Pearl’s causal hierarchy defines three levels of problems in data-driven AI: (1) observation, (2) intervention, and (3) counterfactuals. Problems at higher levels of the hierarchy are more challenging to solve, as they require a deeper understanding of causality that cannot be learned from passively observed data alone. Regarding explainability, these problems and their solutions provide new challenges and opportunities for cross-fertilization with causal studies. In this special track, we focus on these challenges and opportunities. We are interested in the inner causality of models (such as mechanistic interpretability) and external causality (relates to how the model fits with its environment), both of which should enable safety and security analysis. Furthermore, inner causality should enable recognizing and replacing parts of models with coded modules that increase reliability and performance.
Topics
| Causal regularization techniques for xAI |
| Contrastive, interventional, backtracking, diffeomorphic counterfactual explanations |
| Inner causal explanations with mechanistic interpretability and concept-based learning |
| Causal Discovery with representation learning |
| Generative AI as preprocessing for xAI |
| Exploring causality with deep generative models |
| Generative flow networks for explainability |
| Normalizing flow networks for explainable artificial intelligence |
| Observational and interventional causal approaches to XAI |
| Explaining optimization solutions for causal decision intelligence in XAI |
| Explaining sources of infeasibility in optimization models |
| Explainability via amortized optimization models and fitness landscapes |
| Causal transportability |
| Digital twins and simulators for interpretable synthetic data generation |
| Graph neural network causal learning |
| Causal world models, distillation, importance in causal-explainable reinforcement learning |
| Interpretable/Explainable root cause analysis methods |
| Explainable AI with active inference techniques |
| Algorithmic information theory and algorithmic information dynamics for causal discovery |
| XAI methods for causality in spatiotemporal forecasting |
| Interpretable Causal Discovery via affective computing and consciousness studies |
| Observational and Interventional Causal Approaches to XAI |
| Causality and XAI in high-stakes decision-making |
| Topological data analysis and Information geometry for relations between explainability and causality |
| XAI in scientific discovery of mechanistic model |
| Optimal transport for counterfactual models |
