Counterfactual Commonsense: Robust Spatio-Temporal Foundations for Knowledge-Centric Explanatory Intelligence

Consider this: Object A is to the left of B, and B is inside C.What is/are the possible spatial relationship/s between A and C? This seems deceptively easy (for humans); try it with IBM Watson and expect disappointment.

Indeed, (spatial) commonsense is deep: abstract, conceptual, semantic.Now imagine several interacting objects -people, cars, cyclists etc- geometrically abstracted in 2D/3D, e.g., as polygons.

Over time, some objects move or rotate, some don´t; some disappear and reappear; some merge and split, some exhibit motion patterns.

How do we computationally ``abstract and reason´´ about the dynamics and associated human context of such interactions?Thinking about how things exist, might have existed, or could reconfigure is what we call counterfactual (visuospatial) commonsense.

We develop the mathematical and computational foundations to represent, consistently maintain, and reason about dynamic, relational, counterfactual knowledge about space, motion, and interaction in everyday contexts.

Our method emphasises human-centred design: 1) inherent explainability of mechanisms for modelling and reasoning about space-time dynamics; 2) integration of semantic and quantitative data; and 3) handling uncertain and incomplete knowledge.Our foundational methods will be robustly applicable in diverse real-world AI applications.

We contribute directly to the EU Commission proposed Artificial Intelligence Act (2021) stipulating human-centred explicability of next-generation AI.