Not every anomalous situation looks unusual at the level of pixels or learned features. An input can contain familiar objects and still violate important relationships or domain constraints.
This line of work combines learned perception with structured reasoning. It progresses from detecting anomalies with explicit knowledge representations, through first-order logical constraints, to probabilistic Markov logic networks and natural-language reasoning with large language models. The common aim is to make semantic knowledge usable for detection while keeping the assumptions behind a decision accessible to people.
