acceptodds
Withdrawn from ICLR 2027

Attributing Class Recognition to Invariants with Recursive Peeling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A sparse autoencoder splits what a vision model computes into concepts, and most work studies these concepts one at a time. We ask instead what they reveal together. We call them class unique when they fire consistently on one class and shared when they fire consistently on several, representing invariants of different sizes. We propose Multi-level Clustering with Peeling (MCP) to uncover how these invariants organize classes and contribute to recognition. MCP clusters classes at resolutions determined by how widely their concepts are shared. Concepts shared only within one cluster are peeled as cluster unique concepts, exposing further relationships at subsequent levels. Applied to the final-layer class token of the self-supervised DINOv2 on ImageNet, MCP reveals a forest of visual categories that departs from the WordNet hierarchy. The clusters are tighter and more semantically coherent than size-matched random clusters. Ablations with a frozen linear classifier show that removing the concepts peeled from a cluster lowers accuracy within that cluster, with near-zero median effects outside it. Different classes rely on class unique invariants, cluster unique invariants, or both. Removing both kinds produces super-additive accuracy losses, while either kind alone can preserve recognition. These results suggest that the model recognizes classes by parallel abduction, weighing evidence for single classes and for clusters of related classes at once. Class unique invariants distinguish individual classes, while cluster unique invariants support several competing classes. Unlike human vision, which reaches its decision by recursive abduction, the model weighs every candidate in a single pass.

Primary area:
interpretability and explainable AI
Keywords:
Sparse autoencoders, Interpretability, Vision transformers, Concept-based explanations, Causal ablation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_owyFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates how DINOv2 uses visual concepts that are unique to individual classes or shared across multiple classes. Its main methodological contribution is Multi-level Clustering with Peeling (MCP), which repeatedly clusters classes and removes shared concepts to build a multilevel structure of class relationships. The authors evaluate the resulting clusters using semantic similarity measures, clustering baselines, and concept-removal experiments. They find that removing class-unique and shared concepts together often causes a larger accuracy drop than the sum of removing either group alone. Based on these results, the authors make their main scientific claim: DINOv2 uses class-specific and shared evidence simultaneously through a process they call "parallel abduction."

Critical strengths

The authors systematically measure the effects of removing class-unique and shared concepts from DINOv2 representations. The experiments reveal substantial differences across classes in their reliance on these two concept groups. The joint removal experiments also show that both groups can contribute important information to classification. These findings provide useful empirical insights into how visual concepts support recognition, even though they do not establish the proposed parallel abduction mechanism.

Critical weaknesses
  1. The central claim of parallel abduction is not supported by the experiments.

The primary evidence for parallel abduction is the super-additive accuracy loss that occurs when both the class unique and cluster unique concepts are removed. There's actually a far simpler explanation for this phenomenon in terms of the classification margin of a standard linear classifier. The two concept groups additively contribute to the logits. You can remove one of the groups and still have a positive classification margin. Remove both and the margin will go negative and you get a misclassification. For instance, if the original margin is 3 and each concept group contributes 2 to that margin then removing either one of the groups will leave you with a margin of 1 (still positive). Remove both and you end up with a margin of -1 (negative). So you can quite easily get that super-additive degradation in accuracy without any abductive mechanism at all. The experiments do show that the two concept groups are providing complementary information for classification but they don't demonstrate that the model is doing parallel abduction any more than they show that it's doing ordinary discriminative classification. That really undercuts one of the paper's key claims.

  1. The novelty and necessity of MCP are insufficiently established.

MCP is a combination of existing sparse subspace clustering with a set of heuristic rules for picking the clustering resolutions and for iteratively removing concepts. There's no principled rationale given for the choice of these rules and no convincing evidence that the resulting structure is better than simpler alternatives such as graph based clustering or simply filtering concepts by their firing set size. More fundamentally, the key class level ablations in Section 4.2 don't require MCP at all. Under the peeling rule a class stays active until all of its linked concepts have been peeled. As MCP will eventually peel every shared concept, the full set of cluster unique concepts linked to a class is precisely the original set of shared concepts in the biadjacency matrix BB. The main class level ablation results can therefore be derived directly from BB without any need for clustering or peeling. This really does cast serious doubt on what MCP is actually contributing to the paper's main results and whether the added complexity of MCP is worth it.

Questions for authors
  1. Parallel abduction

The super-additive accuracy loss seems to have a much simpler explanation. In a linear classifier you could remove either of the concept groups and still have a positive classification margin. You could remove both and end up with a negative margin. That doesn't require any abductive reasoning at all. What evidence do the authors have to support their interpretation of this as parallel abduction rather than this much simpler explanation?

  1. MCP's role in Section 4.2

Since MCP ultimately strips away all the shared concepts, isn't the full set of cluster-unique concepts associated with a class exactly the same as its original set of shared concepts in BB? If that's the case then the class level ablations in Section 4.2 could be carried out without MCP at all. What exactly is MCP doing in those experiments?

  1. Iterative peeling

Why is iterative peeling necessary? A much simpler approach would be to just filter the concepts on the basis of their firing set sizes and then cluster the classes at various resolutions. Does MCP actually provide any advantage over that approach?

Reviewer_7Fd9OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors find a mapping between the class variable and the sparse features in an auto-encoder. The mapping is not always one-to-one and they use different names for different mappings. They find a hierarchy for the case where a class is represented by more than one feature. They show in this situation, the network is more robust because the class information is shared between different features. Since the features have some kind of hierarchy, they conclude that the used vision transformer is considering this hierarchy all at once in contrast to human vision which is recursive.

Critical strengths

1- Smooth and easy to follow language 2- Comprehensive experiments and ablation studies

Critical weaknesses

1- The word concept is sometimes misused in research papers including this one. A better wording would be feature or something similar. 2- The method is proposed for sparse auto-encoders but only tested on vision transformers. It needs testing on other models as well. 3- The fact that some features hold the information of more than one class is trivial. The contribution of this paper is mainly on how those features are structured conceptually. 4- The word cluster is used in the literature for grouping the data-points not the features. Another wording could reduce the confusion when reading the paper.

Questions for authors

1- The anonymous GitHub link doesn't work. Could you check it? 2- Did you check interpretability/explainability literature. There exist many good resources. For example https://already-clean.goodfire.com/ is a recent blog about these matters. 3- Did you try to find a way so the network adaptively finds the hyper-parameters such as thresholds (equation 1) and scaling factors (appendix 1 - other backbones)?

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

Related papers

Discussion (0)

Sign in to comment.