OCCAM: Open-Set Causal Concept Explanation for Vision Models
A new framework named OCCAM has been developed by researchers to facilitate open-set causal concept explanation and ontology induction within black-box vision models. OCCAM identifies visual concepts without relying on predefined categories, utilizes text-guided segmentation for localization, and executes object-level interventions by eliminating concepts to assess variations in class confidence, thereby estimating the causal impact of each concept. In addition to local explanations, OCCAM compiles interventional evidence from a dataset to create a structured ontology that illustrates how classifiers organize visual concepts on a global scale. Analysis of this ontology reveals consistent relationships between concepts, highlights hidden causal connections, and identifies systematic biases in models. Testing on Broden and ImageNet-S with various classifiers indicates that OCCAM enhances the quality of explanations. This framework effectively tackles the challenge of interpreting deep image classifiers in scenarios where model internals are not accessible.
Key facts
- OCCAM stands for Open-set Causal Concept explAnation and Ontology induction for black-box vision Models.
- The framework discovers visual concepts in an open-set manner.
- It localizes concepts via text-guided segmentation.
- Object-level interventions remove concepts to measure changes in class confidence.
- OCCAM induces a structured concept ontology from interventional evidence.
- Experiments were conducted on Broden and ImageNet-S datasets.
- The framework improves explanation quality across multiple classifiers.
- The research is published on arXiv with ID 2605.18481.
Entities
Institutions
- arXiv