• Better retrieval and clustering pipelines. Since antipodal concept pairs can look dissimilar under cosine similarity, systems built on DINOv2 embeddings may be missing semantically related items or treating signed axes incorrectly.

• Targeted model editing and steering. Knowing which concept subspaces control segmentation versus depth versus scene style allows more surgical interventions, useful in robotics, autonomous driving, and medical imaging where specific failure modes matter.

• Generative and multimodal conditioning. The register-token concepts that capture illumination, blur, and style offer a natural handle for controlling image generation or for aligning visual features with language.

Industry relevance. Any team deploying vision foundation models for perception, content moderation, or multimodal products benefits from a concrete map of what the model internally represents and how reliably those representations can be manipulated.

Future Directions

• Formalize MRH mathematically. The paper proposes Minkowski sums of convex polytopes as the organizing geometry; a rigorous characterization of when this holds and how to estimate the underlying archetypes remains open.

• Extend the dictionary to larger and different backbones. Whether the observed specialization, antipodal pairs, and dense positional outliers generalize to DINOv3, CLIP, or convolutional architectures is untested.

• Build interpretability tools that respect non-local and antipodal structure. Existing heatmap and cosine-similarity methods are shown to be misleading; new tools that account for conditional negations and signed axes are needed.

• Connect archetype geometry to cognitive theory more directly. Gärdenfors' conceptual spaces predict specific convex-region properties; testing those predictions empirically in models could validate both the theory and the proposed representation view.

• Use MRH for steering. If tokens are convex mixtures of archetypes, interventions might be designed as adjustments to mixture weights rather than direction ablations, which could be more robust and semantically cleaner.

Target Audience

Researchers and graduate students in mechanistic interpretability, vision transformer analysis, and representation learning. The paper is also valuable for engineers applying DINOv2 to segmentation, depth, or retrieval who want to understand known failure modes and structure. Cognitive scientists and computational neuroscientists interested in conceptual spaces will find the bridge to Gärdenfors' theory and to monocular depth cue literature particularly relevant.

Authors’ abstract

DINOv2 is routinely deployed to recognize objects, scenes, and actions; yet the nature of what it perceives remains unknown. As a working baseline, we adopt the Linear Representation Hypothesis (LRH) and operationalize it using SAEs, producing a 32,000-unit dictionary that serves as the interpretability backbone of our study, which unfolds in three parts. In the first part, we analyze how different downstream tasks recruit concepts from our learned dictionary, revealing functional specialization: classification exploits "Elsewhere" concepts that fire everywhere except on target objects, implementing learned negations; segmentation relies on boundary detectors forming coherent subspaces; depth estimation draws on three distinct monocular depth cues matching visual neuroscience principles. Following these functional results, we analyze the geometry and statistics of the concepts learned by the SAE. We found that representations are partly dense rather than strictly sparse. The dictionary evolves toward greater coherence and departs from maximally orthogonal ideals (Grassmannian frames). Within an image, tokens occupy a low dimensional, locally connected set persisting after removing position. These signs suggest representations are organized beyond linear sparsity alone. Synthesizing these observations, we propose a refined view: tokens are formed by combining convex mixtures of archetypes (e.g., a rabbit among animals, brown among colors, fluffy among textures). This structure is grounded in Gardenfors' conceptual spaces and in the model's mechanism as multi-head attention produces sums of convex mixtures, defining regions bounded by archetypes. We introduce the Minkowski Representation Hypothesis (MRH) and examine its empirical signatures and implications for interpreting vision-transformer representations.