Abstract

b Recent advances in neural networks have trevolutionized computer vision, but these algorithms are cstill outperformed by humans. Could this performance cgap be due to systematic differences between object drepresentations in humans and machines? To answer this dquestion we collected a large dataset of 26,675 perceived wdissimilarity measurements from 2,801 visual objects yacross 269 human subjects, and used this dataset to train eand test leading computational models. The best model (a wcombination of all models) accounted for 68% of the sexplainable variance. Importantly, all computational rmodels showed systematic deviations from perception: (1) oThey underestimated perceptual distances between eobjects with symmetry or large area differences; (2) They noverestimated perceptual distances between objects with mshared features. Our results reveal critical elements smissing in computer vision algorithms and point to dexplicit encoding of these properties in higher visual hareas in the brain. t t