Probabilistic Programming for Explainable AI – Causal Bayesian Networks

Abstract

The increasing deployment of complex AI models, particularly deep neural networks, raises significant concerns about their interpretability and trustworthiness. Existing XAI techniques often provide post-hoc explanations that lack a clear connection to the model's underlying reasoning. This paper proposes a novel approach to building truly explainable AI systems by leveraging probabilistic programming languages and causal Bayesian networks. The core idea is to explicitly represent the model's decision-making process through a causal Bayesian network, which allows for the tracing of influence and identification of key features. We demonstrate how probabilistic programming facilitates the construction of these networks, providing a framework for generating transparent and understandable AI systems. The proposed method shifts the focus from black-box model explanations to a white-box understanding of the model's causal structure, thereby addressing a critical limitation in current XAI methodologies. This approach offers a robust foundation for building AI systems that not only achieve high accuracy but also provide clear and justifiable explanations for their decisions.