A Computational Method to Find Salient Features

Although it is well known that people selectively attend to salient features in similarity judgment, no clear method of identifying “salient features” has been proposed. In this study, we present a new computational technique to identify salient features. First, we collected behavioral data from human participants, and this data was simulated with machine learning techniques, which determined optimal allocations of weights of candidate features. Results revealed image-specific sets of salient features for similarity perception, and suggested that people exaggerate differences between features while computing similarity.