Concept-Based Abductive and Contrastive Explanations for Behaviors of Vision Models

Researchers have introduced a novel framework for explaining the predictions of deep neural networks in computer vision, termed concept-based abductive and contrastive explanations. While existing methods often lack causal connections or are limited to low-level pixel features, this new approach identifies minimal sets of high-level, human-understandable concepts that are causally relevant to model outcomes. The study presents algorithms that enumerate these minimal explanations using concept erasure procedures to establish causal relationships. By aggregating these explanations, the method enables users to understand model behaviors not just on individual images, but across collections exhibiting specific, user-defined behaviors. Evaluated across multiple models and datasets, the approach demonstrates effectiveness in generating helpful, user-friendly explanations. This work bridges the gap between concept-based interpretability and formal abductive reasoning, offering a more robust tool for analyzing and trusting AI vision systems. The paper was submitted to arXiv under Computer Science > Machine Learning, highlighting advancements in artificial intelligence interpretability and causal inference.