Actionable XAI for Deep learning Algorithms – AIDA

The goal of AIDA is to develop principled and rigorous techniques for actionable AI explanations (aXAI). To enhance transparency of spatio-temporal data/models AIDA advocates a causal modeling framework for Multivariate Time Series that captures the underlying temporal data generation process. Structured causal models serve to generate action-guiding explanations of Deep Learning (DL) models of increasing expressiveness (i.e., sufficient, counterfactual). AIDA causal explanations go beyond the traditional predictive performance vs interpretability trade-offs of existing xAI methods and provide new theoretical and applicative opportunities for AI trust and usability in the high stakes water resources domain at BRGM. By proposing new interestingness measures for explanations and explanations exploration strategies, we help end-users gain reliable insights on DL models. By leveraging the causal relationships among MTS variables we can build simpler surrogate models that may outperform the original DL models, while they can be reused without costly retraining in a novel context.