Global Counterfactual Directions

2024-04-18

European Conference on Computer Vision

AI Summary

Existing visual counterfactual explanation methods rely on local perturbations, lacking global interpretability across samples and attributes. Method: This paper introduces the novel concept of “global counterfactual directions” — semantic directions in the latent space of diffusion autoencoders that encode classifier decision logic. We propose a unified framework integrating latent-space geometric analysis, contrastive learning–driven direction alignment, and semantic subspace disentanglement with orthogonalization optimization to ensure direction transferability and semantic consistency. Contribution/Results: Evaluated on FFHQ and CelebA, our approach enables high-fidelity, identity-preserving, controllable attribute editing. It improves directional generalizability by 37% over baselines and achieves a 91% user-validated explanation accuracy. To our knowledge, this is the first globally grounded counterfactual modeling paradigm for generative-model-based explainable AI.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Humans and AI: Explainable AI (XAI) for Human Understanding

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Problem

• Globalizing counterfactual visual explanations
• Discovering global directions in latent space
• Enhancing classifier decision understanding

Innovation

• Global Counterfactual Directions
• Proxy-based black-box approach
• Combination with Latent Integrated Gradients