acceptodds
Under review as a conference paper at ICLR 2027

Distributional Causal Probing of Vision Models via Self-Controlled Interventions

Abstract

Occlusion studies often summarize model responses by average performance changes, obscuring heterogeneous effects across images and their relationship to representation changes. We introduce Self-Controlled Intervention, a paired design for probing frozen vision models with deterministic interventions at prespecified doses and directions. Building on this design, we establish a statistical framework that characterizes heterogeneous decision effects, representation changes, and their dependence. We conduct extensive numerical experiments in image classification and object detection to evaluate the framework. The results reveal substantial effect heterogeneity and task-dependent directional vulnerability. These findings challenge robustness assessments based on average performance and caution against inferring decision-level vulnerability from representation drift alone.

Primary area:
interpretability and explainable AI
Keywords:
self-controlled intervention, causal probing, model interpretability, representation geometry, quantile intervention effects, vision model robustness, effect heterogeneity

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Ug1YOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes Self-Controlled Intervention (SCI) to study primarily how image occlusion affects vision models. It measures changes in predictions and internal representations, focusing on the distribution of effects. Experiments cover several classifiers and object detectors.

Critical strengths

The distributional analysis provides useful insights that average performance metrics can miss. E.g., some images become more vulnerable than others, and changes in internal representations do not always follow changes in predictions.

Critical weaknesses
  1. Limited Novelty. I feel the proposed framework has components including paired image comparisons, quantile analysis, MMD and HSIC, are well established techniques. But it remains unclear to me what new insights this framework provides scientifically.

  2. Limited causal interpretation. I appreciate the authors' aim to study causal effects through this technique, but I feel the main focus is on model robustness analysis rather than establishing any underlying causal mechanisms. E.g, the foreground/background comparison attempts to separate information loss from artificial effects caused by masking. However, this requires assumptions that are not experimentally established. The authors acknowledge these limitations, which weaken the causal interpretation.

Questions for authors
  1. What important findings could SCI reveal that simpler paired occlusion analysis using per-image effect distributions could not?
  2. How can the authors support the assumptions needed to separate information loss from masking artifacts?
  3. Where is the code you stated in the reproducibility secion?
Reviewer_fLZHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents a probing strategy of vision models based on a masking analysis. Specifically, models are evaluated with metrics that capture the failure distribution more effectively instead of just considering the average. Analysis on image classification and object detection indicate performances that vary across different tasks and masking strategies.

Critical strengths

The paper’s motivation is relevant: More fine-granular masking and metrics allow a more systematic evaluation of vision model robustness.

Critical weaknesses

While several models are evaluated, the analysis and discussion are missing focus. Many references are given (even to many supplementary material results) but the core learnings or not condensed. This blurs the contribution and learning of the paper. The authors are encouraged to give a clear overview of why the work is relevant and what lasting learning the community can take from it. Particularly, it would be interesting to know whether observed effects also hold for more recent models. The evaluations are performed on a small dataset, eg, 10 classes and <3000 images for the classification setting. I am concerned that the measurements can provide sufficient evidence for the drawn conclusions.

Questions for authors

Many results are referenced in the supplementary. Could you provide a focused summary of the most important ones (and their findings). Could the authors provide a concise positioning of their work, comparing to the large body of related works.

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

This paper proposes self-controlled intervention (SCI) to probe the causal robustness of vision models. SCI differs from existing occlusion robustness evaluations in that 1) it measures robustness at image-level by gradually occluding pixels based on the object mask (from SAM 3) along different directions, and 2) it probes the internal representation drift of models in addition to their final predictions. Empirical results are presented by applying SCI to probe frozen image classification and object detection models.

Critical strengths
  • The paper has clear motivation (image-level robustness) and establishes empirical procedures to measure this. The overall intervention design is reasonable and well executed.
  • Empirical analysis is thorough within the proposed experimental framework, including uncertainty intervals via resampling. The observed variation between per-image responses is a good case for reporting finer-grained measurements in occlusion studies.
Critical weaknesses
  • While the empirical findings are mostly valid, the scientific or practical implication is somewhat obscure. The paper would be stronger if further analysis were performed to reveal certain insights, e.g., when SCI flips model ranking vs. conventional aggregate metrics, or how to build better data augmentation for training. The author could choose a setting to highlight the vulnerability of existing robustness evaluations and demonstrate how SCI addresses it; or connect SCI to some behavioral findings of the evaluated models, etc. The insights from these would add to the practical significance of the work.
  • The motivation to the use of placebo, Assumption 1 & 2 (L215-218), are not adequately validated. In fact, ImageNet has substantial label-background correlation (https://arxiv.org/abs/2006.09994) that could invalidate assumption 1. There could be simple experiments to check whether the assumptions hold on the test data, e.g. by measuring classification accuracy on background only or swapping background with another random sample.
  • The paper too frequently refers to figures and tables in the appendix, making it difficult to follow at times. In fact, no figures or tables in the 9 pages of main text address the representation vs. decision drifts, a central claim of the paper.
Questions for authors
  • Can the SAM-based occlusion leak object class information through the shape of the mask (e.g., https://arxiv.org/abs/2202.00449)?
  • The empirical study uses only 10 mini-ImageNet classes for classification & COCO for detection. Have the authors experimented more categories and/or vision tasks?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.