acceptodds
Withdrawn from ICLR 2027

Vocabulary Aligned Explanation for Computer Vision Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Optimizing vision models for accuracy does not force the model to learn and utilize human-understandable and domain-specific concepts and rules. We propose explaining image classifications through domain-specific concepts, visual evidence, and expert knowledge expressed as logical rules. We introduce FLUENT (Framework for Localized, Uncertainty-Aware Explanations in Normative Terminology), a post-hoc framework that ranks expert-authored rules combining jointly required, alternative, and absent concepts. To account for prediction errors, FLUENT combines predicted labels and concept states with empirically estimated likelihoods and concept priors. It scores each rule by calculating support over satisfying assignments to the concepts mentioned in that rule, then normalizes these scores to express relative confidence among candidate explanations. The selected rule is illustrated through occlusion-based highlights for positive concept mentions in the query image and localized examples in retrieved reference images for negative mentions. These illustrations are intended to show end users unfamiliar with the domain what the concepts named in an explanation look like. FLUENT combines domain knowledge and model accuracy, and delivers visual concept highlights and explicit rule confidence to explain an image in the vocabulary of an application domain.

Primary area:
interpretability and explainable AI
Keywords:
Explainable AI, Computer Vision, Concept-based Explanations, Bayesian Reasoning

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_1Wz8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FLUENT is a post-hoc explanation framework that attaches separately trained label and concept predictors to a pretrained vision backbone and selects, among expert-authored propositional rules, the rule best supported by the predictions. Rule scores combine empirically estimated label/concept error likelihoods with concept priors, and are normalized across candidate rules to yield a relative confidence.

Critical strengths
  • The problem is clear and well motivated. Illustrating absent concepts (fig.1) is a simple, intuitive idea.
  • Explicitly modeling predictor error rates lets concept evidence correct weak classifiers
  • The limitations are reported with unusual candor, and no claims are overstated.
Critical weaknesses
  • Technical novelty is limited. The core is a naive bayes style rule scorer combined with standard occlusion and feature-space retrieval.
  • The evaluation is weak. The only real dataset has 3 classes, 4 concepts, and one rule per class, with class-level template rules as targets, so rule acc is essentially a restatement of label accuracy.
  • The central motivation is end-user understanding, yet no user study assesses whether the explanations are actually more understandable or useful than saliency or text baselines.
Questions for authors
  • On bees, rule-implied labels are consistently worse than the classifier. What is the value of the rule explanation in this regime?
  • How sensitive are the occlusion highlights to patch size N, and can localization quality be measured?
  • How does the method handle inputs that satisfy no rule or multiple rules, beyond the all-zero-weight fallback?
Reviewer_fkMyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FLUENT, which tries to explain a decision in terms of which rule is used. A rule is a set of indicators for presence (or absence) of a set of concepts. For example, an image is Category A because Rule R1 is satisfied and Rule R1 is satisfied its logic (Concept 1 present and Concept 2 absent) is satisfied. The rules are per-class, with both rules and concept sets assumed to be provided by domain experts. For an input image, the paper proposes to have a per-concept detector, which gives a binary value of whether the concept is present or not. The presence and absence of concept is then used to see which rule is satisfied. Therefore the rule and the concepts within it can be shown to the user as an explanation.

Critical strengths

s1) The introduction and related work of the paper was fairly well-written and easy to follow. The Fig 1 was also helpful in understanding what is meant by rules and concepts

s2) Working on discrete rules that are defined by experts is a problem that is quite relevant.

Critical weaknesses

W1) My main concern is that I am not sure what exactly to make of the results section and the contribution.

a) In terms of method, the concepts and rules are pre-defined by the user. Therefore this is not a concept discovery or rule extraction method. While this in essence is fine with me (as in the real world often the rules and concepts can be provided from the expert), it makes me question what exactly is the contribution of this “post-hoc framework”. Training a per-concept detector was already done in the very original CBM paper. Running logical inference on the predicate rules also sounds trivial. The only part that was somewhat new to me was using a Bayesian model for rule selection, given the concept confidences. Could the authors clarify if they are claiming this as the core novelty of the method?

b) Assuming that the Bayesian model for rule selection is the novelty, I still did not see any results that would sell the model to the reader. The only results section I found was Table 1, which shows the accuracy on concept, label, and rule prediction of the model across backbones and datasets. As a reader, it is not clear what one should make of this. There are no baselines in this table. This table doesn’t tell me as a reader, if I the bayesian model is working better than simpler approaches, say learning a simple confidence threshold per concept. It doesn’t also tell the reader how doing predictions by (concept detection -> rule selection) compares to the opaque vanilla classifier. Are we losing on performance by doing prediction based on concepts? Also from concepts to prediction, one can just train a linear classifier instead of using pre-defined rules (as done in CBMs). Are we again doing better compared to those by using pre-defined rules? There is unfortunately no comparison in the table hinting at these questions.

c) Aside from quantitative performance, the goal of this paper and its surrounding literature is to provide a more transparent decision making. While I appreciate Fig 1’s example, I currently don’t see any interpretability analysis that the concept detections and rule selections are in any way meaningful to the user (or the expert). That would also need comparison to existing rule extraction methods and CBMs.

Questions for authors

A few mentioned above.

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

FLUENT selects an expert-authored logical rule from predicted labels and binary concepts, using empirical predictor-error likelihoods and concept priors. It illustrates positive literals with query-image patches and negative literals with retrieved examples. Experiments cover twelve backbones on synthetic images and bee photographs.

Critical strengths

The method combines domain terms, simple rules, and visual examples, including reference images for concepts the rule requires to be absent.

Critical weaknesses
  1. The bee evaluation is not independent. The separate disjoint-calibration experiment does not resolve the related-image overlap. These results do not establish performance on independent observations.

  2. The benchmark does not check whether the explanation matches the photograph. Bee concepts are assigned by class, with one rule per class and no independent check of visible cues. Simply returning the classifier's class rule would score 99.7% for CNXv2-L, versus FLUENT's 85.8% RuleAcc. This follows from the metric definition and exposes the weakness of the evaluation.

  3. Equivalent rules can receive different scores. In Eq.7, replacing a with the equivalent rule a AND (a OR b) adds a likelihood factor for b, which can change the selected explanation. Appendix A.2 acknowledges penalties from different evidence sets, but the paper does not justify this dependence on wording or compare against its full-evidence alternative.

Questions for authors
  1. What results remain when training, likelihood fitting, and explanation testing use deduplicated images from separate observations?
  2. Can rule recovery be evaluated against independently checked concepts visible in each photograph, rather than class-level templates?

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

Related papers

Discussion (0)

Sign in to comment.