Zero-shot visual reasoning through probabilistic analogical mapping

Human reasoning excels at identifying abstract commonalities across dissimilar visual inputs. Current algorithms require extensive training, limiting generalization. This paper introduces visiPAM (visual Probabilistic Analogical Mapping), synthesizing learned representations from visual inputs with a similarity-based mapping operation from cognitive science. Without direct training, visiPAM outperforms a state-of-the-art deep learning model on an analogical mapping task and closely matches human performance on a novel 3D object mapping task.

Taylor Webb, Shuhao Fu, Trevor Bihl, Keith J. Holyoak, Hongjing Lu