Physically grounded causal modeling for PLF
Résumé
We propose an approach for efficient, robust learning and reasoning under realistic conditions. Our approach emulates the way students acquire prior knowledge from the combination of practical and theoretical training. This prior knowledge then informs the interpretation of observations on the farm, and provides a causal model for the prediction of future events and the outcome of actions. Grounding the causal model in physical modeling fusing all sensors, enables future sensor data to be predicted. Recognition and segmentation of materials and objects is done on the physical model. This provides descriptions that are definitive, enabling one-shot, and zero-shot learning. Flow of mass, energy, momentum and force, facilitate recognition of actions and causes. Objects and actions are input to abstract causal model search. This allows tangible concepts of anatomy and physical examination to be linked to abstract concepts of physiology, nutrition and disease. The abstract causal model provides explanation in terms of counter-factuals, actions that would alter the outcome. Unlike deep networks, all the components are examinable and explicable
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |