OCCAM: Open-set Causal Concept explAnation and Ontology induction for black-box vision Models

In three linesOCCAM is a framework for explaining black-box image classifier decisions through causal visual concepts. It discovers concepts in open-set manner, localizes them via text-guided segmentation, and measures causal contribution through object-level interventions. OCCAM aggregates interventional evidence to induce a structured ontology revealing concept dependencies and systematic model biases.

Summary generated by Claude — human-verified