Table 1 Analogical mapping with 2D images

From: Zero-shot visual reasoning through probabilistic analogical mapping

 

Within-category

Between-category

 

Animals

Vehicles

Animals

VisiPAM

63.2% (59.1%)

69.5% (58.8%)

67.9% (59.9%)

VisiPAM (nodes only)

55.5% (50.6%)

61.6% (48.1%)

62.3% (52.9%)

VisiPAM (edges only)

47.8% (42%)

63.7% (50.9%)

51.5% (39.4%)

SSMN

46.6% (40.7%)

  

Random

10%

26%

20%

  1. Mapping accuracy on part-matching task proposed by Choi et al.25. Original task in ref. 25 involved the evaluation on within-category animal comparisons after training with 37,330 mapping problems. VisiPAM significantly outperformed SSMN, the previous state-of-the-art, despite having no direct training on mapping. VisiPAM also performed well on new problems involving within-category vehicle comparisons, and between-category animal comparisons (e.g., mapping from cat to horse). VisiPAM performed best when mapping was based on both node and edge similarity (as indicated by bold text). ‘Random’ denotes chance performance (determined by average number of part comparisons). Values in parentheses reflect chance-normalized performance (percentage of the range between chance performance and 100% accuracy). VisiPAM’s performance is roughly comparable across all conditions once chance performance is taken into account.
Back to article page