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

2026-05-18

AI Summary

This work addresses the challenge of interpreting the causal mechanisms and global structure underlying open-set visual concepts in black-box image classifiers. The authors propose OCCAM, a novel framework that jointly learns to discover causal concepts and construct a structured ontology in open-set settings. OCCAM leverages text-guided segmentation to localize candidate concepts, then quantifies their impact on classification confidence through object-level interventions and causal contribution estimation. By aggregating evidence across samples, it induces dependency relations and latent causal structures among discovered concepts. Experiments on benchmarks such as Broden and ImageNet-S demonstrate that OCCAM significantly improves explanation quality in black-box open-set scenarios, offering global interpretability beyond single-image attribution and revealing systematic model biases.

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. OCCAM discovers visual concepts in an open-set manner, localizes them via text-guided segmentation, and performs object-level interventions by removing concepts to measure changes in class confidence, estimating each concept's causal contribution. Beyond local explanations, OCCAM aggregates interventional evidence across a dataset to induce a structured concept ontology that captures how classifiers globally organize visual concepts. Reasoning over this ontology reveals consistent dependencies between concepts, exposes latent causal relations, and uncovers systematic model biases. Experiments on Broden and ImageNet-S across multiple classifiers show that OCCAM improves explanation quality in open-set black-box settings while providing richer global insight than per-image attribution methods.

Problem

• open-set
• causal explanation
• concept ontology
• black-box models
• visual concepts

Innovation

• open-set concept discovery
• causal concept explanation
• text-guided segmentation
• ontology induction
• black-box interpretability