OCCAM: Open-set Causal Concept explAnation and Ontology induction for black-box vision Models
The paper introduces OCCAM, a framework designed for open-set causal concept explanation and ontology induction in black-box vision models. OCCAM aims to enhance the interpretability of deep image classifiers by discovering and localizing visual concepts and measuring their causal contributions. Experimental results demonstrate that OCCAM improves explanation quality and provides a more comprehensive understanding of visual concepts compared to traditional methods.
- ▪OCCAM discovers visual concepts in an open-set manner and localizes them through text-guided segmentation.
- ▪The framework performs object-level interventions to assess changes in class confidence and estimate causal contributions.
- ▪OCCAM aggregates evidence across datasets to create a structured concept ontology, revealing dependencies and biases in classifiers.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.18481 (cs) [Submitted on 18 May 2026] Title:OCCAM: Open-set Causal Concept explAnation and Ontology induction for black-box vision Models Authors:Chiara Maria Russo, Simone Carnemolla, Simone Palazzo, Daniela Giordano, Concetto Spampinato, Matteo Pennisi View a PDF of the paper titled OCCAM: Open-set Causal Concept explAnation and Ontology induction for black-box vision Models, by Chiara Maria Russo and 5 other authors View PDF HTML (experimental) Abstract:Interpreting the decisions of deep image classifiers remains challenging, particularly in black-box settings where model internals are inaccessible. We introduce OCCAM, a framework for open-set causal concept explanation and ontology induction in vision models.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
