OCCAM: Open-set Causal Concept explAnation and Ontology induction for black-box vision Models
Researchers have developed OCCAM, a new framework designed to explain the decisions of black-box image classifiers. OCCAM identifies visual concepts, localizes them using text guidance, and measures their causal impact by removing them to observe changes in model confidence. This approach not only provides per-image explanations but also induces a structured concept ontology to reveal global model biases and dependencies between concepts. AI
IMPACT Provides a new method for understanding and debugging vision models, potentially improving trust and identifying biases.