A SAT-centered XAI method for Deep Learning based Video Understanding

Formal guarantees improve trustworthiness: Unlike existing explanation methods for video AI systems, this approach uses SAT (satisfiability) solving to provide logic-based explanations with formal guarantees, meaning the explanations can be mathematically verified rather than just approximated.

Addresses a critical gap in video AI: The method specifically tackles explainability challenges in video understanding systems, which are increasingly used in security, surveillance, and content moderation—domains where understanding why an AI made a decision is particularly important.

Enables both positive and negative reasoning: The framework formulates both "Why?" and "Why not?" questions, allowing stakeholders to understand not just what the system detected in a video, but also why it rejected alternative interpretations—crucial for identifying potential biases or errors.

Computational complexity remains a practical challenge: The paper acknowledges that encoding video data and deep learning models into logical frameworks creates significant computational demands, which could limit real-world deployment in time-sensitive applications like live video monitoring.

Propositional logic has representational limits: The approach's reliance on propositional logic may struggle to capture the full complexity of how deep learning models actually process video, potentially oversimplifying explanations despite their formal guarantees.